Chapter 8 Optimization

Table of Contents

8.1 Optimization Overview
8.2 Optimizing SQL Statements
8.2.1 Optimizing SELECT Statements
8.2.2 Optimizing DML Statements
8.2.3 Optimizing Database Privileges
8.2.4 Optimizing INFORMATION_SCHEMA Queries
8.2.5 Other Optimization Tips
8.3 Optimization and Indexes
8.3.1 How MySQL Uses Indexes
8.3.2 Using Primary Keys
8.3.3 Using Foreign Keys
8.3.4 Column Indexes
8.3.5 Multiple-Column Indexes
8.3.6 Verifying Index Usage
8.3.7 InnoDB and MyISAM Index Statistics Collection
8.3.8 Comparison of B-Tree and Hash Indexes
8.3.9 Optimizer Use of Generated Column Indexes
8.4 Optimizing Database Structure
8.4.1 Optimizing Data Size
8.4.2 Optimizing MySQL Data Types
8.4.3 Optimizing for Many Tables
8.4.4 Internal Temporary Table Use in MySQL
8.5 Optimizing for InnoDB Tables
8.5.1 Optimizing Storage Layout for InnoDB Tables
8.5.2 Optimizing InnoDB Transaction Management
8.5.3 Optimizing InnoDB Read-Only Transactions
8.5.4 Optimizing InnoDB Redo Logging
8.5.5 Bulk Data Loading for InnoDB Tables
8.5.6 Optimizing InnoDB Queries
8.5.7 Optimizing InnoDB DDL Operations
8.5.8 Optimizing InnoDB Disk I/O
8.5.9 Optimizing InnoDB Configuration Variables
8.5.10 Optimizing InnoDB for Systems with Many Tables
8.6 Optimizing for MyISAM Tables
8.6.1 Optimizing MyISAM Queries
8.6.2 Bulk Data Loading for MyISAM Tables
8.6.3 Speed of REPAIR TABLE Statements
8.7 Optimizing for MEMORY Tables
8.8 Understanding the Query Execution Plan
8.8.1 Optimizing Queries with EXPLAIN
8.8.2 EXPLAIN Output Format
8.8.3 EXPLAIN EXTENDED Output Format
8.8.4 Obtaining Execution Plan Information for a Named Connection
8.8.5 Estimating Query Performance
8.9 Controlling the Query Optimizer
8.9.1 Controlling Query Plan Evaluation
8.9.2 Controlling Switchable Optimizations
8.9.3 Optimizer Hints
8.9.4 Index Hints
8.9.5 The Optimizer Cost Model
8.10 Buffering and Caching
8.10.1 The InnoDB Buffer Pool
8.10.2 The MyISAM Key Cache
8.10.3 The MySQL Query Cache
8.10.4 Caching of Prepared Statements and Stored Programs
8.11 Optimizing Locking Operations
8.11.1 Internal Locking Methods
8.11.2 Table Locking Issues
8.11.3 Concurrent Inserts
8.11.4 Metadata Locking
8.11.5 External Locking
8.12 Optimizing the MySQL Server
8.12.1 System Factors and Startup Parameter Tuning
8.12.2 Tuning Server Parameters
8.12.3 Optimizing Disk I/O
8.12.4 Using Symbolic Links
8.12.5 Optimizing Memory Use
8.12.6 Optimizing Network Use
8.12.7 The Thread Pool Plugin
8.13 Measuring Performance (Benchmarking)
8.13.1 Measuring the Speed of Expressions and Functions
8.13.2 Using Your Own Benchmarks
8.13.3 Measuring Performance with performance_schema
8.14 Examining Thread Information
8.14.1 Thread Command Values
8.14.2 General Thread States
8.14.3 Query Cache Thread States
8.14.4 Replication Master Thread States
8.14.5 Replication Slave I/O Thread States
8.14.6 Replication Slave SQL Thread States
8.14.7 Replication Slave Connection Thread States
8.14.8 MySQL Cluster Thread States
8.14.9 Event Scheduler Thread States

This chapter explains how to optimize MySQL performance and provides examples. Optimization involves configuring, tuning, and measuring performance, at several levels. Depending on your job role (developer, DBA, or a combination of both), you might optimize at the level of individual SQL statements, entire applications, a single database server, or multiple networked database servers. Sometimes you can be proactive and plan in advance for performance, while other times you might troubleshoot a configuration or code issue after a problem occurs. Optimizing CPU and memory usage can also improve scalability, allowing the database to handle more load without slowing down.

8.1 Optimization Overview

Database performance depends on several factors at the database level, such as tables, queries, and configuration settings. These software constructs result in CPU and I/O operations at the hardware level, which you must minimize and make as efficient as possible. As you work on database performance, you start by learning the high-level rules and guidelines for the software side, and measuring performance using wall-clock time. As you become an expert, you learn more about what happens internally, and start measuring things such as CPU cycles and I/O operations.

Typical users aim to get the best database performance out of their existing software and hardware configurations. Advanced users look for opportunities to improve the MySQL software itself, or develop their own storage engines and hardware appliances to expand the MySQL ecosystem.

Optimizing at the Database Level

The most important factor in making a database application fast is its basic design:

  • Are the tables structured properly? In particular, do the columns have the right data types, and does each table have the appropriate columns for the type of work? For example, applications that perform frequent updates often have many tables with few columns, while applications that analyze large amounts of data often have few tables with many columns.

  • Are the right indexes in place to make queries efficient?

  • Are you using the appropriate storage engine for each table, and taking advantage of the strengths and features of each storage engine you use? In particular, the choice of a transactional storage engine such as InnoDB or a nontransactional one such as MyISAM can be very important for performance and scalability.

    Note

    In MySQL 5.5 and higher, InnoDB is the default storage engine for new tables. In practice, the advanced InnoDB performance features mean that InnoDB tables often outperform the simpler MyISAM tables, especially for a busy database.

  • Does each table use an appropriate row format? This choice also depends on the storage engine used for the table. In particular, compressed tables use less disk space and so require less disk I/O to read and write the data. Compression is available for all kinds of workloads with InnoDB tables, and for read-only MyISAM tables.

  • Does the application use an appropriate locking strategy? For example, by allowing shared access when possible so that database operations can run concurrently, and requesting exclusive access when appropriate so that critical operations get top priority. Again, the choice of storage engine is significant. The InnoDB storage engine handles most locking issues without involvement from you, allowing for better concurrency in the database and reducing the amount of experimentation and tuning for your code.

  • Are all memory areas used for caching sized correctly? That is, large enough to hold frequently accessed data, but not so large that they overload physical memory and cause paging. The main memory areas to configure are the InnoDB buffer pool, the MyISAM key cache, and the MySQL query cache.

Optimizing at the Hardware Level

Any database application eventually hits hardware limits as the database becomes more and more busy. A DBA must evaluate whether it is possible to tune the application or reconfigure the server to avoid these bottlenecks, or whether more hardware resources are required. System bottlenecks typically arise from these sources:

  • Disk seeks. It takes time for the disk to find a piece of data. With modern disks, the mean time for this is usually lower than 10ms, so we can in theory do about 100 seeks a second. This time improves slowly with new disks and is very hard to optimize for a single table. The way to optimize seek time is to distribute the data onto more than one disk.

  • Disk reading and writing. When the disk is at the correct position, we need to read or write the data. With modern disks, one disk delivers at least 10–20MB/s throughput. This is easier to optimize than seeks because you can read in parallel from multiple disks.

  • CPU cycles. When the data is in main memory, we must process it to get our result. Having large tables compared to the amount of memory is the most common limiting factor. But with small tables, speed is usually not the problem.

  • Memory bandwidth. When the CPU needs more data than can fit in the CPU cache, main memory bandwidth becomes a bottleneck. This is an uncommon bottleneck for most systems, but one to be aware of.

Balancing Portability and Performance

To use performance-oriented SQL extensions in a portable MySQL program, you can wrap MySQL-specific keywords in a statement within /*! */ comment delimiters. Other SQL servers ignore the commented keywords. For information about writing comments, see Section 9.6, “Comment Syntax”.

8.2 Optimizing SQL Statements

The core logic of a database application is performed through SQL statements, whether issued directly through an interpreter or submitted behind the scenes through an API. The tuning guidelines in this section help to speed up all kinds of MySQL applications. The guidelines cover SQL operations that read and write data, the behind-the-scenes overhead for SQL operations in general, and operations used in specific scenarios such as database monitoring.

8.2.1 Optimizing SELECT Statements

Queries, in the form of SELECT statements, perform all the lookup operations in the database. Tuning these statements is a top priority, whether to achieve sub-second response times for dynamic web pages, or to chop hours off the time to generate huge overnight reports.

Besides SELECT statements, the tuning techniques for queries also apply to constructs such as CREATE TABLE...AS SELECT, INSERT INTO...SELECT, and WHERE clauses in DELETE statements. Those statements have additional performance considerations because they combine write operations with the read-oriented query operations.

MySQL Cluster supports a join pushdown optimization whereby a qualifying join is sent in its entirety to MySQL Cluster data nodes, where it can be distributed among them and executed in parallel. For more information about this optimization, see Conditions for NDB pushdown joins,

8.2.1.1 Speed of SELECT Statements

The main considerations for optimizing queries are:

  • To make a slow SELECT ... WHERE query faster, the first thing to check is whether you can add an index. Set up indexes on columns used in the WHERE clause, to speed up evaluation, filtering, and the final retrieval of results. To avoid wasted disk space, construct a small set of indexes that speed up many related queries used in your application.

    Indexes are especially important for queries that reference different tables, using features such as joins and foreign keys. You can use the EXPLAIN statement to determine which indexes are used for a SELECT. See Section 8.3.1, “How MySQL Uses Indexes” and Section 8.8.1, “Optimizing Queries with EXPLAIN”.

  • Isolate and tune any part of the query, such as a function call, that takes excessive time. Depending on how the query is structured, a function could be called once for every row in the result set, or even once for every row in the table, greatly magnifying any inefficiency.

  • Minimize the number of full table scans in your queries, particularly for big tables.

  • Keep table statistics up to date by using the ANALYZE TABLE statement periodically, so the optimizer has the information needed to construct an efficient execution plan.

  • Learn the tuning techniques, indexing techniques, and configuration parameters that are specific to the storage engine for each table. Both InnoDB and MyISAM have sets of guidelines for enabling and sustaining high performance in queries. For details, see Section 8.5.6, “Optimizing InnoDB Queries” and Section 8.6.1, “Optimizing MyISAM Queries”.

  • You can optimize single-query transactions for InnoDB tables, using the technique in Section 8.5.3, “Optimizing InnoDB Read-Only Transactions”.

  • Avoid transforming the query in ways that make it hard to understand, especially if the optimizer does some of the same transformations automatically.

  • If a performance issue is not easily solved by one of the basic guidelines, investigate the internal details of the specific query by reading the EXPLAIN plan and adjusting your indexes, WHERE clauses, join clauses, and so on. (When you reach a certain level of expertise, reading the EXPLAIN plan might be your first step for every query.)

  • Adjust the size and properties of the memory areas that MySQL uses for caching. With efficient use of the InnoDB buffer pool, MyISAM key cache, and the MySQL query cache, repeated queries run faster because the results are retrieved from memory the second and subsequent times.

  • Even for a query that runs fast using the cache memory areas, you might still optimize further so that they require less cache memory, making your application more scalable. Scalability means that your application can handle more simultaneous users, larger requests, and so on without experiencing a big drop in performance.

  • Deal with locking issues, where the speed of your query might be affected by other sessions accessing the tables at the same time.

8.2.1.2 How MySQL Optimizes WHERE Clauses

This section discusses optimizations that can be made for processing WHERE clauses. The examples use SELECT statements, but the same optimizations apply for WHERE clauses in DELETE and UPDATE statements.

Note

Because work on the MySQL optimizer is ongoing, not all of the optimizations that MySQL performs are documented here.

You might be tempted to rewrite your queries to make arithmetic operations faster, while sacrificing readability. Because MySQL does similar optimizations automatically, you can often avoid this work, and leave the query in a more understandable and maintainable form. Some of the optimizations performed by MySQL follow:

  • Removal of unnecessary parentheses:

       ((a AND b) AND c OR (((a AND b) AND (c AND d))))
    -> (a AND b AND c) OR (a AND b AND c AND d)
    
  • Constant folding:

       (a<b AND b=c) AND a=5
    -> b>5 AND b=c AND a=5
    
  • Constant condition removal (needed because of constant folding):

       (B>=5 AND B=5) OR (B=6 AND 5=5) OR (B=7 AND 5=6)
    -> B=5 OR B=6
    
  • Constant expressions used by indexes are evaluated only once.

  • COUNT(*) on a single table without a WHERE is retrieved directly from the table information for MyISAM and MEMORY tables. This is also done for any NOT NULL expression when used with only one table.

  • Early detection of invalid constant expressions. MySQL quickly detects that some SELECT statements are impossible and returns no rows.

  • HAVING is merged with WHERE if you do not use GROUP BY or aggregate functions (COUNT(), MIN(), and so on).

  • For each table in a join, a simpler WHERE is constructed to get a fast WHERE evaluation for the table and also to skip rows as soon as possible.

  • All constant tables are read first before any other tables in the query. A constant table is any of the following:

    • An empty table or a table with one row.

    • A table that is used with a WHERE clause on a PRIMARY KEY or a UNIQUE index, where all index parts are compared to constant expressions and are defined as NOT NULL.

    All of the following tables are used as constant tables:

    SELECT * FROM t WHERE primary_key=1;
    SELECT * FROM t1,t2
      WHERE t1.primary_key=1 AND t2.primary_key=t1.id;
    
  • The best join combination for joining the tables is found by trying all possibilities. If all columns in ORDER BY and GROUP BY clauses come from the same table, that table is preferred first when joining.

  • If there is an ORDER BY clause and a different GROUP BY clause, or if the ORDER BY or GROUP BY contains columns from tables other than the first table in the join queue, a temporary table is created.

  • If you use the SQL_SMALL_RESULT option, MySQL uses an in-memory temporary table.

  • Each table index is queried, and the best index is used unless the optimizer believes that it is more efficient to use a table scan. At one time, a scan was used based on whether the best index spanned more than 30% of the table, but a fixed percentage no longer determines the choice between using an index or a scan. The optimizer now is more complex and bases its estimate on additional factors such as table size, number of rows, and I/O block size.

  • In some cases, MySQL can read rows from the index without even consulting the data file. If all columns used from the index are numeric, only the index tree is used to resolve the query.

  • Before each row is output, those that do not match the HAVING clause are skipped.

Some examples of queries that are very fast:

SELECT COUNT(*) FROM tbl_name;

SELECT MIN(key_part1),MAX(key_part1) FROM tbl_name;

SELECT MAX(key_part2) FROM tbl_name
  WHERE key_part1=constant;

SELECT ... FROM tbl_name
  ORDER BY key_part1,key_part2,... LIMIT 10;

SELECT ... FROM tbl_name
  ORDER BY key_part1 DESC, key_part2 DESC, ... LIMIT 10;

MySQL resolves the following queries using only the index tree, assuming that the indexed columns are numeric:

SELECT key_part1,key_part2 FROM tbl_name WHERE key_part1=val;

SELECT COUNT(*) FROM tbl_name
  WHERE key_part1=val1 AND key_part2=val2;

SELECT key_part2 FROM tbl_name GROUP BY key_part1;

The following queries use indexing to retrieve the rows in sorted order without a separate sorting pass:

SELECT ... FROM tbl_name
  ORDER BY key_part1,key_part2,... ;

SELECT ... FROM tbl_name
  ORDER BY key_part1 DESC, key_part2 DESC, ... ;

8.2.1.3 Range Optimization

The range access method uses a single index to retrieve a subset of table rows that are contained within one or several index value intervals. It can be used for a single-part or multiple-part index. The following sections give descriptions of conditions under which the optimizer uses range access.

8.2.1.3.1 The Range Access Method for Single-Part Indexes

For a single-part index, index value intervals can be conveniently represented by corresponding conditions in the WHERE clause, denoted as range conditions rather than intervals.

The definition of a range condition for a single-part index is as follows:

  • For both BTREE and HASH indexes, comparison of a key part with a constant value is a range condition when using the =, <=>, IN(), IS NULL, or IS NOT NULL operators.

  • Additionally, for BTREE indexes, comparison of a key part with a constant value is a range condition when using the >, <, >=, <=, BETWEEN, !=, or <> operators, or LIKE comparisons if the argument to LIKE is a constant string that does not start with a wildcard character.

  • For all index types, multiple range conditions combined with OR or AND form a range condition.

Constant value in the preceding descriptions means one of the following:

  • A constant from the query string

  • A column of a const or system table from the same join

  • The result of an uncorrelated subquery

  • Any expression composed entirely from subexpressions of the preceding types

Here are some examples of queries with range conditions in the WHERE clause:

SELECT * FROM t1
  WHERE key_col > 1
  AND key_col < 10;

SELECT * FROM t1
  WHERE key_col = 1
  OR key_col IN (15,18,20);

SELECT * FROM t1
  WHERE key_col LIKE 'ab%'
  OR key_col BETWEEN 'bar' AND 'foo';

Some nonconstant values may be converted to constants during the optimizer constant propagation phase.

MySQL tries to extract range conditions from the WHERE clause for each of the possible indexes. During the extraction process, conditions that cannot be used for constructing the range condition are dropped, conditions that produce overlapping ranges are combined, and conditions that produce empty ranges are removed.

Consider the following statement, where key1 is an indexed column and nonkey is not indexed:

SELECT * FROM t1 WHERE
  (key1 < 'abc' AND (key1 LIKE 'abcde%' OR key1 LIKE '%b')) OR
  (key1 < 'bar' AND nonkey = 4) OR
  (key1 < 'uux' AND key1 > 'z');

The extraction process for key key1 is as follows:

  1. Start with original WHERE clause:

    (key1 < 'abc' AND (key1 LIKE 'abcde%' OR key1 LIKE '%b')) OR
    (key1 < 'bar' AND nonkey = 4) OR
    (key1 < 'uux' AND key1 > 'z')
    
  2. Remove nonkey = 4 and key1 LIKE '%b' because they cannot be used for a range scan. The correct way to remove them is to replace them with TRUE, so that we do not miss any matching rows when doing the range scan. Having replaced them with TRUE, we get:

    (key1 < 'abc' AND (key1 LIKE 'abcde%' OR TRUE)) OR
    (key1 < 'bar' AND TRUE) OR
    (key1 < 'uux' AND key1 > 'z')
    
  3. Collapse conditions that are always true or false:

    • (key1 LIKE 'abcde%' OR TRUE) is always true

    • (key1 < 'uux' AND key1 > 'z') is always false

    Replacing these conditions with constants, we get:

    (key1 < 'abc' AND TRUE) OR (key1 < 'bar' AND TRUE) OR (FALSE)
    

    Removing unnecessary TRUE and FALSE constants, we obtain:

    (key1 < 'abc') OR (key1 < 'bar')
    
  4. Combining overlapping intervals into one yields the final condition to be used for the range scan:

    (key1 < 'bar')
    

In general (and as demonstrated by the preceding example), the condition used for a range scan is less restrictive than the WHERE clause. MySQL performs an additional check to filter out rows that satisfy the range condition but not the full WHERE clause.

The range condition extraction algorithm can handle nested AND/OR constructs of arbitrary depth, and its output does not depend on the order in which conditions appear in WHERE clause.

MySQL does not support merging multiple ranges for the range access method for spatial indexes. To work around this limitation, you can use a UNION with identical SELECT statements, except that you put each spatial predicate in a different SELECT.

8.2.1.3.2 The Range Access Method for Multiple-Part Indexes

Range conditions on a multiple-part index are an extension of range conditions for a single-part index. A range condition on a multiple-part index restricts index rows to lie within one or several key tuple intervals. Key tuple intervals are defined over a set of key tuples, using ordering from the index.

For example, consider a multiple-part index defined as key1(key_part1, key_part2, key_part3), and the following set of key tuples listed in key order:

key_part1  key_part2  key_part3
  NULL       1          'abc'
  NULL       1          'xyz'
  NULL       2          'foo'
   1         1          'abc'
   1         1          'xyz'
   1         2          'abc'
   2         1          'aaa'

The condition key_part1 = 1 defines this interval:

(1,-inf,-inf) <= (key_part1,key_part2,key_part3) < (1,+inf,+inf)

The interval covers the 4th, 5th, and 6th tuples in the preceding data set and can be used by the range access method.

By contrast, the condition key_part3 = 'abc' does not define a single interval and cannot be used by the range access method.

The following descriptions indicate how range conditions work for multiple-part indexes in greater detail.

  • For HASH indexes, each interval containing identical values can be used. This means that the interval can be produced only for conditions in the following form:

        key_part1 cmp const1
    AND key_part2 cmp const2
    AND ...
    AND key_partN cmp constN;
    

    Here, const1, const2, … are constants, cmp is one of the =, <=>, or IS NULL comparison operators, and the conditions cover all index parts. (That is, there are N conditions, one for each part of an N-part index.) For example, the following is a range condition for a three-part HASH index:

    key_part1 = 1 AND key_part2 IS NULL AND key_part3 = 'foo'
    

    For the definition of what is considered to be a constant, see Section 8.2.1.3.1, “The Range Access Method for Single-Part Indexes”.

  • For a BTREE index, an interval might be usable for conditions combined with AND, where each condition compares a key part with a constant value using =, <=>, IS NULL, >, <, >=, <=, !=, <>, BETWEEN, or LIKE 'pattern' (where 'pattern' does not start with a wildcard). An interval can be used as long as it is possible to determine a single key tuple containing all rows that match the condition (or two intervals if <> or != is used).

    The optimizer attempts to use additional key parts to determine the interval as long as the comparison operator is =, <=>, or IS NULL. If the operator is >, <, >=, <=, !=, <>, BETWEEN, or LIKE, the optimizer uses it but considers no more key parts. For the following expression, the optimizer uses = from the first comparison. It also uses >= from the second comparison but considers no further key parts and does not use the third comparison for interval construction:

    key_part1 = 'foo' AND key_part2 >= 10 AND key_part3 > 10
    

    The single interval is:

    ('foo',10,-inf) < (key_part1,key_part2,key_part3) < ('foo',+inf,+inf)
    

    It is possible that the created interval contains more rows than the initial condition. For example, the preceding interval includes the value ('foo', 11, 0), which does not satisfy the original condition.

  • If conditions that cover sets of rows contained within intervals are combined with OR, they form a condition that covers a set of rows contained within the union of their intervals. If the conditions are combined with AND, they form a condition that covers a set of rows contained within the intersection of their intervals. For example, for this condition on a two-part index:

    (key_part1 = 1 AND key_part2 < 2) OR (key_part1 > 5)
    

    The intervals are:

    (1,-inf) < (key_part1,key_part2) < (1,2)
    (5,-inf) < (key_part1,key_part2)
    

    In this example, the interval on the first line uses one key part for the left bound and two key parts for the right bound. The interval on the second line uses only one key part. The key_len column in the EXPLAIN output indicates the maximum length of the key prefix used.

    In some cases, key_len may indicate that a key part was used, but that might be not what you would expect. Suppose that key_part1 and key_part2 can be NULL. Then the key_len column displays two key part lengths for the following condition:

    key_part1 >= 1 AND key_part2 < 2
    

    But, in fact, the condition is converted to this:

    key_part1 >= 1 AND key_part2 IS NOT NULL
    

Section 8.2.1.3.1, “The Range Access Method for Single-Part Indexes”, describes how optimizations are performed to combine or eliminate intervals for range conditions on a single-part index. Analogous steps are performed for range conditions on multiple-part indexes.

8.2.1.3.3 Equality Range Optimization of Many-Valued Comparisons

Consider these expressions, where col_name is an indexed column:

col_name IN(val1, ..., valN)
col_name = val1 OR ... OR col_name = valN

Each expression is true if col_name is equal to any of several values. These comparisons are equality range comparisons (where the range is a single value). The optimizer estimates the cost of reading qualifying rows for equality range comparisons as follows:

  • If there is a unique index on col_name, the row estimate for each range is 1 because at most one row can have the given value.

  • Otherwise, any index on col_name is nonunique and the optimizer can estimate the row count for each range using dives into the index or index statistics.

With index dives, the optimizer makes a dive at each end of a range and uses the number of rows in the range as the estimate. For example, the expression col_name IN (10, 20, 30) has three equality ranges and the optimizer makes two dives per range to generate a row estimate. Each pair of dives yields an estimate of the number of rows that have the given value.

Index dives provide accurate row estimates, but as the number of comparison values in the expression increases, the optimizer takes longer to generate a row estimate. Use of index statistics is less accurate than index dives but permits faster row estimation for large value lists.

The eq_range_index_dive_limit system variable enables you to configure the number of values at which the optimizer switches from one row estimation strategy to the other. To permit use of index dives for comparisons of up to N equality ranges, set eq_range_index_dive_limit to N + 1. To disable use of statistics and always use index dives regardless of N, set eq_range_index_dive_limit to 0.

To update table index statistics for best estimates, use ANALYZE TABLE.

8.2.1.3.4 Limiting Memory Use for Range Optimization

To control the memory available to the range optimizer, use the range_optimizer_max_mem_size system variable:

  • A value of 0 means no limit.

  • With a value greater than 0, the optimizer tracks the memory consumed when considering the range access method. If the specified limit is about to be exceeded, the range access method is abandoned and other methods, including a full table scan, are considered instead. This could be less optimal. If this happens, the following warning occurs (where N is the current range_optimizer_max_mem_size value):

    Warning    3170    Memory capacity of N bytes for
                       'range_optimizer_max_mem_size' exceeded. Range
                       optimization was not done for this query.
    

For individual queries that exceed the available range optimization memory and for which the optimizer falls back to less optimal plans, increasing the range_optimizer_max_mem_size value may improve performance.

To estimate the amount of memory needed to process a range expression, use these guidelines:

  • For a simple query such as the following, where there is one candidate key for for the range access method, each predicate combined with OR uses approximately 230 bytes:

    SELECT COUNT(*) FROM t
    WHERE a=1 OR a=2 OR a=3 OR .. . a=N;
    
  • Similarly for a query such as the following, each predicate combined with AND uses approximately 125 bytes:

    SELECT COUNT(*) FROM t
    WHERE a=1 AND b=1 AND c=1 ... N;
    
  • For a query with IN() predicates:

    SELECT COUNT(*) FROM t
    WHERE a IN (1,2, ..., M) AND b IN (1,2, ..., N);
    

    Each literal value in an IN() list counts as a predicate combined with OR. If there are two IN() lists, the number of predicates combined with OR is the product of the number of literal values in each list. Thus, the number of predicates combined with OR in the preceding case is M × N.

Before 5.7.11, the number of bytes per predicate combined with OR was higher, approximately 700 bytes.

Before MySQL 5.7.9 (that is, before the introduction of range_optimizer_max_mem_size), the optimizer had a hardcoded limit of 16,000 items that could be present in a WHERE clause. When this limit was reached, the optimizer abandoned the plan and considered other plans. A range_optimizer_max_mem_size value of 4MB in current MySQL versions corresponds approximately to the same 16,000 item limit applicable before MySQL 5.7.9.

However, note that before MySQL 5.7.9, if a query was written in a certain way, the limitation on the number of items used in a WHERE clause could be bypassed. For such queries as of 5.7.9, it may be necessary to increase the range_optimizer_max_mem_size value to make the optimizer consider the range access method again. Also, memory utilized for items in a WHERE clause might vary depending on the way the condition is written. Hence, the preceding approximations should be used as guidelines only.

8.2.1.3.5 Range Optimization of Row Constructor Expressions

As of MySQL 5.7.3, the optimizer is able to apply the range scan access method to queries of this form:

SELECT ... FROM t1 WHERE ( col_1, col_2 ) IN (( 'a', 'b' ), ( 'c', 'd' ));

Previously, for range scans to be used, it was necessary to write the query as:

SELECT ... FROM t1 WHERE ( col_1 = 'a' AND col_2 = 'b' )
OR ( col_1 = 'c' AND col_2 = 'd' );

For the optimizer to use a range scan, queries must satisfy these conditions:

  • Only IN() predicates are used, not NOT IN().

  • On the left side of the IN() predicate, the row constructor contains only column references.

  • On the right side of the IN() predicate, row constructors contain only runtime constants, which are either literals or local column references that are bound to constants during execution.

  • On the right side of the IN() predicate, there is more than one row constructor.

Compared to similar queries executed before MySQL 5.7.3, EXPLAIN output for applicable queries changes from full table scan or index scan to range scan. Changes are also visible by checking the values of the Handler_read_first, Handler_read_key, and Handler_read_next status variables.

For more information about the optimizer and row constructors, see Section 8.2.1.20, “Row Constructor Expression Optimization”

8.2.1.4 Index Merge Optimization

The Index Merge method is used to retrieve rows with several range scans and to merge their results into one. The merge can produce unions, intersections, or unions-of-intersections of its underlying scans. This access method merges index scans from a single table; it does not merge scans across multiple tables.

In EXPLAIN output, the Index Merge method appears as index_merge in the type column. In this case, the key column contains a list of indexes used, and key_len contains a list of the longest key parts for those indexes.

Examples:

SELECT * FROM tbl_name WHERE key1 = 10 OR key2 = 20;

SELECT * FROM tbl_name
  WHERE (key1 = 10 OR key2 = 20) AND non_key=30;

SELECT * FROM t1, t2
  WHERE (t1.key1 IN (1,2) OR t1.key2 LIKE 'value%')
  AND t2.key1=t1.some_col;

SELECT * FROM t1, t2
  WHERE t1.key1=1
  AND (t2.key1=t1.some_col OR t2.key2=t1.some_col2);

The Index Merge method has several access algorithms (seen in the Extra field of EXPLAIN output):

  • Using intersect(...)

  • Using union(...)

  • Using sort_union(...)

The following sections describe these methods in greater detail.

Note

The Index Merge optimization algorithm has the following known deficiencies:

  • If your query has a complex WHERE clause with deep AND/OR nesting and MySQL does not choose the optimal plan, try distributing terms using the following identity laws:

    (x AND y) OR z = (x OR z) AND (y OR z)
    (x OR y) AND z = (x AND z) OR (y AND z)
    
  • Index Merge is not applicable to full-text indexes. We plan to extend it to cover these in a future MySQL release.

The choice between different possible variants of the Index Merge access method and other access methods is based on cost estimates of various available options.

8.2.1.4.1 The Index Merge Intersection Access Algorithm

This access algorithm can be employed when a WHERE clause was converted to several range conditions on different keys combined with AND, and each condition is one of the following:

  • In this form, where the index has exactly N parts (that is, all index parts are covered):

    key_part1=const1 AND key_part2=const2 ... AND key_partN=constN
    
  • Any range condition over a primary key of an InnoDB table.

Examples:

SELECT * FROM innodb_table WHERE primary_key < 10 AND key_col1=20;

SELECT * FROM tbl_name
  WHERE (key1_part1=1 AND key1_part2=2) AND key2=2;

The Index Merge intersection algorithm performs simultaneous scans on all used indexes and produces the intersection of row sequences that it receives from the merged index scans.

If all columns used in the query are covered by the used indexes, full table rows are not retrieved (EXPLAIN output contains Using index in Extra field in this case). Here is an example of such a query:

SELECT COUNT(*) FROM t1 WHERE key1=1 AND key2=1;

If the used indexes do not cover all columns used in the query, full rows are retrieved only when the range conditions for all used keys are satisfied.

If one of the merged conditions is a condition over a primary key of an InnoDB table, it is not used for row retrieval, but is used to filter out rows retrieved using other conditions.

8.2.1.4.2 The Index Merge Union Access Algorithm

The applicability criteria for this algorithm are similar to those for the Index Merge method intersection algorithm. The algorithm can be employed when the table's WHERE clause was converted to several range conditions on different keys combined with OR, and each condition is one of the following:

  • In this form, where the index has exactly N parts (that is, all index parts are covered):

    key_part1=const1 AND key_part2=const2 ... AND key_partN=constN
    
  • Any range condition over a primary key of an InnoDB table.

  • A condition for which the Index Merge method intersection algorithm is applicable.

Examples:

SELECT * FROM t1 WHERE key1=1 OR key2=2 OR key3=3;

SELECT * FROM innodb_table WHERE (key1=1 AND key2=2) OR
  (key3='foo' AND key4='bar') AND key5=5;
8.2.1.4.3 The Index Merge Sort-Union Access Algorithm

This access algorithm is employed when the WHERE clause was converted to several range conditions combined by OR, but for which the Index Merge method union algorithm is not applicable.

Examples:

SELECT * FROM tbl_name WHERE key_col1 < 10 OR key_col2 < 20;

SELECT * FROM tbl_name
  WHERE (key_col1 > 10 OR key_col2 = 20) AND nonkey_col=30;

The difference between the sort-union algorithm and the union algorithm is that the sort-union algorithm must first fetch row IDs for all rows and sort them before returning any rows.

8.2.1.5 Engine Condition Pushdown Optimization

This optimization improves the efficiency of direct comparisons between a nonindexed column and a constant. In such cases, the condition is pushed down to the storage engine for evaluation. This optimization can be used only by the NDB storage engine.

For MySQL Cluster, this optimization can eliminate the need to send nonmatching rows over the network between the cluster's data nodes and the MySQL Server that issued the query, and can speed up queries where it is used by a factor of 5 to 10 times over cases where condition pushdown could be but is not used.

Suppose that a MySQL Cluster table is defined as follows:

CREATE TABLE t1 (
    a INT,
    b INT,
    KEY(a)
) ENGINE=NDB;

Condition pushdown can be used with queries such as the one shown here, which includes a comparison between a nonindexed column and a constant:

SELECT a, b FROM t1 WHERE b = 10;

The use of condition pushdown can be seen in the output of EXPLAIN:

mysql> EXPLAIN SELECT a,b FROM t1 WHERE b = 10\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: t1
         type: ALL
possible_keys: NULL
          key: NULL
      key_len: NULL
          ref: NULL
         rows: 10
        Extra: Using where with pushed condition

However, condition pushdown cannot be used with either of these two queries:

SELECT a,b FROM t1 WHERE a = 10;
SELECT a,b FROM t1 WHERE b + 1 = 10;

Condition pushdown is not applicable to the first query because an index exists on column a. (An index access method would be more efficient and so would be chosen in preference to condition pushdown.) Condition pushdown cannot be employed for the second query because the comparison involving the nonindexed column b is indirect. (However, condition pushdown could be applied if you were to reduce b + 1 = 10 to b = 9 in the WHERE clause.)

Condition pushdown may also be employed when an indexed column is compared with a constant using a > or < operator:

mysql> EXPLAIN SELECT a, b FROM t1 WHERE a < 2\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: t1
         type: range
possible_keys: a
          key: a
      key_len: 5
          ref: NULL
         rows: 2
        Extra: Using where with pushed condition

Other supported comparisons for condition pushdown include the following:

  • column [NOT] LIKE pattern

    pattern must be a string literal containing the pattern to be matched; for syntax, see Section 12.5.1, “String Comparison Functions”.

  • column IS [NOT] NULL

  • column IN (value_list)

    Each item in the value_list must be a constant, literal value.

  • column BETWEEN constant1 AND constant2

    constant1 and constant2 must each be a constant, literal value.

In all of the cases in the preceding list, it is possible for the condition to be converted into the form of one or more direct comparisons between a column and a constant.

Engine condition pushdown is enabled by default. To disable it at server startup, set the optimizer_switch system variable. For example, in a my.cnf file, use these lines:

[mysqld]
optimizer_switch=engine_condition_pushdown=off

At runtime, enable condition pushdown like this:

SET optimizer_switch='engine_condition_pushdown=off';

Limitations.  Engine condition pushdown is subject to the following limitations:

  • Condition pushdown is supported only by the NDB storage engine.

  • Columns may be compared with constants only; however, this includes expressions which evaluate to constant values.

  • Columns used in comparisons cannot be of any of the BLOB or TEXT types.

  • A string value to be compared with a column must use the same collation as the column.

  • Joins are not directly supported; conditions involving multiple tables are pushed separately where possible. Use EXPLAIN EXTENDED to determine which conditions are actually pushed down.

8.2.1.6 Index Condition Pushdown Optimization

Index Condition Pushdown (ICP) is an optimization for the case where MySQL retrieves rows from a table using an index. Without ICP, the storage engine traverses the index to locate rows in the base table and returns them to the MySQL server which evaluates the WHERE condition for the rows. With ICP enabled, and if parts of the WHERE condition can be evaluated by using only fields from the index, the MySQL server pushes this part of the WHERE condition down to the storage engine. The storage engine then evaluates the pushed index condition by using the index entry and only if this is satisfied is the row read from the table. ICP can reduce the number of times the storage engine must access the base table and the number of times the MySQL server must access the storage engine.

Index Condition Pushdown optimization is used for the range, ref, eq_ref, and ref_or_null access methods when there is a need to access full table rows. This strategy can be used for InnoDB and MyISAM tables. Beginning with MySQL 5.7.3, it can also be used with partitioned InnoDB and MyISAM tables (Bug #17306882, Bug #70001). For InnoDB tables, however, ICP is used only for secondary indexes. The goal of ICP is to reduce the number of full-record reads and thereby reduce IO operations. For InnoDB clustered indexes, the complete record is already read into the InnoDB buffer. Using ICP in this case does not reduce IO.

The ICP optimization is not supported with secondary indexes created on virtual generated columns. InnoDB supports secondary indexes on virtual generated columns as of MySQL 5.7.8.

To see how this optimization works, consider first how an index scan proceeds when Index Condition Pushdown is not used:

  1. Get the next row, first by reading the index tuple, and then by using the index tuple to locate and read the full table row.

  2. Test the part of the WHERE condition that applies to this table. Accept or reject the row based on the test result.

When Index Condition Pushdown is used, the scan proceeds like this instead:

  1. Get the next row's index tuple (but not the full table row).

  2. Test the part of the WHERE condition that applies to this table and can be checked using only index columns. If the condition is not satisfied, proceed to the index tuple for the next row.

  3. If the condition is satisfied, use the index tuple to locate and read the full table row.

  4. Test the remaining part of the WHERE condition that applies to this table. Accept or reject the row based on the test result.

When Index Condition Pushdown is used, the Extra column in EXPLAIN output shows Using index condition. It will not show Index only because that does not apply when full table rows must be read.

Suppose that we have a table containing information about people and their addresses and that the table has an index defined as INDEX (zipcode, lastname, firstname). If we know a person's zipcode value but are not sure about the last name, we can search like this:

SELECT * FROM people
  WHERE zipcode='95054'
  AND lastname LIKE '%etrunia%'
  AND address LIKE '%Main Street%';

MySQL can use the index to scan through people with zipcode='95054'. The second part (lastname LIKE '%etrunia%') cannot be used to limit the number of rows that must be scanned, so without Index Condition Pushdown, this query must retrieve full table rows for all the people who have zipcode='95054'.

With Index Condition Pushdown, MySQL will check the lastname LIKE '%etrunia%' part before reading the full table row. This avoids reading full rows corresponding to all index tuples that do not match the lastname condition.

Index Condition Pushdown is enabled by default; it can be controlled with the optimizer_switch system variable by setting the index_condition_pushdown flag. See Section 8.9.2, “Controlling Switchable Optimizations”.

8.2.1.7 Use of Index Extensions

InnoDB automatically extends each secondary index by appending the primary key columns to it. Consider this table definition:

CREATE TABLE t1 (
  i1 INT NOT NULL DEFAULT 0,
  i2 INT NOT NULL DEFAULT 0,
  d DATE DEFAULT NULL,
  PRIMARY KEY (i1, i2),
  INDEX k_d (d)
) ENGINE = InnoDB;

This table defines the primary key on columns (i1, i2). It also defines a secondary index k_d on column (d), but internally InnoDB extends this index and treats it as columns (d, i1, i2).

The optimizer takes into account the primary key columns of the extended secondary index when determining how and whether to use that index. This can result in more efficient query execution plans and better performance.

The optimizer can use extended secondary indexes for ref, range, and index_merge index access, for loose index scans, for join and sorting optimization, and for MIN()/MAX() optimization.

The following example shows how execution plans are affected by whether the optimizer uses extended secondary indexes. Suppose that t1 is populated with these rows:

INSERT INTO t1 VALUES
(1, 1, '1998-01-01'), (1, 2, '1999-01-01'),
(1, 3, '2000-01-01'), (1, 4, '2001-01-01'),
(1, 5, '2002-01-01'), (2, 1, '1998-01-01'),
(2, 2, '1999-01-01'), (2, 3, '2000-01-01'),
(2, 4, '2001-01-01'), (2, 5, '2002-01-01'),
(3, 1, '1998-01-01'), (3, 2, '1999-01-01'),
(3, 3, '2000-01-01'), (3, 4, '2001-01-01'),
(3, 5, '2002-01-01'), (4, 1, '1998-01-01'),
(4, 2, '1999-01-01'), (4, 3, '2000-01-01'),
(4, 4, '2001-01-01'), (4, 5, '2002-01-01'),
(5, 1, '1998-01-01'), (5, 2, '1999-01-01'),
(5, 3, '2000-01-01'), (5, 4, '2001-01-01'),
(5, 5, '2002-01-01');

Now consider this query:

EXPLAIN SELECT COUNT(*) FROM t1 WHERE i1 = 3 AND d = '2000-01-01'

The optimizer cannot use the primary key in this case because that comprises columns (i1, i2) and the query does not refer to i2. Instead, the optimizer can use the secondary index k_d on (d), and the execution plan depends on whether the extended index is used.

When the optimizer does not consider index extensions, it treats the index k_d as only (d). EXPLAIN for the query produces this result:

mysql> EXPLAIN SELECT COUNT(*) FROM t1 WHERE i1 = 3 AND d = '2000-01-01'\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: t1
         type: ref
possible_keys: PRIMARY,k_d
          key: k_d
      key_len: 4
          ref: const
         rows: 5
        Extra: Using where; Using index

When the optimizer takes index extensions into account, it treats k_d as (d, i1, i2). In this case, it can use the leftmost index prefix (d, i1) to produce a better execution plan:

mysql> EXPLAIN SELECT COUNT(*) FROM t1 WHERE i1 = 3 AND d = '2000-01-01'\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: t1
         type: ref
possible_keys: PRIMARY,k_d
          key: k_d
      key_len: 8
          ref: const,const
         rows: 1
        Extra: Using index

In both cases, key indicates that the optimizer will use secondary index k_d but the EXPLAIN output shows these improvements from using the extended index:

  • key_len goes from 4 bytes to 8 bytes, indicating that key lookups use columns d and i1, not just d.

  • The ref value changes from const to const,const because the key lookup uses two key parts, not one.

  • The rows count decreases from 5 to 1, indicating that InnoDB should need to examine fewer rows to produce the result.

  • The Extra value changes from Using where; Using index to Using index. This means that rows can be read using only the index, without consulting columns in the data row.

Differences in optimizer behavior for use of extended indexes can also be seen with SHOW STATUS:

FLUSH TABLE t1;
FLUSH STATUS;
SELECT COUNT(*) FROM t1 WHERE i1 = 3 AND d = '2000-01-01';
SHOW STATUS LIKE 'handler_read%'

The preceding statements include FLUSH TABLE and FLUSH STATUS to flush the table cache and clear the status counters.

Without index extensions, SHOW STATUS produces this result:

+-----------------------+-------+
| Variable_name         | Value |
+-----------------------+-------+
| Handler_read_first    | 0     |
| Handler_read_key      | 1     |
| Handler_read_last     | 0     |
| Handler_read_next     | 5     |
| Handler_read_prev     | 0     |
| Handler_read_rnd      | 0     |
| Handler_read_rnd_next | 0     |
+-----------------------+-------+

With index extensions, SHOW STATUS produces this result. The Handler_read_next value decreases from 5 to 1, indicating more efficient use of the index:

+-----------------------+-------+
| Variable_name         | Value |
+-----------------------+-------+
| Handler_read_first    | 0     |
| Handler_read_key      | 1     |
| Handler_read_last     | 0     |
| Handler_read_next     | 1     |
| Handler_read_prev     | 0     |
| Handler_read_rnd      | 0     |
| Handler_read_rnd_next | 0     |
+-----------------------+-------+

The use_index_extensions flag of the optimizer_switch system variable permits control over whether the optimizer takes the primary key columns into account when determining how to use an InnoDB table's secondary indexes. By default, use_index_extensions is enabled. To check whether disabling use of index extensions will improve performance, use this statement:

SET optimizer_switch = 'use_index_extensions=off';

Use of index extensions by the optimizer is subject to the usual limits on the number of key parts in an index (16) and the maximum key length (3072 bytes).

8.2.1.8 IS NULL Optimization

MySQL can perform the same optimization on col_name IS NULL that it can use for col_name = constant_value. For example, MySQL can use indexes and ranges to search for NULL with IS NULL.

Examples:

SELECT * FROM tbl_name WHERE key_col IS NULL;

SELECT * FROM tbl_name WHERE key_col <=> NULL;

SELECT * FROM tbl_name
  WHERE key_col=const1 OR key_col=const2 OR key_col IS NULL;

If a WHERE clause includes a col_name IS NULL condition for a column that is declared as NOT NULL, that expression is optimized away. This optimization does not occur in cases when the column might produce NULL anyway; for example, if it comes from a table on the right side of a LEFT JOIN.

MySQL can also optimize the combination col_name = expr OR col_name IS NULL, a form that is common in resolved subqueries. EXPLAIN shows ref_or_null when this optimization is used.

This optimization can handle one IS NULL for any key part.

Some examples of queries that are optimized, assuming that there is an index on columns a and b of table t2:

SELECT * FROM t1 WHERE t1.a=expr OR t1.a IS NULL;

SELECT * FROM t1, t2 WHERE t1.a=t2.a OR t2.a IS NULL;

SELECT * FROM t1, t2
  WHERE (t1.a=t2.a OR t2.a IS NULL) AND t2.b=t1.b;

SELECT * FROM t1, t2
  WHERE t1.a=t2.a AND (t2.b=t1.b OR t2.b IS NULL);

SELECT * FROM t1, t2
  WHERE (t1.a=t2.a AND t2.a IS NULL AND ...)
  OR (t1.a=t2.a AND t2.a IS NULL AND ...);

ref_or_null works by first doing a read on the reference key, and then a separate search for rows with a NULL key value.

The optimization can handle only one IS NULL level. In the following query, MySQL uses key lookups only on the expression (t1.a=t2.a AND t2.a IS NULL) and is not able to use the key part on b:

SELECT * FROM t1, t2
  WHERE (t1.a=t2.a AND t2.a IS NULL)
  OR (t1.b=t2.b AND t2.b IS NULL);

8.2.1.9 LEFT JOIN and RIGHT JOIN Optimization

MySQL implements an A LEFT JOIN B join_condition as follows:

  • Table B is set to depend on table A and all tables on which A depends.

  • Table A is set to depend on all tables (except B) that are used in the LEFT JOIN condition.

  • The LEFT JOIN condition is used to decide how to retrieve rows from table B. (In other words, any condition in the WHERE clause is not used.)

  • All standard join optimizations are performed, with the exception that a table is always read after all tables on which it depends. If there is a circular dependence, MySQL issues an error.

  • All standard WHERE optimizations are performed.

  • If there is a row in A that matches the WHERE clause, but there is no row in B that matches the ON condition, an extra B row is generated with all columns set to NULL.

  • If you use LEFT JOIN to find rows that do not exist in some table and you have the following test: col_name IS NULL in the WHERE part, where col_name is a column that is declared as NOT NULL, MySQL stops searching for more rows (for a particular key combination) after it has found one row that matches the LEFT JOIN condition.

The implementation of RIGHT JOIN is analogous to that of LEFT JOIN with the roles of the tables reversed.

The join optimizer calculates the order in which tables should be joined. The table read order forced by LEFT JOIN or STRAIGHT_JOIN helps the join optimizer do its work much more quickly, because there are fewer table permutations to check. Note that this means that if you do a query of the following type, MySQL does a full scan on b because the LEFT JOIN forces it to be read before d:

SELECT *
  FROM a JOIN b LEFT JOIN c ON (c.key=a.key)
  LEFT JOIN d ON (d.key=a.key)
  WHERE b.key=d.key;

The fix in this case is reverse the order in which a and b are listed in the FROM clause:

SELECT *
  FROM b JOIN a LEFT JOIN c ON (c.key=a.key)
  LEFT JOIN d ON (d.key=a.key)
  WHERE b.key=d.key;

For a LEFT JOIN, if the WHERE condition is always false for the generated NULL row, the LEFT JOIN is changed to a normal join. For example, the WHERE clause would be false in the following query if t2.column1 were NULL:

SELECT * FROM t1 LEFT JOIN t2 ON (column1) WHERE t2.column2=5;

Therefore, it is safe to convert the query to a normal join:

SELECT * FROM t1, t2 WHERE t2.column2=5 AND t1.column1=t2.column1;

This can be made faster because MySQL can use table t2 before table t1 if doing so would result in a better query plan. To provide a hint about the table join order, use STRAIGHT_JOIN. (See Section 13.2.9, “SELECT Syntax”.) (However, STRAIGHT_JOIN may prevent indexes from being used because it disables semi-join transformations. See Section 8.2.1.18.1, “Optimizing Subqueries with Semi-Join Transformations”.)

8.2.1.10 Nested-Loop Join Algorithms

MySQL executes joins between tables using a nested-loop algorithm or variations on it.

Nested-Loop Join Algorithm

A simple nested-loop join (NLJ) algorithm reads rows from the first table in a loop one at a time, passing each row to a nested loop that processes the next table in the join. This process is repeated as many times as there remain tables to be joined.

Assume that a join between three tables t1, t2, and t3 is to be executed using the following join types:

Table   Join Type
t1      range
t2      ref
t3      ALL

If a simple NLJ algorithm is used, the join is processed like this:

for each row in t1 matching range {
  for each row in t2 matching reference key {
    for each row in t3 {
      if row satisfies join conditions,
      send to client
    }
  }
}

Because the NLJ algorithm passes rows one at a time from outer loops to inner loops, it typically reads tables processed in the inner loops many times.

Block Nested-Loop Join Algorithm

A Block Nested-Loop (BNL) join algorithm uses buffering of rows read in outer loops to reduce the number of times that tables in inner loops must be read. For example, if 10 rows are read into a buffer and the buffer is passed to the next inner loop, each row read in the inner loop can be compared against all 10 rows in the buffer. The reduces the number of times the inner table must be read by an order of magnitude.

MySQL uses join buffering under these conditions:

  • The join_buffer_size system variable determines the size of each join buffer.

  • Join buffering can be used when the join is of type ALL or index (in other words, when no possible keys can be used, and a full scan is done, of either the data or index rows, respectively), or range. Use of buffering is also applicable to outer joins, as described in Section 8.2.1.14, “Block Nested-Loop and Batched Key Access Joins”.

  • One buffer is allocated for each join that can be buffered, so a given query might be processed using multiple join buffers.

  • A join buffer is never allocated for the first nonconst table, even if it would be of type ALL or index.

  • A join buffer is allocated prior to executing the join and freed after the query is done.

  • Only columns of interest to the join are stored in the join buffer, not whole rows.

For the example join described previously for the NLJ algorithm (without buffering), the join is done as follow using join buffering:

for each row in t1 matching range {
  for each row in t2 matching reference key {
    store used columns from t1, t2 in join buffer
    if buffer is full {
      for each row in t3 {
        for each t1, t2 combination in join buffer {
          if row satisfies join conditions,
          send to client
        }
      }
      empty buffer
    }
  }
}

if buffer is not empty {
  for each row in t3 {
    for each t1, t2 combination in join buffer {
      if row satisfies join conditions,
      send to client
    }
  }
}

If S is the size of each stored t1, t2 combination is the join buffer and C is the number of combinations in the buffer, the number of times table t3 is scanned is:

(S * C)/join_buffer_size + 1

The number of t3 scans decreases as the value of join_buffer_size increases, up to the point when join_buffer_size is large enough to hold all previous row combinations. At that point, there is no speed to be gained by making it larger.

8.2.1.11 Nested Join Optimization

The syntax for expressing joins permits nested joins. The following discussion refers to the join syntax described in Section 13.2.9.2, “JOIN Syntax”.

The syntax of table_factor is extended in comparison with the SQL Standard. The latter accepts only table_reference, not a list of them inside a pair of parentheses. This is a conservative extension if we consider each comma in a list of table_reference items as equivalent to an inner join. For example:

SELECT * FROM t1 LEFT JOIN (t2, t3, t4)
                 ON (t2.a=t1.a AND t3.b=t1.b AND t4.c=t1.c)

is equivalent to:

SELECT * FROM t1 LEFT JOIN (t2 CROSS JOIN t3 CROSS JOIN t4)
                 ON (t2.a=t1.a AND t3.b=t1.b AND t4.c=t1.c)

In MySQL, CROSS JOIN is a syntactic equivalent to INNER JOIN (they can replace each other). In standard SQL, they are not equivalent. INNER JOIN is used with an ON clause; CROSS JOIN is used otherwise.

In general, parentheses can be ignored in join expressions containing only inner join operations. After removing parentheses and grouping operations to the left, the join expression:

t1 LEFT JOIN (t2 LEFT JOIN t3 ON t2.b=t3.b OR t2.b IS NULL)
   ON t1.a=t2.a

transforms into the expression:

(t1 LEFT JOIN t2 ON t1.a=t2.a) LEFT JOIN t3
    ON t2.b=t3.b OR t2.b IS NULL

Yet, the two expressions are not equivalent. To see this, suppose that the tables t1, t2, and t3 have the following state:

  • Table t1 contains rows (1), (2)

  • Table t2 contains row (1,101)

  • Table t3 contains row (101)

In this case, the first expression returns a result set including the rows (1,1,101,101), (2,NULL,NULL,NULL), whereas the second expression returns the rows (1,1,101,101), (2,NULL,NULL,101):

mysql> SELECT *
    -> FROM t1
    ->      LEFT JOIN
    ->      (t2 LEFT JOIN t3 ON t2.b=t3.b OR t2.b IS NULL)
    ->      ON t1.a=t2.a;
+------+------+------+------+
| a    | a    | b    | b    |
+------+------+------+------+
|    1 |    1 |  101 |  101 |
|    2 | NULL | NULL | NULL |
+------+------+------+------+

mysql> SELECT *
    -> FROM (t1 LEFT JOIN t2 ON t1.a=t2.a)
    ->      LEFT JOIN t3
    ->      ON t2.b=t3.b OR t2.b IS NULL;
+------+------+------+------+
| a    | a    | b    | b    |
+------+------+------+------+
|    1 |    1 |  101 |  101 |
|    2 | NULL | NULL |  101 |
+------+------+------+------+

In the following example, an outer join operation is used together with an inner join operation:

t1 LEFT JOIN (t2, t3) ON t1.a=t2.a

That expression cannot be transformed into the following expression:

t1 LEFT JOIN t2 ON t1.a=t2.a, t3.

For the given table states, the two expressions return different sets of rows:

mysql> SELECT *
    -> FROM t1 LEFT JOIN (t2, t3) ON t1.a=t2.a;
+------+------+------+------+
| a    | a    | b    | b    |
+------+------+------+------+
|    1 |    1 |  101 |  101 |
|    2 | NULL | NULL | NULL |
+------+------+------+------+

mysql> SELECT *
    -> FROM t1 LEFT JOIN t2 ON t1.a=t2.a, t3;
+------+------+------+------+
| a    | a    | b    | b    |
+------+------+------+------+
|    1 |    1 |  101 |  101 |
|    2 | NULL | NULL |  101 |
+------+------+------+------+

Therefore, if we omit parentheses in a join expression with outer join operators, we might change the result set for the original expression.

More exactly, we cannot ignore parentheses in the right operand of the left outer join operation and in the left operand of a right join operation. In other words, we cannot ignore parentheses for the inner table expressions of outer join operations. Parentheses for the other operand (operand for the outer table) can be ignored.

The following expression:

(t1,t2) LEFT JOIN t3 ON P(t2.b,t3.b)

is equivalent to this expression:

t1, t2 LEFT JOIN t3 ON P(t2.b,t3.b)

for any tables t1,t2,t3 and any condition P over attributes t2.b and t3.b.

Whenever the order of execution of the join operations in a join expression (join_table) is not from left to right, we talk about nested joins. Consider the following queries:

SELECT * FROM t1 LEFT JOIN (t2 LEFT JOIN t3 ON t2.b=t3.b) ON t1.a=t2.a
  WHERE t1.a > 1

SELECT * FROM t1 LEFT JOIN (t2, t3) ON t1.a=t2.a
  WHERE (t2.b=t3.b OR t2.b IS NULL) AND t1.a > 1

Those queries are considered to contain these nested joins:

t2 LEFT JOIN t3 ON t2.b=t3.b
t2, t3

The nested join is formed in the first query with a left join operation, whereas in the second query it is formed with an inner join operation.

In the first query, the parentheses can be omitted: The grammatical structure of the join expression will dictate the same order of execution for join operations. For the second query, the parentheses cannot be omitted, although the join expression here can be interpreted unambiguously without them. (In our extended syntax the parentheses in (t2, t3) of the second query are required, although theoretically the query could be parsed without them: We still would have unambiguous syntactical structure for the query because LEFT JOIN and ON would play the role of the left and right delimiters for the expression (t2,t3).)

The preceding examples demonstrate these points:

  • For join expressions involving only inner joins (and not outer joins), parentheses can be removed. You can remove parentheses and evaluate left to right (or, in fact, you can evaluate the tables in any order).

  • The same is not true, in general, for outer joins or for outer joins mixed with inner joins. Removal of parentheses may change the result.

Queries with nested outer joins are executed in the same pipeline manner as queries with inner joins. More exactly, a variation of the nested-loop join algorithm is exploited. Recall by what algorithmic schema the nested-loop join executes a query. Suppose that we have a join query over 3 tables T1,T2,T3 of the form:

SELECT * FROM T1 INNER JOIN T2 ON P1(T1,T2)
                 INNER JOIN T3 ON P2(T2,T3)
  WHERE P(T1,T2,T3).

Here, P1(T1,T2) and P2(T3,T3) are some join conditions (on expressions), whereas P(T1,T2,T3) is a condition over columns of tables T1,T2,T3.

The nested-loop join algorithm would execute this query in the following manner:

FOR each row t1 in T1 {
  FOR each row t2 in T2 such that P1(t1,t2) {
    FOR each row t3 in T3 such that P2(t2,t3) {
      IF P(t1,t2,t3) {
         t:=t1||t2||t3; OUTPUT t;
      }
    }
  }
}

The notation t1||t2||t3 means a row constructed by concatenating the columns of rows t1, t2, and t3. In some of the following examples, NULL where a row name appears means that NULL is used for each column of that row. For example, t1||t2||NULL means a row constructed by concatenating the columns of rows t1 and t2, and NULL for each column of t3.

Now let's consider a query with nested outer joins:

SELECT * FROM T1 LEFT JOIN
              (T2 LEFT JOIN T3 ON P2(T2,T3))
              ON P1(T1,T2)
  WHERE P(T1,T2,T3).

For this query, we modify the nested-loop pattern to get:

FOR each row t1 in T1 {
  BOOL f1:=FALSE;
  FOR each row t2 in T2 such that P1(t1,t2) {
    BOOL f2:=FALSE;
    FOR each row t3 in T3 such that P2(t2,t3) {
      IF P(t1,t2,t3) {
        t:=t1||t2||t3; OUTPUT t;
      }
      f2=TRUE;
      f1=TRUE;
    }
    IF (!f2) {
      IF P(t1,t2,NULL) {
        t:=t1||t2||NULL; OUTPUT t;
      }
      f1=TRUE;
    }
  }
  IF (!f1) {
    IF P(t1,NULL,NULL) {
      t:=t1||NULL||NULL; OUTPUT t;
    }
  }
}

In general, for any nested loop for the first inner table in an outer join operation, a flag is introduced that is turned off before the loop and is checked after the loop. The flag is turned on when for the current row from the outer table a match from the table representing the inner operand is found. If at the end of the loop cycle the flag is still off, no match has been found for the current row of the outer table. In this case, the row is complemented by NULL values for the columns of the inner tables. The result row is passed to the final check for the output or into the next nested loop, but only if the row satisfies the join condition of all embedded outer joins.

In our example, the outer join table expressed by the following expression is embedded:

(T2 LEFT JOIN T3 ON P2(T2,T3))

For the query with inner joins, the optimizer could choose a different order of nested loops, such as this one:

FOR each row t3 in T3 {
  FOR each row t2 in T2 such that P2(t2,t3) {
    FOR each row t1 in T1 such that P1(t1,t2) {
      IF P(t1,t2,t3) {
         t:=t1||t2||t3; OUTPUT t;
      }
    }
  }
}

For the queries with outer joins, the optimizer can choose only such an order where loops for outer tables precede loops for inner tables. Thus, for our query with outer joins, only one nesting order is possible. For the following query, the optimizer will evaluate two different nestings:

SELECT * T1 LEFT JOIN (T2,T3) ON P1(T1,T2) AND P2(T1,T3)
  WHERE P(T1,T2,T3)

The nestings are these:

FOR each row t1 in T1 {
  BOOL f1:=FALSE;
  FOR each row t2 in T2 such that P1(t1,t2) {
    FOR each row t3 in T3 such that P2(t1,t3) {
      IF P(t1,t2,t3) {
        t:=t1||t2||t3; OUTPUT t;
      }
      f1:=TRUE
    }
  }
  IF (!f1) {
    IF P(t1,NULL,NULL) {
      t:=t1||NULL||NULL; OUTPUT t;
    }
  }
}

and:

FOR each row t1 in T1 {
  BOOL f1:=FALSE;
  FOR each row t3 in T3 such that P2(t1,t3) {
    FOR each row t2 in T2 such that P1(t1,t2) {
      IF P(t1,t2,t3) {
        t:=t1||t2||t3; OUTPUT t;
      }
      f1:=TRUE
    }
  }
  IF (!f1) {
    IF P(t1,NULL,NULL) {
      t:=t1||NULL||NULL; OUTPUT t;
    }
  }
}

In both nestings, T1 must be processed in the outer loop because it is used in an outer join. T2 and T3 are used in an inner join, so that join must be processed in the inner loop. However, because the join is an inner join, T2 and T3 can be processed in either order.

When discussing the nested-loop algorithm for inner joins, we omitted some details whose impact on the performance of query execution may be huge. We did not mention so-called pushed-down conditions. Suppose that our WHERE condition P(T1,T2,T3) can be represented by a conjunctive formula:

P(T1,T2,T2) = C1(T1) AND C2(T2) AND C3(T3).

In this case, MySQL actually uses the following nested-loop schema for the execution of the query with inner joins:

FOR each row t1 in T1 such that C1(t1) {
  FOR each row t2 in T2 such that P1(t1,t2) AND C2(t2)  {
    FOR each row t3 in T3 such that P2(t2,t3) AND C3(t3) {
      IF P(t1,t2,t3) {
         t:=t1||t2||t3; OUTPUT t;
      }
    }
  }
}

You see that each of the conjuncts C1(T1), C2(T2), C3(T3) are pushed out of the most inner loop to the most outer loop where it can be evaluated. If C1(T1) is a very restrictive condition, this condition pushdown may greatly reduce the number of rows from table T1 passed to the inner loops. As a result, the execution time for the query may improve immensely.

For a query with outer joins, the WHERE condition is to be checked only after it has been found that the current row from the outer table has a match in the inner tables. Thus, the optimization of pushing conditions out of the inner nested loops cannot be applied directly to queries with outer joins. Here we have to introduce conditional pushed-down predicates guarded by the flags that are turned on when a match has been encountered.

For our example with outer joins with:

P(T1,T2,T3)=C1(T1) AND C(T2) AND C3(T3)

the nested-loop schema using guarded pushed-down conditions looks like this:

FOR each row t1 in T1 such that C1(t1) {
  BOOL f1:=FALSE;
  FOR each row t2 in T2
      such that P1(t1,t2) AND (f1?C2(t2):TRUE) {
    BOOL f2:=FALSE;
    FOR each row t3 in T3
        such that P2(t2,t3) AND (f1&&f2?C3(t3):TRUE) {
      IF (f1&&f2?TRUE:(C2(t2) AND C3(t3))) {
        t:=t1||t2||t3; OUTPUT t;
      }
      f2=TRUE;
      f1=TRUE;
    }
    IF (!f2) {
      IF (f1?TRUE:C2(t2) && P(t1,t2,NULL)) {
        t:=t1||t2||NULL; OUTPUT t;
      }
      f1=TRUE;
    }
  }
  IF (!f1 && P(t1,NULL,NULL)) {
      t:=t1||NULL||NULL; OUTPUT t;
  }
}

In general, pushed-down predicates can be extracted from join conditions such as P1(T1,T2) and P(T2,T3). In this case, a pushed-down predicate is guarded also by a flag that prevents checking the predicate for the NULL-complemented row generated by the corresponding outer join operation.

Access by key from one inner table to another in the same nested join is prohibited if it is induced by a predicate from the WHERE condition. (We could use conditional key access in this case, but this technique is not employed yet in MySQL.)

8.2.1.12 Outer Join Simplification

Table expressions in the FROM clause of a query are simplified in many cases.

At the parser stage, queries with right outer joins operations are converted to equivalent queries containing only left join operations. In the general case, the conversion is performed according to the following rule:

(T1, ...) RIGHT JOIN (T2,...) ON P(T1,...,T2,...) =
(T2, ...) LEFT JOIN (T1,...) ON P(T1,...,T2,...)

All inner join expressions of the form T1 INNER JOIN T2 ON P(T1,T2) are replaced by the list T1,T2, P(T1,T2) being joined as a conjunct to the WHERE condition (or to the join condition of the embedding join, if there is any).

When the optimizer evaluates plans for join queries with outer join operation, it takes into consideration only the plans where, for each such operation, the outer tables are accessed before the inner tables. The optimizer options are limited because only such plans enables us to execute queries with outer joins operations by the nested loop schema.

Suppose that we have a query of the form:

SELECT * T1 LEFT JOIN T2 ON P1(T1,T2)
  WHERE P(T1,T2) AND R(T2)

with R(T2) narrowing greatly the number of matching rows from table T2. If we executed the query as it is, the optimizer would have no other choice besides to access table T1 before table T2 that may lead to a very inefficient execution plan.

Fortunately, MySQL converts such a query into a query without an outer join operation if the WHERE condition is null-rejected. A condition is called null-rejected for an outer join operation if it evaluates to FALSE or to UNKNOWN for any NULL-complemented row built for the operation.

Thus, for this outer join:

T1 LEFT JOIN T2 ON T1.A=T2.A

Conditions such as these are null-rejected:

T2.B IS NOT NULL,
T2.B > 3,
T2.C <= T1.C,
T2.B < 2 OR T2.C > 1

Conditions such as these are not null-rejected:

T2.B IS NULL,
T1.B < 3 OR T2.B IS NOT NULL,
T1.B < 3 OR T2.B > 3

The general rules for checking whether a condition is null-rejected for an outer join operation are simple. A condition is null-rejected in the following cases:

  • If it is of the form A IS NOT NULL, where A is an attribute of any of the inner tables

  • If it is a predicate containing a reference to an inner table that evaluates to UNKNOWN when one of its arguments is NULL

  • If it is a conjunction containing a null-rejected condition as a conjunct

  • If it is a disjunction of null-rejected conditions

A condition can be null-rejected for one outer join operation in a query and not null-rejected for another. In the query:

SELECT * FROM T1 LEFT JOIN T2 ON T2.A=T1.A
                 LEFT JOIN T3 ON T3.B=T1.B
  WHERE T3.C > 0

the WHERE condition is null-rejected for the second outer join operation but is not null-rejected for the first one.

If the WHERE condition is null-rejected for an outer join operation in a query, the outer join operation is replaced by an inner join operation.

For example, the preceding query is replaced with the query:

SELECT * FROM T1 LEFT JOIN T2 ON T2.A=T1.A
                 INNER JOIN T3 ON T3.B=T1.B
  WHERE T3.C > 0

For the original query, the optimizer would evaluate plans compatible with only one access order T1,T2,T3. For the replacing query, it additionally considers the access sequence T3,T1,T2.

A conversion of one outer join operation may trigger a conversion of another. Thus, the query:

SELECT * FROM T1 LEFT JOIN T2 ON T2.A=T1.A
                 LEFT JOIN T3 ON T3.B=T2.B
  WHERE T3.C > 0

will be first converted to the query:

SELECT * FROM T1 LEFT JOIN T2 ON T2.A=T1.A
                 INNER JOIN T3 ON T3.B=T2.B
  WHERE T3.C > 0

which is equivalent to the query:

SELECT * FROM (T1 LEFT JOIN T2 ON T2.A=T1.A), T3
  WHERE T3.C > 0 AND T3.B=T2.B

Now the remaining outer join operation can be replaced by an inner join, too, because the condition T3.B=T2.B is null-rejected and we get a query without outer joins at all:

SELECT * FROM (T1 INNER JOIN T2 ON T2.A=T1.A), T3
  WHERE T3.C > 0 AND T3.B=T2.B

Sometimes we succeed in replacing an embedded outer join operation, but cannot convert the embedding outer join. The following query:

SELECT * FROM T1 LEFT JOIN
              (T2 LEFT JOIN T3 ON T3.B=T2.B)
              ON T2.A=T1.A
  WHERE T3.C > 0

is converted to:

SELECT * FROM T1 LEFT JOIN
              (T2 INNER JOIN T3 ON T3.B=T2.B)
              ON T2.A=T1.A
  WHERE T3.C > 0,

That can be rewritten only to the form still containing the embedding outer join operation:

SELECT * FROM T1 LEFT JOIN
              (T2,T3)
              ON (T2.A=T1.A AND T3.B=T2.B)
  WHERE T3.C > 0.

When trying to convert an embedded outer join operation in a query, we must take into account the join condition for the embedding outer join together with the WHERE condition. In the query:

SELECT * FROM T1 LEFT JOIN
              (T2 LEFT JOIN T3 ON T3.B=T2.B)
              ON T2.A=T1.A AND T3.C=T1.C
  WHERE T3.D > 0 OR T1.D > 0

the WHERE condition is not null-rejected for the embedded outer join, but the join condition of the embedding outer join T2.A=T1.A AND T3.C=T1.C is null-rejected. So the query can be converted to:

SELECT * FROM T1 LEFT JOIN
              (T2, T3)
              ON T2.A=T1.A AND T3.C=T1.C AND T3.B=T2.B
  WHERE T3.D > 0 OR T1.D > 0

8.2.1.13 Multi-Range Read Optimization

Reading rows using a range scan on a secondary index can result in many random disk accesses to the base table when the table is large and not stored in the storage engine's cache. With the Disk-Sweep Multi-Range Read (MRR) optimization, MySQL tries to reduce the number of random disk access for range scans by first scanning the index only and collecting the keys for the relevant rows. Then the keys are sorted and finally the rows are retrieved from the base table using the order of the primary key. The motivation for Disk-sweep MRR is to reduce the number of random disk accesses and instead achieve a more sequential scan of the base table data.

The Multi-Range Read optimization provides these benefits:

  • MRR enables data rows to be accessed sequentially rather than in random order, based on index tuples. The server obtains a set of index tuples that satisfy the query conditions, sorts them according to data row ID order, and uses the sorted tuples to retrieve data rows in order. This makes data access more efficient and less expensive.

  • MRR enables batch processing of requests for key access for operations that require access to data rows through index tuples, such as range index scans and equi-joins that use an index for the join attribute. MRR iterates over a sequence of index ranges to obtain qualifying index tuples. As these results accumulate, they are used to access the corresponding data rows. It is not necessary to acquire all index tuples before starting to read data rows.

The MRR optimization is not supported with secondary indexes created on virtual generated columns. InnoDB supports secondary indexes on virtual generated columns as of MySQL 5.7.8.

The following scenarios illustrate when MRR optimization can be advantageous:

Scenario A: MRR can be used for InnoDB and MyISAM tables for index range scans and equi-join operations.

  1. A portion of the index tuples are accumulated in a buffer.

  2. The tuples in the buffer are sorted by their data row ID.

  3. Data rows are accessed according to the sorted index tuple sequence.

Scenario B: MRR can be used for NDB tables for multiple-range index scans or when performing an equi-join by an attribute.

  1. A portion of ranges, possibly single-key ranges, is accumulated in a buffer on the central node where the query is submitted.

  2. The ranges are sent to the execution nodes that access data rows.

  3. The accessed rows are packed into packages and sent back to the central node.

  4. The received packages with data rows are placed in a buffer.

  5. Data rows are read from the buffer.

When MRR is used, the Extra column in EXPLAIN output shows Using MRR.

InnoDB and MyISAM do not use MRR if full table rows need not be accessed to produce the query result. This is the case if results can be produced entirely on the basis on information in the index tuples (through a covering index); MRR provides no benefit.

Example query for which MRR can be used, assuming that there is an index on (key_part1, key_part2):

SELECT * FROM t
  WHERE key_part1 >= 1000 AND key_part1 < 2000
  AND key_part2 = 10000;

The index consists of tuples of (key_part1, key_part2) values, ordered first by key_part1 and then by key_part2.

Without MRR, an index scan covers all index tuples for the key_part1 range from 1000 up to 2000, regardless of the key_part2 value in these tuples. The scan does extra work to the extent that tuples in the range contain key_part2 values other than 10000.

With MRR, the scan is broken up into multiple ranges, each for a single value of key_part1 (1000, 1001, ... , 1999). Each of these scans need look only for tuples with key_part2 = 10000. If the index contains many tuples for which key_part2 is not 10000, MRR results in many fewer index tuples being read.

To express this using interval notation, the non-MRR scan must examine the index range [{1000, 10000}, {2000, MIN_INT}), which may include many tuples other than those for which key_part2 = 10000. The MRR scan examines multiple single-point intervals [{1000, 10000}], ..., [{1999, 10000}], which includes only tuples with key_part2 = 10000.

Two optimizer_switch system variable flags provide an interface to the use of MRR optimization. The mrr flag controls whether MRR is enabled. If mrr is enabled (on), the mrr_cost_based flag controls whether the optimizer attempts to make a cost-based choice between using and not using MRR (on) or uses MRR whenever possible (off). By default, mrr is on and mrr_cost_based is on. See Section 8.9.2, “Controlling Switchable Optimizations”.

For MRR, a storage engine uses the value of the read_rnd_buffer_size system variable as a guideline for how much memory it can allocate for its buffer. The engine uses up to read_rnd_buffer_size bytes and determines the number of ranges to process in a single pass.

8.2.1.14 Block Nested-Loop and Batched Key Access Joins

In MySQL, a Batched Key Access (BKA) Join algorithm is available that uses both index access to the joined table and a join buffer. The BKA algorithm supports inner join, outer join, and semi-join operations, including nested outer joins. Benefits of BKA include improved join performance due to more efficient table scanning. Also, the Block Nested-Loop (BNL) Join algorithm previously used only for inner joins is extended and can be employed for outer join and semi-join operations, including nested outer joins.

The following sections discuss the join buffer management that underlies the extension of the original BNL algorithm, the extended BNL algorithm, and the BKA algorithm. For information about semi-join strategies, see Section 8.2.1.18.1, “Optimizing Subqueries with Semi-Join Transformations”

8.2.1.14.1 Join Buffer Management for Block Nested-Loop and Batched Key Access Algorithms

MySQL Server can employ join buffers to execute not only inner joins without index access to the inner table, but also outer joins and semi-joins that appear after subquery flattening. Moreover, a join buffer can be effectively used when there is an index access to the inner table.

The join buffer management code slightly more efficiently utilizes join buffer space when storing the values of the interesting row columns: No additional bytes are allocated in buffers for a row column if its value is NULL, and the minimum number of bytes is allocated for any value of the VARCHAR type.

The code supports two types of buffers, regular and incremental. Suppose that join buffer B1 is employed to join tables t1 and t2 and the result of this operation is joined with table t3 using join buffer B2:

  • A regular join buffer contains columns from each join operand. If B2 is a regular join buffer, each row r put into B2 is composed of the columns of a row r1 from B1 and the interesting columns of a matching row r2 from table t3.

  • An incremental join buffer contains only columns from rows of the table produced by the second join operand. That is, it is incremental to a row from the first operand buffer. If B2 is an incremental join buffer, it contains the interesting columns of the row r2 together with a link to the row r1 from B1.

Incremental join buffers are always incremental relative to a join buffer from an earlier join operation, so the buffer from the first join operation is always a regular buffer. In the example just given, the buffer B1 used to join tables t1 and t2 must be a regular buffer.

Each row of the incremental buffer used for a join operation contains only the interesting columns of a row from the table to be joined. These columns are augmented with a reference to the interesting columns of the matched row from the table produced by the first join operand. Several rows in the incremental buffer can refer to the same row r whose columns are stored in the previous join buffers insofar as all these rows match row r.

Incremental buffers enable less frequent copying of columns from buffers used for previous join operations. This provides a savings in buffer space because in the general case a row produced by the first join operand can be matched by several rows produced by the second join operand. It is unnecessary to make several copies of a row from the first operand. Incremental buffers also provide a savings in processing time due to the reduction in copying time.

The block_nested_loop and batched_key_access flags of the optimizer_switch system variable control how the optimizer uses the Block Nested-Loop and Batched Key Access join algorithms. By default, block_nested_loop is on and batched_key_access is off. See Section 8.9.2, “Controlling Switchable Optimizations”.

For information about semi-join strategies, see Section 8.2.1.18.1, “Optimizing Subqueries with Semi-Join Transformations”

8.2.1.14.2 Block Nested-Loop Algorithm for Outer Joins and Semi-Joins

The original implementation of the MySQL BNL algorithm is extended to support outer join and semi-join operations.

When these operations are executed with a join buffer, each row put into the buffer is supplied with a match flag.

If an outer join operation is executed using a join buffer, each row of the table produced by the second operand is checked for a match against each row in the join buffer. When a match is found, a new extended row is formed (the original row plus columns from the second operand) and sent for further extensions by the remaining join operations. In addition, the match flag of the matched row in the buffer is enabled. After all rows of the table to be joined have been examined, the join buffer is scanned. Each row from the buffer that does not have its match flag enabled is extended by NULL complements (NULL values for each column in the second operand) and sent for further extensions by the remaining join operations.

The block_nested_loop flag of the optimizer_switch system variable controls how the optimizer uses the Block Nested-Loop algorithm. By default, block_nested_loop is on. See Section 8.9.2, “Controlling Switchable Optimizations”.

In EXPLAIN output, use of BNL for a table is signified when the Extra value contains Using join buffer (Block Nested Loop) and the type value is ALL, index, or range.

For information about semi-join strategies, see Section 8.2.1.18.1, “Optimizing Subqueries with Semi-Join Transformations”

8.2.1.14.3 Batched Key Access Joins

MySQL Server implements a method of joining tables called the Batched Key Access (BKA) join algorithm. BKA can be applied when there is an index access to the table produced by the second join operand. Like the BNL join algorithm, the BKA join algorithm employs a join buffer to accumulate the interesting columns of the rows produced by the first operand of the join operation. Then the BKA algorithm builds keys to access the table to be joined for all rows in the buffer and submits these keys in a batch to the database engine for index lookups. The keys are submitted to the engine through the Multi-Range Read (MRR) interface (see Section 8.2.1.13, “Multi-Range Read Optimization”). After submission of the keys, the MRR engine functions perform lookups in the index in an optimal way, fetching the rows of the joined table found by these keys, and starts feeding the BKA join algorithm with matching rows. Each matching row is coupled with a reference to a row in the join buffer.

When BKA is used, the value of join_buffer_size defines how large the batch of keys is in each request to the storage engine. The larger the buffer, the more sequential access will be to the right hand table of a join operation, which can significantly improve performance.

For BKA to be used, the batched_key_access flag of the optimizer_switch system variable must be set to on. BKA uses MRR, so the mrr flag must also be on. Currently, the cost estimation for MRR is too pessimistic. Hence, it is also necessary for mrr_cost_based to be off for BKA to be used. The following setting enables BKA:

mysql> SET optimizer_switch='mrr=on,mrr_cost_based=off,batched_key_access=on';

There are two scenarios by which MRR functions execute:

  • The first scenario is used for conventional disk-based storage engines such as InnoDB and MyISAM. For these engines, usually the keys for all rows from the join buffer are submitted to the MRR interface at once. Engine-specific MRR functions perform index lookups for the submitted keys, get row IDs (or primary keys) from them, and then fetch rows for all these selected row IDs one by one by request from BKA algorithm. Every row is returned with an association reference that enables access to the matched row in the join buffer. The rows are fetched by the MRR functions in an optimal way: They are fetched in the row ID (primary key) order. This improves performance because reads are in disk order rather than random order.

  • The second scenario is used for remote storage engines such as NDB. A package of keys for a portion of rows from the join buffer, together with their associations, is sent by a MySQL Server (SQL node) to MySQL Cluster data nodes. In return, the SQL node receives a package (or several packages) of matching rows coupled with corresponding associations. The BKA join algorithm takes these rows and builds new joined rows. Then a new set of keys is sent to the data nodes and the rows from the returned packages are used to build new joined rows. The process continues until the last keys from the join buffer are sent to the data nodes, and the SQL node has received and joined all rows matching these keys. This improves performance because fewer key-bearing packages sent by the SQL node to the data nodes means fewer round trips between it and the data nodes to perform the join operation.

With the first scenario, a portion of the join buffer is reserved to store row IDs (primary keys) selected by index lookups and passed as a parameter to the MRR functions.

There is no special buffer to store keys built for rows from the join buffer. Instead, a function that builds the key for the next row in the buffer is passed as a parameter to the MRR functions.

In EXPLAIN output, use of BKA for a table is signified when the Extra value contains Using join buffer (Batched Key Access) and the type value is ref or eq_ref.

8.2.1.15 ORDER BY Optimization

In some cases, MySQL can use an index to satisfy an ORDER BY clause without doing extra sorting.

The index can also be used even if the ORDER BY does not match the index exactly, as long as all unused portions of the index and all extra ORDER BY columns are constants in the WHERE clause. The following queries use the index to resolve the ORDER BY part:

SELECT * FROM t1
  ORDER BY key_part1,key_part2,... ;

SELECT * FROM t1
  WHERE key_part1 = constant
  ORDER BY key_part2;

SELECT * FROM t1
  ORDER BY key_part1 DESC, key_part2 DESC;

SELECT * FROM t1
  WHERE key_part1 = 1
  ORDER BY key_part1 DESC, key_part2 DESC;

SELECT * FROM t1
  WHERE key_part1 > constant
  ORDER BY key_part1 ASC;

SELECT * FROM t1
  WHERE key_part1 < constant
  ORDER BY key_part1 DESC;

SELECT * FROM t1
  WHERE key_part1 = constant1 AND key_part2 > constant2
  ORDER BY key_part2;

In some cases, MySQL cannot use indexes to resolve the ORDER BY, although it still uses indexes to find the rows that match the WHERE clause. These cases include the following:

  • The query uses ORDER BY on different indexes:

    SELECT * FROM t1 ORDER BY key1, key2;
    
  • The query uses ORDER BY on nonconsecutive parts of an index:

    SELECT * FROM t1 WHERE key2=constant ORDER BY key_part2;
    
  • The query mixes ASC and DESC:

    SELECT * FROM t1 ORDER BY key_part1 DESC, key_part2 ASC;
    
  • The index used to fetch the rows differs from the one used in the ORDER BY:

    SELECT * FROM t1 WHERE key2=constant ORDER BY key1;
    
  • The query uses ORDER BY with an expression that includes terms other than the index column name:

    SELECT * FROM t1 ORDER BY ABS(key);
    SELECT * FROM t1 ORDER BY -key;
    
  • The query joins many tables, and the columns in the ORDER BY are not all from the first nonconstant table that is used to retrieve rows. (This is the first table in the EXPLAIN output that does not have a const join type.)

  • The query has different ORDER BY and GROUP BY expressions.

  • There is an index on only a prefix of a column named in the ORDER BY clause. In this case, the index cannot be used to fully resolve the sort order. For example, if only the first 10 bytes of a CHAR(20) column are indexed, the index cannot distinguish values past the 10th byte and a filesort will be needed.

  • The index does not store rows in order. For example, this is true for a HASH index in a MEMORY table.

Availability of an index for sorting may be affected by the use of column aliases. Suppose that the column t1.a is indexed. In this statement, the name of the column in the select list is a. It refers to t1.a, so for the reference to a in the ORDER BY, the index can be used:

SELECT a FROM t1 ORDER BY a;

In this statement, the name of the column in the select list is also a, but it is the alias name. It refers to ABS(a), so for the reference to a in the ORDER BY, the index cannot be used:

SELECT ABS(a) AS a FROM t1 ORDER BY a;

In the following statement, the ORDER BY refers to a name that is not the name of a column in the select list. But there is a column in t1 named a, so the ORDER BY uses that and the index can be used. (The resulting sort order may be completely different from the order for ABS(a), of course.)

SELECT ABS(a) AS b FROM t1 ORDER BY a;

By default, MySQL sorts all GROUP BY col1, col2, ... queries as if you specified ORDER BY col1, col2, ... in the query as well. If you include an explicit ORDER BY clause that contains the same column list, MySQL optimizes it away without any speed penalty, although the sorting still occurs.

Note

Relying on implicit GROUP BY sorting is deprecated. To achieve a specific sort order of grouped results, it is preferable to use an explicit ORDER BY clause. GROUP BY sorting is a MySQL extension that may change in a future release; for example, to make it possible for the optimizer to order groupings in whatever manner it deems most efficient and to avoid the sorting overhead.

If a query includes GROUP BY but you want to avoid the overhead of sorting the result, you can suppress sorting by specifying ORDER BY NULL. For example:

INSERT INTO foo
SELECT a, COUNT(*) FROM bar GROUP BY a ORDER BY NULL;

The optimizer may still choose to use sorting to implement grouping operations. ORDER BY NULL suppresses sorting of the result, not prior sorting done by grouping operations to determine the result.

With EXPLAIN SELECT ... ORDER BY, you can check whether MySQL can use indexes to resolve the query. It cannot if you see Using filesort in the Extra column. See Section 8.8.1, “Optimizing Queries with EXPLAIN”. Filesort uses a fixed-length row-storage format similar to that used by the MEMORY storage engine. Variable-length types such as VARCHAR are stored using a fixed length.

MySQL has two filesort algorithms for sorting and retrieving results. The original method uses only the ORDER BY columns. The modified method uses not just the ORDER BY columns, but all the columns referenced by the query.

The optimizer selects which filesort algorithm to use. It normally uses the modified algorithm except when BLOB or TEXT columns are involved, in which case it uses the original algorithm. For both algorithms, the sort buffer size is the sort_buffer_size system variable value.

The original filesort algorithm works as follows:

  1. Read all rows according to key or by table scanning. Skip rows that do not match the WHERE clause.

  2. For each row, store in the sort buffer a tuple consisting of a pair of values (the sort key value and the row ID).

  3. If all pairs fit into the sort buffer, no temporary file is created. Otherwise, when the sort buffer becomes full, run a qsort (quicksort) on it in memory and write it to a temporary file. Save a pointer to the sorted block.

  4. Repeat the preceding steps until all rows have been read.

  5. Do a multi-merge of up to MERGEBUFF (7) regions to one block in another temporary file. Repeat until all blocks from the first file are in the second file.

  6. Repeat the following until there are fewer than MERGEBUFF2 (15) blocks left.

  7. On the last multi-merge, only the row ID (the last part of the value pair) is written to a result file.

  8. Read the rows in sorted order using the row IDs in the result file. To optimize this, read in a large block of row IDs, sort them, and use them to read the rows in sorted order into a row buffer. The row buffer size is the read_rnd_buffer_size system variable value. The code for this step is in the sql/records.cc source file.

One problem with this approach is that it reads rows twice: One time during WHERE clause evaluation, and again after sorting the value pairs. And even if the rows were accessed successively the first time (for example, if a table scan is done), the second time they are accessed randomly. (The sort keys are ordered, but the row positions are not.)

The modified filesort algorithm incorporates an optimization to avoid reading the rows twice: It records the sort key value, but instead of the row ID, it records the columns referenced by the query. The modified filesort algorithm works like this:

  1. Read the rows that match the WHERE clause.

  2. For each row, store in the sort buffer a tuple consisting of the sort key value and the columns referenced by the query.

  3. When the sort buffer becomes full, sort the tuples by sort key value in memory and write it to a temporary file.

  4. After merge-sorting the temporary file, retrieve the rows in sorted order, but read the columns required by the query directly from the sorted tuples rather than by accessing the table a second time.

The tuples used by the modified filesort algorithm are longer than the pairs used by the original algorithm, and fewer of them fit in the sort buffer. As a result, it is possible for the extra I/O to make the modified approach slower, not faster. To avoid a slowdown, the optimizer uses the modified algorithm only if the total size of the extra columns in the sort tuple does not exceed the value of the max_length_for_sort_data system variable. (A symptom of setting the value of this variable too high is a combination of high disk activity and low CPU activity.)

As of MySQL 5.7.3, the modified filesort algorithm includes an additional optimization designed to enable more tuples to fit into the sort buffer: For additional columns of type CHAR or VARCHAR, or any nullable fixed-size data type, the values are packed. For example, without packing, a VARCHAR(255) column value containing only 3 characters takes 255 characters in the sort buffer. With packing, the value requires only 3 characters plus a two-byte length indicator. NULL values require only a bitmask.

For data containing packable strings shorter than the maximum column length or many NULL values, more records fit into the sort buffer. This improves in-memory sorting of the sort buffer and performance of disk-based temporary file merge sorting.

In edge cases, packing may be disadvantageous: If packable strings are the maximum column length or there are few NULL values, the space required for the length indicators reduces the number of records that fit into the sort buffer and sorting is slower in memory and on disk.

If a filesort is done, EXPLAIN output includes Using filesort in the Extra column. Also, optimizer trace output includes a filesort_summary block. For example:

"filesort_summary": {
  "rows": 100,
  "examined_rows": 100,
  "number_of_tmp_files": 0,
  "sort_buffer_size": 25192,
  "sort_mode": "<sort_key, packed_additional_fields>"
}

The sort_mode value provides information about the filesort algorithm used and the contents of tuples in the sort buffer:

  • <sort_key, rowid>: This indicates use of the original algorithm. Sort buffer tuples are pairs that contain the sort key value and row ID of the original table row. Tuples are sorted by sort key value and the row ID is used to read the row from the table.

  • <sort_key, additional_fields>: This indicates use of the modified algorithm. Sort buffer tuples contain the sort key value and columns referenced by the query. Tuples are sorted by sort key value and column values are read directly from the tuple.

  • <sort_key, packed_additional_fields>: This indicates use of the modified algorithm. Sort buffer tuples contain the sort key value and packed columns referenced by the query. Tuples are sorted by sort key value and column values are read directly from the tuple.

For information about the optimizer trace, see MySQL Internals: Tracing the Optimizer.

Suppose that a table t1 has four VARCHAR columns a, b, c, and d and that the optimizer uses filesort for this query:

SELECT * FROM t1 ORDER BY a, b;

The query sorts by a and b, but returns all columns, so the columns referenced by the query are a, b, c, and d. Depending on which filesort algorithm the optimizer chooses, the query executes as follows:

For the original algorithm, sort buffer tuples have these contents:

(fixed size a value, fixed size b value,
row ID into t1)

The optimizer sorts on the fixed size values. After sorting, the optimizer reads the tuples in order and uses the row ID in each tuple to read rows from t1 to obtain the select list column values.

For the modified algorithm without packing, sort buffer tuples have these contents:

(fixed size a value, fixed size b value,
a value, b value, c value, d value)

The optimizer sorts on the fixed size values. After sorting, the optimizer reads the tuples in order and uses the values for a, b, c, and d to obtain the select list column values without reading t1 again.

For the modified algorithm with packing, sort buffer tuples have these contents:

(fixed size a value, fixed size b value,
a length, packed a value, b length, packed b value,
c length, packed c value, d length, packed d value)

If any of a, b, c, or d are NULL, they take no space in the sort buffer other than in the bitmask.

The optimizer sorts on the fixed size values. After sorting, the optimizer reads the tuples in order and uses the values for a, b, c, and d to obtain the select list column values without reading t1 again.

For slow queries for which filesort is not used, try lowering max_length_for_sort_data to a value that is appropriate to trigger a filesort.

To increase ORDER BY speed, check whether you can get MySQL to use indexes rather than an extra sorting phase. If this is not possible, you can try the following strategies:

  • Increase the sort_buffer_size variable value. Ideally, the value should be large enough for the entire result set to fit in the sort buffer (to avoid writes to disk and merge passes), but at minimum the value must be large enough to accommodate fifteen tuples.

    Take into account that the size of column values stored in the sort buffer is affected by the max_sort_length system variable value. For example, if tuples store values of long string columns and you increase the value of max_sort_length, the size of sort buffer tuples increases as well and may require you to increase sort_buffer_size. For column values calculated as a result of string expressions (such as those that invoke a string-valued function), the filesort algorithm cannot tell the maximum length of expression values, so it must allocate max_sort_length bytes for each tuple.

    To monitor the number of merge passes, check the Sort_merge_passes status variable.

  • Increase the read_rnd_buffer_size variable value.

  • Use less RAM per row by declaring columns only as large as they need to be to hold the values stored in them. For example, CHAR(16) is better than CHAR(200) if values never exceed 16 characters.

  • Change the tmpdir system variable to point to a dedicated file system with large amounts of free space. The variable value can list several paths that are used in round-robin fashion; you can use this feature to spread the load across several directories. Paths should be separated by colon characters (:) on Unix and semicolon characters (;) on Windows. The paths should name directories in file systems located on different physical disks, not different partitions on the same disk.

If an index is not used for ORDER BY but a LIMIT clause is also present, the optimizer may be able to avoid using a merge file and sort the rows in memory. For details, see Section 8.2.1.19, “LIMIT Query Optimization”.

8.2.1.16 GROUP BY Optimization

The most general way to satisfy a GROUP BY clause is to scan the whole table and create a new temporary table where all rows from each group are consecutive, and then use this temporary table to discover groups and apply aggregate functions (if any). In some cases, MySQL is able to do much better than that and to avoid creation of temporary tables by using index access.

The most important preconditions for using indexes for GROUP BY are that all GROUP BY columns reference attributes from the same index, and that the index stores its keys in order (for example, this is a BTREE index and not a HASH index). Whether use of temporary tables can be replaced by index access also depends on which parts of an index are used in a query, the conditions specified for these parts, and the selected aggregate functions.

There are two ways to execute a GROUP BY query through index access, as detailed in the following sections. In the first method, the grouping operation is applied together with all range predicates (if any). The second method first performs a range scan, and then groups the resulting tuples.

In MySQL, GROUP BY is used for sorting, so the server may also apply ORDER BY optimizations to grouping. See Section 8.2.1.15, “ORDER BY Optimization”.

8.2.1.16.1 Loose Index Scan

The most efficient way to process GROUP BY is when an index is used to directly retrieve the grouping columns. With this access method, MySQL uses the property of some index types that the keys are ordered (for example, BTREE). This property enables use of lookup groups in an index without having to consider all keys in the index that satisfy all WHERE conditions. This access method considers only a fraction of the keys in an index, so it is called a loose index scan. When there is no WHERE clause, a loose index scan reads as many keys as the number of groups, which may be a much smaller number than that of all keys. If the WHERE clause contains range predicates (see the discussion of the range join type in Section 8.8.1, “Optimizing Queries with EXPLAIN”), a loose index scan looks up the first key of each group that satisfies the range conditions, and again reads the least possible number of keys. This is possible under the following conditions:

  • The query is over a single table.

  • The GROUP BY names only columns that form a leftmost prefix of the index and no other columns. (If, instead of GROUP BY, the query has a DISTINCT clause, all distinct attributes refer to columns that form a leftmost prefix of the index.) For example, if a table t1 has an index on (c1,c2,c3), loose index scan is applicable if the query has GROUP BY c1, c2,. It is not applicable if the query has GROUP BY c2, c3 (the columns are not a leftmost prefix) or GROUP BY c1, c2, c4 (c4 is not in the index).

  • The only aggregate functions used in the select list (if any) are MIN() and MAX(), and all of them refer to the same column. The column must be in the index and must follow the columns in the GROUP BY.

  • Any other parts of the index than those from the GROUP BY referenced in the query must be constants (that is, they must be referenced in equalities with constants), except for the argument of MIN() or MAX() functions.

  • For columns in the index, full column values must be indexed, not just a prefix. For example, with c1 VARCHAR(20), INDEX (c1(10)), the index cannot be used for loose index scan.

If loose index scan is applicable to a query, the EXPLAIN output shows Using index for group-by in the Extra column.

Assume that there is an index idx(c1,c2,c3) on table t1(c1,c2,c3,c4). The loose index scan access method can be used for the following queries:

SELECT c1, c2 FROM t1 GROUP BY c1, c2;
SELECT DISTINCT c1, c2 FROM t1;
SELECT c1, MIN(c2) FROM t1 GROUP BY c1;
SELECT c1, c2 FROM t1 WHERE c1 < const GROUP BY c1, c2;
SELECT MAX(c3), MIN(c3), c1, c2 FROM t1 WHERE c2 > const GROUP BY c1, c2;
SELECT c2 FROM t1 WHERE c1 < const GROUP BY c1, c2;
SELECT c1, c2 FROM t1 WHERE c3 = const GROUP BY c1, c2;

The following queries cannot be executed with this quick select method, for the reasons given:

  • There are aggregate functions other than MIN() or MAX():

    SELECT c1, SUM(c2) FROM t1 GROUP BY c1;
    
  • The columns in the GROUP BY clause do not form a leftmost prefix of the index:

    SELECT c1, c2 FROM t1 GROUP BY c2, c3;
    
  • The query refers to a part of a key that comes after the GROUP BY part, and for which there is no equality with a constant:

    SELECT c1, c3 FROM t1 GROUP BY c1, c2;
    

    Were the query to include WHERE c3 = const, loose index scan could be used.

The loose index scan access method can be applied to other forms of aggregate function references in the select list, in addition to the MIN() and MAX() references already supported:

Assume that there is an index idx(c1,c2,c3) on table t1(c1,c2,c3,c4). The loose index scan access method can be used for the following queries:

SELECT COUNT(DISTINCT c1), SUM(DISTINCT c1) FROM t1;

SELECT COUNT(DISTINCT c1, c2), COUNT(DISTINCT c2, c1) FROM t1;

Loose index scan is not applicable for the following queries:

SELECT DISTINCT COUNT(DISTINCT c1) FROM t1;

SELECT COUNT(DISTINCT c1) FROM t1 GROUP BY c1;
8.2.1.16.2 Tight Index Scan

A tight index scan may be either a full index scan or a range index scan, depending on the query conditions.

When the conditions for a loose index scan are not met, it still may be possible to avoid creation of temporary tables for GROUP BY queries. If there are range conditions in the WHERE clause, this method reads only the keys that satisfy these conditions. Otherwise, it performs an index scan. Because this method reads all keys in each range defined by the WHERE clause, or scans the whole index if there are no range conditions, we term it a tight index scan. With a tight index scan, the grouping operation is performed only after all keys that satisfy the range conditions have been found.

For this method to work, it is sufficient that there is a constant equality condition for all columns in a query referring to parts of the key coming before or in between parts of the GROUP BY key. The constants from the equality conditions fill in any gaps in the search keys so that it is possible to form complete prefixes of the index. These index prefixes then can be used for index lookups. If we require sorting of the GROUP BY result, and it is possible to form search keys that are prefixes of the index, MySQL also avoids extra sorting operations because searching with prefixes in an ordered index already retrieves all the keys in order.

Assume that there is an index idx(c1,c2,c3) on table t1(c1,c2,c3,c4). The following queries do not work with the loose index scan access method described earlier, but still work with the tight index scan access method.

  • There is a gap in the GROUP BY, but it is covered by the condition c2 = 'a':

    SELECT c1, c2, c3 FROM t1 WHERE c2 = 'a' GROUP BY c1, c3;
    
  • The GROUP BY does not begin with the first part of the key, but there is a condition that provides a constant for that part:

    SELECT c1, c2, c3 FROM t1 WHERE c1 = 'a' GROUP BY c2, c3;
    

8.2.1.17 DISTINCT Optimization

DISTINCT combined with ORDER BY needs a temporary table in many cases.

Because DISTINCT may use GROUP BY, learn how MySQL works with columns in ORDER BY or HAVING clauses that are not part of the selected columns. See Section 12.20.3, “MySQL Handling of GROUP BY”.

In most cases, a DISTINCT clause can be considered as a special case of GROUP BY. For example, the following two queries are equivalent:

SELECT DISTINCT c1, c2, c3 FROM t1
WHERE c1 > const;

SELECT c1, c2, c3 FROM t1
WHERE c1 > const GROUP BY c1, c2, c3;

Due to this equivalence, the optimizations applicable to GROUP BY queries can be also applied to queries with a DISTINCT clause. Thus, for more details on the optimization possibilities for DISTINCT queries, see Section 8.2.1.16, “GROUP BY Optimization”.

When combining LIMIT row_count with DISTINCT, MySQL stops as soon as it finds row_count unique rows.

If you do not use columns from all tables named in a query, MySQL stops scanning any unused tables as soon as it finds the first match. In the following case, assuming that t1 is used before t2 (which you can check with EXPLAIN), MySQL stops reading from t2 (for any particular row in t1) when it finds the first row in t2:

SELECT DISTINCT t1.a FROM t1, t2 where t1.a=t2.a;

8.2.1.18 Subquery Optimization

The MySQL query optimizer has different strategies available to evaluate subqueries. For IN (or =ANY) subqueries, the optimizer has these choices:

  • Semi-join

  • Materialization

  • EXISTS strategy

For NOT IN (or <>ALL) subqueries, the optimizer has these choices:

  • Materialization

  • EXISTS strategy

For derived tables (subqueries in the FROM clause) and view references, the optimizer has these choices:

  • Merge the derived table or view into the outer query block

  • Materialize the derived table or view to an internal temporary table

The following discussion provides more information about these optimization strategies.

Note

A limitation on UPDATE and DELETE statements that use a subquery to modify a single table is that the optimizer does not use semi-join or materialization subquery optimizations. As a workaround, try rewriting them as multiple-table UPDATE and DELETE statements that use a join rather than a subquery.

8.2.1.18.1 Optimizing Subqueries with Semi-Join Transformations

The optimizer uses semi-join strategies to improve subquery execution, as described in this section.

For an inner join between two tables, the join returns a row from one table as many times as there are matches in the other table. But for some questions, the only information that matters is whether there is a match, not the number of matches. Suppose that there are tables named class and roster that list classes in a course curriculum and class rosters (students enrolled in each class), respectively. To list the classes that actually have students enrolled, you could use this join:

SELECT class.class_num, class.class_name
FROM class INNER JOIN roster
WHERE class.class_num = roster.class_num;

However, the result lists each class once for each enrolled student. For the question being asked, this is unnecessary duplication of information.

Assuming that class_num is a primary key in the class table, duplicate suppression could be achieved by using SELECT DISTINCT, but it is inefficient to generate all matching rows first only to eliminate duplicates later.

The same duplicate-free result can be obtained by using a subquery:

SELECT class_num, class_name
FROM class
WHERE class_num IN (SELECT class_num FROM roster);

Here, the optimizer can recognize that the IN clause requires the subquery to return only one instance of each class number from the roster table. In this case, the query can be executed as a semi-join—that is, an operation that returns only one instance of each row in class that is matched by rows in roster.

Outer join and inner join syntax is permitted in the outer query specification, and table references may be base tables or views.

In MySQL, a subquery must satisfy these criteria to be handled as a semi-join:

  • It must be an IN (or =ANY) subquery that appears at the top level of the WHERE or ON clause, possibly as a term in an AND expression. For example:

    SELECT ...
    FROM ot1, ...
    WHERE (oe1, ...) IN (SELECT ie1, ... FROM it1, ... WHERE ...);
    

    Here, ot_i and it_i represent tables in the outer and inner parts of the query, and oe_i and ie_i represent expressions that refer to columns in the outer and inner tables.

  • It must be a single SELECT without UNION constructs.

  • It must not contain a GROUP BY or HAVING clause.

  • It must not be implicitly grouped (it must contain no aggregate functions).

  • It must not have ORDER BY with LIMIT.

  • It must not have STRAIGHT_JOIN in the outer query.

  • The number of outer and inner tables together must be less than the maximum number of tables permitted in a join.

  • The STRAIGHT_JOIN qualifier is not present.

The subquery may be correlated or uncorrelated. DISTINCT is permitted, as is LIMIT unless ORDER BY is also used.

If a subquery meets the preceding criteria, MySQL converts it to a semi-join and makes a cost-based choice from these strategies:

  • Convert the subquery to a join, or use table pullout and run the query as an inner join between subquery tables and outer tables. Table pullout pulls a table out from the subquery to the outer query.

  • Duplicate Weedout: Run the semi-join as if it was a join and remove duplicate records using a temporary table.

  • FirstMatch: When scanning the inner tables for row combinations and there are multiple instances of a given value group, choose one rather than returning them all. This "shortcuts" scanning and eliminates production of unnecessary rows.

  • LooseScan: Scan a subquery table using an index that enables a single value to be chosen from each subquery's value group.

  • Materialize the subquery into a temporary table with an index and use the temporary table to perform a join. The index is used to remove duplicates. The index might also be used later for lookups when joining the temporary table with the outer tables; if not, the table is scanned.

Each of these strategies can be enabled or disabled using the optimizer_switch system variable. The semijoin flag controls whether semi-joins are used. If it is set to on, the firstmatch, loosescan, duplicateweedout (added in MySQL 5.7.8), and materialization flags enable finer control over the permitted semi-join strategies. These flags are on by default. See Section 8.9.2, “Controlling Switchable Optimizations”.

If the duplicateweedout semi-join strategy is disabled, it is not used unless all other applicable strategies are also disabled.

If duplicateweedout is disabled, on occasion the optimizer may generate a query plan that is far from optimal. This occurs due to heuristic pruning during greedy search, which can be avoided by setting optimizer_prune_level=0.

As of MySQL 5.7.6, the optimizer minimizes differences in handling of views and subqueries in the FROM clause. This affects queries with the STRAIGHT_JOIN modifier and a view with an IN subquery that can be converted into a semi-join. The following query illustrates this because the change in processing causes a change in transformation, and thus a different execution strategy:

CREATE VIEW v AS
SELECT *
FROM t1
WHERE a IN (SELECT b
           FROM t2);

SELECT STRAIGHT_JOIN *
FROM t3 JOIN v ON t3.x = v.a;

Before 5.7.6, the optimizer first merges the view v into the outer query. When deciding whether to convert the IN subquery into a semi-join, it notices the STRAIGHT_JOIN and refuses the conversion.

As of 5.7.6, the optimizer first looks at the view and converts the IN subquery into a semi-join, then checks whether it is possible to merge the view into the outer query. Because the STRAIGHT_JOIN modifier in the outer query prevents semi-join, the optimizer refuses the merge, causing the derived table to be evaluated using a materialized table.

The use of semi-join strategies is indicated in EXPLAIN output as follows:

  • Semi-joined tables show up in the outer select. EXPLAIN EXTENDED plus SHOW WARNINGS shows the rewritten query, which displays the semi-join structure. From this you can get an idea about which tables were pulled out of the semi-join. If a subquery was converted to a semi-join, you will see that the subquery predicate is gone and its tables and WHERE clause were merged into the outer query join list and WHERE clause.

  • Temporary table use for Duplicate Weedout is indicated by Start temporary and End temporary in the Extra column. Tables that were not pulled out and are in the range of EXPLAIN output rows covered by Start temporary and End temporary will have their rowid in the temporary table.

  • FirstMatch(tbl_name) in the Extra column indicates join shortcutting.

  • LooseScan(m..n) in the Extra column indicates use of the LooseScan strategy. m and n are key part numbers.

  • Temporary table use for materialization is indicated by rows with a select_type value of MATERIALIZED and rows with a table value of <subqueryN>.

8.2.1.18.2 Optimizing Subqueries with Subquery Materialization

The optimizer uses subquery materialization as a strategy that enables more efficient subquery processing. Materialization speeds up query execution by generating a subquery result as a temporary table, normally in memory. The first time MySQL needs the subquery result, it materializes that result into a temporary table. Any subsequent time the result is needed, MySQL refers again to the temporary table. The table is indexed with a hash index to make lookups fast and inexpensive. The index is unique, which makes the table smaller because it has no duplicates.

Subquery materialization attempts to use an in-memory temporary table when possible, falling back to on-disk storage if the table becomes too large. See Section 8.4.4, “Internal Temporary Table Use in MySQL”.

If materialization is not used, the optimizer sometimes rewrites a noncorrelated subquery as a correlated subquery. For example, the following IN subquery is noncorrelated (where_condition involves only columns from t2 and not t1):

SELECT * FROM t1
WHERE t1.a IN (SELECT t2.b FROM t2 WHERE where_condition);

The optimizer might rewrite this as an EXISTS correlated subquery:

SELECT * FROM t1
WHERE EXISTS (SELECT t2.b FROM t2 WHERE where_condition AND t1.a=t2.b);

Subquery materialization using a temporary table avoids such rewrites and makes it possible to execute the subquery only once rather than once per row of the outer query.

For subquery materialization to be used in MySQL, the materialization flag of the optimizer_switch system variable must be on. Materialization then applies to subquery predicates that appear anywhere (in the select list, WHERE, ON, GROUP BY, HAVING, or ORDER BY), for predicates that fall into any of these use cases:

  • The predicate has this form, when no outer expression oe_i or inner expression ie_i is nullable. N can be 1 or larger.

    (oe_1, oe_2, ..., oe_N) [NOT] IN (SELECT ie_1, i_2, ..., ie_N ...)
    
  • The predicate has this form, when there is a single outer expression oe and inner expression ie. The expressions can be nullable.

    oe [NOT] IN (SELECT ie ...)
    
  • The predicate is IN or NOT IN and a result of UNKNOWN (NULL) has the same meaning as a result of FALSE.

The following examples illustrate how the requirement for equivalence of UNKNOWN and FALSE predicate evaluation affects whether subquery materialization can be used. Assume that where_condition involves columns only from t2 and not t1 so that the subquery is noncorrelated.

This query is subject to materialization:

SELECT * FROM t1
WHERE t1.a IN (SELECT t2.b FROM t2 WHERE where_condition);

Here, it does not matter whether the IN predicate returns UNKNOWN or FALSE. Either way, the row from t1 is not included in the query result.

An example where subquery materialization will not be used is the following query, where t2.b is a nullable column.

SELECT * FROM t1
WHERE (t1.a,t1.b) NOT IN (SELECT t2.a,t2.b FROM t2
                          WHERE where_condition);

The following restrictions apply to the use of subquery materialization:

  • The types of the inner and outer expressions must match. For example, the optimizer might be able to use materialization if both expressions are integer or both are decimal. The optimizer cannot use materialization if one expression is integer and the other is decimal.

  • The inner expression cannot be a BLOB.

Use of EXPLAIN with a query can give some indication of whether the optimizer uses subquery materialization. Compared to query execution that does not use materialization, select_type may change from DEPENDENT SUBQUERY to SUBQUERY. This indicates that, for a subquery that would be executed once per outer row, materialization enables the subquery to be executed just once. In addition, for EXPLAIN EXTENDED, the text displayed by a following SHOW WARNINGS will include materialize materialize and materialized-subquery.

8.2.1.18.3 Optimizing Derived Tables and View References

The optimizer can handle derived tables (subqueries in the FROM clause) and view references using two strategies:

  • Merge the derived table or view into the outer query block

  • Materialize the derived table or view to an internal temporary table

Example 1:

SELECT * FROM (SELECT * FROM t1) AS derived_t1;

With merging, that query is executed similar to:

SELECT * FROM t1;

Example 2:

SELECT *
  FROM t1 JOIN (SELECT t2.f1 FROM t2) AS derived_t2 ON t1.f2=derived_t2.f1
  WHERE t1.f1 > 0;

With merging, that query is executed similar to:

SELECT t1.*, t2.f1
  FROM t1 JOIN t2 ON t1.f2=t2.f1
  WHERE t1.f1 > 0;

With materialization, derived_t1 and derived_t2 are treated as a separate table within their respective queries.

As of MySQL 5.7.6, the optimizer handles derived tables and view references the same way: It avoids unnecessary materialization whenever possible, which enables pushing down conditions from the outer query to derived tables and produces more efficient execution plans. (For an example, see Section 8.2.1.18.2, “Optimizing Subqueries with Subquery Materialization”.) Before MySQL 5.7.6, derived tables were always materialized, whereas equivalent view references were sometimes materialized and sometimes merged. This inconsistent treatment of equivalent queries could lead to performance problems: Unnecessary derived table materialization takes time and prevents the optimizer from pushing down conditions to derived tables.

If merging would result in an outer query block that references more than 61 base tables, the optimizer chooses materialization instead.

As of MySQL 5.7.6, the optimizer handles propagation of an ORDER BY clause in a derived table or view reference to the outer query block by propagating the ORDER BY clause if the following conditions apply: The outer query is not grouped or aggregated; does not specify DISTINCT, HAVING, or ORDER BY; and has this derived table or view reference as the only source in the FROM clause. Otherwise, the optimizer ignores the ORDER BY clause. Before MySQL 5.7.6, the optimizer always propagated ORDER BY, even if it was irrelevant or resulted in an invalid query.

For statements such as DELETE or UPDATE that modify tables, using the merge strategy for a derived table that prior to MySQL 5.7.6 was materialized can result in an ER_UPDATE_TABLE_USED error:

mysql> DELETE FROM t1
    -> WHERE id IN (SELECT id
    ->              FROM (SELECT t1.id
    ->                    FROM t1 INNER JOIN t2 USING (id)
    ->                    WHERE t2.status = 0) AS t);
ERROR 1093 (HY000): You can't specify target table 't1'
for update in FROM clause

The error occurs when merging a derived table into the outer query block results in a statement that both selects from and modifies a table. (Materialization does not cause the problem because, in effect, it converts the derived table to a separate table.) To avoid this error, disable the derived_merge flag of the optimizer_switch system variable before executing the statement:

mysql> SET optimizer_switch = 'derived_merge=off';

The derived_merge flag controls whether the optimizer attempts to merge derived tables and views into the outer query block, assuming that no other rule prevents merging. By default, the flag is on to enable merging. Setting the flag to off prevents merging and avoids the error just described. Other workarounds include using in the subquery any constructs that prevent merging, although these are not as explicit in their effect on materialization. Constructs that prevent merging are the same as those that prevent merging in views. Examples are SELECT DISTINCT or LIMIT in the subquery. For details, see Section 20.5.2, “View Processing Algorithms”.

The derived_merge flag also applies to views that contain no ALGORITHM clause. Thus, if an ER_UPDATE_TABLE_USED error occurs for a view reference that uses an expression equivalent to the subquery, adding ALGORITHM=TEMPTABLE to the view definition prevents merging and takes precedence over the current derived_merge value.

If the optimizer chooses the materialization strategy for a derived table, it handles the query as follows:

  • The optimizer postpones materialization of subqueries in the FROM clause until their contents are needed during query execution. This improves performance because delay of materialization may result in not having to do it at all. Consider a query that joins the result of a subquery in the FROM clause to another table: If the optimizer processes that other table first and finds that it returns no rows, the join need not be carried out further and the optimizer can completely skip materializing the subquery.

  • During query execution, the optimizer may add an index to a derived table to speed up row retrieval from it.

Consider the following EXPLAIN statement, for which a subquery appears in the FROM clause of a SELECT query:

EXPLAIN SELECT * FROM (SELECT * FROM t1) AS derived_t1;

The optimizer avoids materializing the subquery by delaying it until the result is needed during SELECT execution. In this case, the query is not executed, so the result is never needed.

Even for queries that are executed, delay of subquery materialization may enable the optimizer to avoid materialization entirely. When this happens, query execution is quicker by the time needed to perform materialization. Consider the following query, which joins the result of a subquery in the FROM clause to another table:

SELECT *
  FROM t1 JOIN (SELECT t2.f1 FROM t2) AS derived_t2
          ON t1.f2=derived_t2.f1
  WHERE t1.f1 > 0;

If the optimization processes t1 first and the WHERE clause produces an empty result, the join must necessarily be empty and the subquery need not be materialized.

For cases when a derived table requires materialization, the optimizer may speed up access to the result by adding an index to the materialized table. If such an index enables ref access to the table, it can greatly reduce amount of data that must be read during query execution. Consider the following query:

SELECT *
 FROM t1 JOIN (SELECT DISTINCT f1 FROM t2) AS derived_t2
         ON t1.f1=derived_t2.f1;

The optimizer constructs an index over column f1 from derived_t2 if doing so would enable use of ref access for the lowest cost execution plan. After adding the index, the optimizer can treat the materialized derived table the same as a regular table with an index, and it benefits similarly from the generated index. The overhead of index creation is negligible compared to the cost of query execution without the index. If ref access would result in higher cost than some other access method, the optimizer creates no index and loses nothing.

8.2.1.18.4 Optimizing Subqueries with EXISTS Strategy

Certain optimizations are applicable to comparisons that use the IN operator to test subquery results (or that use =ANY, which is equivalent). This section discusses these optimizations, particularly with regard to the challenges that NULL values present. The last part of the discussion includes suggestions on what you can do to help the optimizer.

Consider the following subquery comparison:

outer_expr IN (SELECT inner_expr FROM ... WHERE subquery_where)

MySQL evaluates queries from outside to inside. That is, it first obtains the value of the outer expression outer_expr, and then runs the subquery and captures the rows that it produces.

A very useful optimization is to inform the subquery that the only rows of interest are those where the inner expression inner_expr is equal to outer_expr. This is done by pushing down an appropriate equality into the subquery's WHERE clause. That is, the comparison is converted to this:

EXISTS (SELECT 1 FROM ... WHERE subquery_where AND outer_expr=inner_expr)

After the conversion, MySQL can use the pushed-down equality to limit the number of rows that it must examine when evaluating the subquery.

More generally, a comparison of N values to a subquery that returns N-value rows is subject to the same conversion. If oe_i and ie_i represent corresponding outer and inner expression values, this subquery comparison:

(oe_1, ..., oe_N) IN
  (SELECT ie_1, ..., ie_N FROM ... WHERE subquery_where)

Becomes:

EXISTS (SELECT 1 FROM ... WHERE subquery_where
                          AND oe_1 = ie_1
                          AND ...
                          AND oe_N = ie_N)

For simplicity, the following discussion assumes a single pair of outer and inner expression values.

The conversion just described has its limitations. It is valid only if we ignore possible NULL values. That is, the pushdown strategy works as long as both of these two conditions are true:

  • outer_expr and inner_expr cannot be NULL.

  • You do not need to distinguish NULL from FALSE subquery results. If the subquery is a part of an OR or AND expression in the WHERE clause, MySQL assumes that you do not care. Another instance where the optimizer notices that NULL and FALSE subquery results need not be distinguished is this construct:

    ... WHERE outer_expr IN (subquery)
    

    In this case, the WHERE clause rejects the row whether IN (subquery) returns NULL or FALSE.

When either or both of those conditions do not hold, optimization is more complex.

Suppose that outer_expr is known to be a non-NULL value but the subquery does not produce a row such that outer_expr = inner_expr. Then outer_expr IN (SELECT ...) evaluates as follows:

  • NULL, if the SELECT produces any row where inner_expr is NULL

  • FALSE, if the SELECT produces only non-NULL values or produces nothing

In this situation, the approach of looking for rows with outer_expr = inner_expr is no longer valid. It is necessary to look for such rows, but if none are found, also look for rows where inner_expr is NULL. Roughly speaking, the subquery can be converted to something like this:

EXISTS (SELECT 1 FROM ... WHERE subquery_where AND
        (outer_expr=inner_expr OR inner_expr IS NULL))

The need to evaluate the extra IS NULL condition is why MySQL has the ref_or_null access method:

mysql> EXPLAIN
    -> SELECT outer_expr IN (SELECT t2.maybe_null_key
    ->                       FROM t2, t3 WHERE ...)
    -> FROM t1;
*************************** 1. row ***************************
           id: 1
  select_type: PRIMARY
        table: t1
...
*************************** 2. row ***************************
           id: 2
  select_type: DEPENDENT SUBQUERY
        table: t2
         type: ref_or_null
possible_keys: maybe_null_key
          key: maybe_null_key
      key_len: 5
          ref: func
         rows: 2
        Extra: Using where; Using index
...

The unique_subquery and index_subquery subquery-specific access methods also have or NULL variants. However, prior to MySQL 5.7.3, they are not visible in EXPLAIN output, so you must use EXPLAIN EXTENDED followed by SHOW WARNINGS (note the checking NULL in the warning message):

mysql> EXPLAIN EXTENDED
    -> SELECT outer_expr IN (SELECT maybe_null_key FROM t2) FROM t1\G
*************************** 1. row ***************************
           id: 1
  select_type: PRIMARY
        table: t1
...
*************************** 2. row ***************************
           id: 2
  select_type: DEPENDENT SUBQUERY
        table: t2
         type: index_subquery
possible_keys: maybe_null_key
          key: maybe_null_key
      key_len: 5
          ref: func
         rows: 2
        Extra: Using index

mysql> SHOW WARNINGS\G
*************************** 1. row ***************************
  Level: Note
   Code: 1003
Message: select (`test`.`t1`.`outer_expr`,
         (((`test`.`t1`.`outer_expr`) in t2 on
         maybe_null_key checking NULL))) AS `outer_expr IN (SELECT
         maybe_null_key FROM t2)` from `test`.`t1`

The additional OR ... IS NULL condition makes query execution slightly more complicated (and some optimizations within the subquery become inapplicable), but generally this is tolerable.

The situation is much worse when outer_expr can be NULL. According to the SQL interpretation of NULL as unknown value, NULL IN (SELECT inner_expr ...) should evaluate to:

  • NULL, if the SELECT produces any rows

  • FALSE, if the SELECT produces no rows

For proper evaluation, it is necessary to be able to check whether the SELECT has produced any rows at all, so outer_expr = inner_expr cannot be pushed down into the subquery. This is a problem, because many real world subqueries become very slow unless the equality can be pushed down.

Essentially, there must be different ways to execute the subquery depending on the value of outer_expr.

The optimizer chooses SQL compliance over speed, so it accounts for the possibility that outer_expr might be NULL.

If outer_expr is NULL, to evaluate the following expression, it is necessary to run the SELECT to determine whether it produces any rows:

NULL IN (SELECT inner_expr FROM ... WHERE subquery_where)

It is necessary to run the original SELECT here, without any pushed-down equalities of the kind mentioned earlier.

On the other hand, when outer_expr is not NULL, it is absolutely essential that this comparison:

outer_expr IN (SELECT inner_expr FROM ... WHERE subquery_where)

be converted to this expression that uses a pushed-down condition:

EXISTS (SELECT 1 FROM ... WHERE subquery_where AND outer_expr=inner_expr)

Without this conversion, subqueries will be slow. To solve the dilemma of whether to push down or not push down conditions into the subquery, the conditions are wrapped in trigger functions. Thus, an expression of the following form:

outer_expr IN (SELECT inner_expr FROM ... WHERE subquery_where)

is converted into:

EXISTS (SELECT 1 FROM ... WHERE subquery_where
                          AND trigcond(outer_expr=inner_expr))

More generally, if the subquery comparison is based on several pairs of outer and inner expressions, the conversion takes this comparison:

(oe_1, ..., oe_N) IN (SELECT ie_1, ..., ie_N FROM ... WHERE subquery_where)

and converts it to this expression:

EXISTS (SELECT 1 FROM ... WHERE subquery_where
                          AND trigcond(oe_1=ie_1)
                          AND ...
                          AND trigcond(oe_N=ie_N)
       )

Each trigcond(X) is a special function that evaluates to the following values:

  • X when the linked outer expression oe_i is not NULL

  • TRUE when the linked outer expression oe_i is NULL

Note

Trigger functions are not triggers of the kind that you create with CREATE TRIGGER.

Equalities that are wrapped into trigcond() functions are not first class predicates for the query optimizer. Most optimizations cannot deal with predicates that may be turned on and off at query execution time, so they assume any trigcond(X) to be an unknown function and ignore it. At the moment, triggered equalities can be used by those optimizations:

  • Reference optimizations: trigcond(X=Y [OR Y IS NULL]) can be used to construct ref, eq_ref, or ref_or_null table accesses.

  • Index lookup-based subquery execution engines: trigcond(X=Y) can be used to construct unique_subquery or index_subquery accesses.

  • Table-condition generator: If the subquery is a join of several tables, the triggered condition will be checked as soon as possible.

When the optimizer uses a triggered condition to create some kind of index lookup-based access (as for the first two items of the preceding list), it must have a fallback strategy for the case when the condition is turned off. This fallback strategy is always the same: Do a full table scan. In EXPLAIN output, the fallback shows up as Full scan on NULL key in the Extra column:

mysql> EXPLAIN SELECT t1.col1,
    -> t1.col1 IN (SELECT t2.key1 FROM t2 WHERE t2.col2=t1.col2) FROM t1\G
*************************** 1. row ***************************
           id: 1
  select_type: PRIMARY
        table: t1
        ...
*************************** 2. row ***************************
           id: 2
  select_type: DEPENDENT SUBQUERY
        table: t2
         type: index_subquery
possible_keys: key1
          key: key1
      key_len: 5
          ref: func
         rows: 2
        Extra: Using where; Full scan on NULL key

If you run EXPLAIN EXTENDED followed by SHOW WARNINGS, you can see the triggered condition:

*************************** 1. row ***************************
  Level: Note
   Code: 1003
Message: select `test`.`t1`.`col1` AS `col1`,
         <in_optimizer>(`test`.`t1`.`col1`,
         <exists>(<index_lookup>(<cache>(`test`.`t1`.`col1`) in t2
         on key1 checking NULL
         where (`test`.`t2`.`col2` = `test`.`t1`.`col2`) having
         trigcond(<is_not_null_test>(`test`.`t2`.`key1`))))) AS
         `t1.col1 IN (select t2.key1 from t2 where t2.col2=t1.col2)`
         from `test`.`t1`

The use of triggered conditions has some performance implications. A NULL IN (SELECT ...) expression now may cause a full table scan (which is slow) when it previously did not. This is the price paid for correct results (the goal of the trigger-condition strategy was to improve compliance and not speed).

For multiple-table subqueries, execution of NULL IN (SELECT ...) will be particularly slow because the join optimizer does not optimize for the case where the outer expression is NULL. It assumes that subquery evaluations with NULL on the left side are very rare, even if there are statistics that indicate otherwise. On the other hand, if the outer expression might be NULL but never actually is, there is no performance penalty.

To help the query optimizer better execute your queries, use these tips:

  • Declare a column as NOT NULL if it really is. (This also helps other aspects of the optimizer by simplifying condition testing for the column.)

  • If you do not need to distinguish a NULL from FALSE subquery result, you can easily avoid the slow execution path. Replace a comparison that looks like this:

    outer_expr IN (SELECT inner_expr FROM ...)
    

    with this expression:

    (outer_expr IS NOT NULL) AND (outer_expr IN (SELECT inner_expr FROM ...))
    

    Then NULL IN (SELECT ...) is never evaluated because MySQL stops evaluating AND parts as soon as the expression result is clear.

    Another possible rewrite:

    EXISTS (SELECT inner_expr FROM ...
            WHERE inner_expr=outer_expr)
    

    This would apply when you need not distinguish NULL from FALSE subquery results, in which case you may actually want EXISTS.

The subquery_materialization_cost_based flag enables control over the choice between subquery materialization and IN-to-EXISTS subquery transformation. See Section 8.9.2, “Controlling Switchable Optimizations”.

8.2.1.19 LIMIT Query Optimization

If you need only a specified number of rows from a result set, use a LIMIT clause in the query, rather than fetching the whole result set and throwing away the extra data.

MySQL sometimes optimizes a query that has a LIMIT row_count clause and no HAVING clause:

  • If you select only a few rows with LIMIT, MySQL uses indexes in some cases when normally it would prefer to do a full table scan.

  • If you combine LIMIT row_count with ORDER BY, MySQL ends the sorting as soon as it has found the first row_count rows of the sorted result, rather than sorting the entire result. If ordering is done by using an index, this is very fast. If a filesort must be done, all rows that match the query without the LIMIT clause are selected, and most or all of them are sorted, before the first row_count are found. After the initial rows have been found, MySQL does not sort any remainder of the result set.

    One manifestation of this behavior is that an ORDER BY query with and without LIMIT may return rows in different order, as described later in this section.

  • If you combine LIMIT row_count with DISTINCT, MySQL stops as soon as it finds row_count unique rows.

  • In some cases, a GROUP BY can be resolved by reading the index in order (or doing a sort on the index) and then calculating summaries until the index value changes. In this case, LIMIT row_count does not calculate any unnecessary GROUP BY values.

  • As soon as MySQL has sent the required number of rows to the client, it aborts the query unless you are using SQL_CALC_FOUND_ROWS. The number of rows can then be retrieved with SELECT FOUND_ROWS(). See Section 12.14, “Information Functions”.

  • LIMIT 0 quickly returns an empty set. This can be useful for checking the validity of a query. It can also be employed to obtain the types of the result columns if you are using a MySQL API that makes result set metadata available. With the mysql client program, you can use the --column-type-info option to display result column types.

  • If the server uses temporary tables to resolve the query, it uses the LIMIT row_count clause to calculate how much space is required.

If multiple rows have identical values in the ORDER BY columns, the server is free to return those rows in any order, and may do so differently depending on the overall execution plan. In other words, the sort order of those rows is nondeterministic with respect to the nonordered columns.

One factor that affects the execution plan is LIMIT, so an ORDER BY query with and without LIMIT may return rows in different orders. Consider this query, which is sorted by the category column but nondeterministic with respect to the id and rating columns:

mysql> SELECT * FROM ratings ORDER BY category;
+----+----------+--------+
| id | category | rating |
+----+----------+--------+
|  1 |        1 |    4.5 |
|  5 |        1 |    3.2 |
|  3 |        2 |    3.7 |
|  4 |        2 |    3.5 |
|  6 |        2 |    3.5 |
|  2 |        3 |    5.0 |
|  7 |        3 |    2.7 |
+----+----------+--------+

Including LIMIT may affect order of rows within each category value. For example, this is a valid query result:

mysql> SELECT * FROM ratings ORDER BY category LIMIT 5;
+----+----------+--------+
| id | category | rating |
+----+----------+--------+
|  1 |        1 |    4.5 |
|  5 |        1 |    3.2 |
|  4 |        2 |    3.5 |
|  3 |        2 |    3.7 |
|  6 |        2 |    3.5 |
+----+----------+--------+

In each case, the rows are sorted by the ORDER BY column, which is all that is required by the SQL standard.

If it is important to ensure the same row order with and without LIMIT, include additional columns in the ORDER BY clause to make the order deterministic. For example, if id values are unique, you can make rows for a given category value appear in id order by sorting like this:

mysql> SELECT * FROM ratings ORDER BY category, id;
+----+----------+--------+
| id | category | rating |
+----+----------+--------+
|  1 |        1 |    4.5 |
|  5 |        1 |    3.2 |
|  3 |        2 |    3.7 |
|  4 |        2 |    3.5 |
|  6 |        2 |    3.5 |
|  2 |        3 |    5.0 |
|  7 |        3 |    2.7 |
+----+----------+--------+

mysql> SELECT * FROM ratings ORDER BY category, id LIMIT 5;
+----+----------+--------+
| id | category | rating |
+----+----------+--------+
|  1 |        1 |    4.5 |
|  5 |        1 |    3.2 |
|  3 |        2 |    3.7 |
|  4 |        2 |    3.5 |
|  6 |        2 |    3.5 |
+----+----------+--------+

The optimizer does handle queries (and subqueries) of the following form:

SELECT ... FROM single_table ... ORDER BY non_index_column [DESC] LIMIT [M,]N;

That type of query is common in web applications that display only a few rows from a larger result set. For example:

SELECT col1, ... FROM t1 ... ORDER BY name LIMIT 10;
SELECT col1, ... FROM t1 ... ORDER BY RAND() LIMIT 15;

The sort buffer has a size of sort_buffer_size. If the sort elements for N rows are small enough to fit in the sort buffer (M+N rows if M was specified), the server can avoid using a merge file and perform the sort entirely in memory by treating the sort buffer as a priority queue:

  • Scan the table, inserting the select list columns from each selected row in sorted order in the queue. If the queue is full, bump out the last row in the sort order.

  • Return the first N rows from the queue. (If M was specified, skip the first M rows and return the next N rows.)

Previously, the server performed this operation by using a merge file for the sort:

  • Scan the table, repeating these steps through the end of the table:

    • Select rows until the sort buffer is filled.

    • Write the first N rows in the buffer (M+N rows if M was specified) to a merge file.

  • Sort the merge file and return the first N rows. (If M was specified, skip the first M rows and return the next N rows.)

The cost of the table scan is the same for the queue and merge-file methods, so the optimizer chooses between methods based on other costs:

  • The queue method involves more CPU for inserting rows into the queue in order

  • The merge-file method has I/O costs to write and read the file and CPU cost to sort it

The optimizer considers the balance between these factors for particular values of N and the row size.

8.2.1.20 Row Constructor Expression Optimization

Row constructors permit simultaneous comparisons of multiple values. For example, these two statements are semantically equivalent:

SELECT * FROM t1 WHERE (column1,column2) = (1,1);
SELECT * FROM t1 WHERE column1 = 1 AND column2 = 1;

In addition, the optimizer handles both expressions the same way.

The optimizer is less likely to use available indexes if the row constructor columns do not cover the prefix of an index. Consider the following table, which has a primary key on (c1, c2, c3):

CREATE TABLE t1 (
  c1 INT, c2 INT, c3 INT, c4 CHAR(100),
  PRIMARY KEY(c1,c2,c3)
);

In this query, the WHERE clause uses all columns in the index. However, the row constructor itself does not cover an index prefix, with the result that the optimizer uses only c1 (key_len=4, the size of c1):

mysql> EXPLAIN SELECT * FROM t1
    -> WHERE c1=1 AND (c2,c3) > (1,1)\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: t1
   partitions: NULL
         type: ref
possible_keys: PRIMARY
          key: PRIMARY
      key_len: 4
          ref: const
         rows: 3
     filtered: 100.00
        Extra: Using where

In such cases, rewriting the row constructor expression using an equivalent nonconstructor expression may result in more complete index use. For the given query, the row constructor and equivalent nonconstructor expressions are:

(c2,c3) > (1,1)
c2 > 1 OR ((c2 = 1) AND (c3 > 1))

Rewriting the query to use the nonconstructor expression results in the optimizer using all three columns in the index (key_len=12):

mysql> EXPLAIN SELECT * FROM t1
    -> WHERE c1 = 1 AND (c2 > 1 OR ((c2 = 1) AND (c3 > 1)))\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: t1
   partitions: NULL
         type: range
possible_keys: PRIMARY
          key: PRIMARY
      key_len: 12
          ref: NULL
         rows: 3
     filtered: 100.00
        Extra: Using where

Under certain conditions, the optimizer can apply the range access method to IN() expressions that have row constructor arguments. See Section 8.2.1.3.5, “Range Optimization of Row Constructor Expressions”.

8.2.1.21 How to Avoid Full Table Scans

The output from EXPLAIN shows ALL in the type column when MySQL uses a full table scan to resolve a query. This usually happens under the following conditions:

  • The table is so small that it is faster to perform a table scan than to bother with a key lookup. This is common for tables with fewer than 10 rows and a short row length.

  • There are no usable restrictions in the ON or WHERE clause for indexed columns.

  • You are comparing indexed columns with constant values and MySQL has calculated (based on the index tree) that the constants cover too large a part of the table and that a table scan would be faster. See Section 8.2.1.2, “How MySQL Optimizes WHERE Clauses”.

  • You are using a key with low cardinality (many rows match the key value) through another column. In this case, MySQL assumes that by using the key it probably will do many key lookups and that a table scan would be faster.

For small tables, a table scan often is appropriate and the performance impact is negligible. For large tables, try the following techniques to avoid having the optimizer incorrectly choose a table scan:

8.2.2 Optimizing DML Statements

This section explains how to speed up the data manipulation language (DML) statements, INSERT, UPDATE, and DELETE. Traditional OLTP applications and modern web applications typically do many small DML operations, where concurrency is vital. Data analysis and reporting applications typically run DML operations that affect many rows at once, where the main considerations is the I/O to write large amounts of data and keep indexes up-to-date. For inserting and updating large volumes of data (known in the industry as ETL, for extract-transform-load), sometimes you use other SQL statements or external commands, that mimic the effects of INSERT, UPDATE, and DELETE statements.

8.2.2.1 Speed of INSERT Statements

To optimize insert speed, combine many small operations into a single large operation. Ideally, you make a single connection, send the data for many new rows at once, and delay all index updates and consistency checking until the very end.

The time required for inserting a row is determined by the following factors, where the numbers indicate approximate proportions:

  • Connecting: (3)

  • Sending query to server: (2)

  • Parsing query: (2)

  • Inserting row: (1 × size of row)

  • Inserting indexes: (1 × number of indexes)

  • Closing: (1)

This does not take into consideration the initial overhead to open tables, which is done once for each concurrently running query.

The size of the table slows down the insertion of indexes by log N, assuming B-tree indexes.

You can use the following methods to speed up inserts:

8.2.2.2 Speed of UPDATE Statements

An update statement is optimized like a SELECT query with the additional overhead of a write. The speed of the write depends on the amount of data being updated and the number of indexes that are updated. Indexes that are not changed do not get updated.

Another way to get fast updates is to delay updates and then do many updates in a row later. Performing multiple updates together is much quicker than doing one at a time if you lock the table.

For a MyISAM table that uses dynamic row format, updating a row to a longer total length may split the row. If you do this often, it is very important to use OPTIMIZE TABLE occasionally. See Section 13.7.2.4, “OPTIMIZE TABLE Syntax”.

8.2.2.3 Speed of DELETE Statements

The time required to delete individual rows in a MyISAM table is exactly proportional to the number of indexes. To delete rows more quickly, you can increase the size of the key cache by increasing the key_buffer_size system variable. See Section 8.12.2, “Tuning Server Parameters”.

To delete all rows from a MyISAM table, TRUNCATE TABLE tbl_name is faster than DELETE FROM tbl_name. Truncate operations are not transaction-safe; an error occurs when attempting one in the course of an active transaction or active table lock. See Section 13.1.34, “TRUNCATE TABLE Syntax”.

8.2.3 Optimizing Database Privileges

The more complex your privilege setup, the more overhead applies to all SQL statements. Simplifying the privileges established by GRANT statements enables MySQL to reduce permission-checking overhead when clients execute statements. For example, if you do not grant any table-level or column-level privileges, the server need not ever check the contents of the tables_priv and columns_priv tables. Similarly, if you place no resource limits on any accounts, the server does not have to perform resource counting. If you have a very high statement-processing load, consider using a simplified grant structure to reduce permission-checking overhead.

8.2.4 Optimizing INFORMATION_SCHEMA Queries

Applications that monitor the database can make frequent use of the INFORMATION_SCHEMA tables. Certain types of queries for INFORMATION_SCHEMA tables can be optimized to execute more quickly. The goal is to minimize file operations (for example, scanning a directory or opening a table file) to collect the information that makes up these dynamic tables. These optimizations do have an effect on how collations are used for searches in INFORMATION_SCHEMA tables. For more information, see Section 10.1.7.9, “Collation and INFORMATION_SCHEMA Searches”.

1) Try to use constant lookup values for database and table names in the WHERE clause

You can take advantage of this principle as follows:

  • To look up databases or tables, use expressions that evaluate to a constant, such as literal values, functions that return a constant, or scalar subqueries.

  • Avoid queries that use a nonconstant database name lookup value (or no lookup value) because they require a scan of the data directory to find matching database directory names.

  • Within a database, avoid queries that use a nonconstant table name lookup value (or no lookup value) because they require a scan of the database directory to find matching table files.

This principle applies to the INFORMATION_SCHEMA tables shown in the following table, which shows the columns for which a constant lookup value enables the server to avoid a directory scan. For example, if you are selecting from TABLES, using a constant lookup value for TABLE_SCHEMA in the WHERE clause enables a data directory scan to be avoided.

TableColumn to specify to avoid data directory scanColumn to specify to avoid database directory scan
COLUMNSTABLE_SCHEMATABLE_NAME
KEY_COLUMN_USAGETABLE_SCHEMATABLE_NAME
PARTITIONSTABLE_SCHEMATABLE_NAME
REFERENTIAL_CONSTRAINTSCONSTRAINT_SCHEMATABLE_NAME
STATISTICSTABLE_SCHEMATABLE_NAME
TABLESTABLE_SCHEMATABLE_NAME
TABLE_CONSTRAINTSTABLE_SCHEMATABLE_NAME
TRIGGERSEVENT_OBJECT_SCHEMAEVENT_OBJECT_TABLE
VIEWSTABLE_SCHEMATABLE_NAME

The benefit of a query that is limited to a specific constant database name is that checks need be made only for the named database directory. Example:

SELECT TABLE_NAME FROM INFORMATION_SCHEMA.TABLES
WHERE TABLE_SCHEMA = 'test';

Use of the literal database name test enables the server to check only the test database directory, regardless of how many databases there might be. By contrast, the following query is less efficient because it requires a scan of the data directory to determine which database names match the pattern 'test%':

SELECT TABLE_NAME FROM INFORMATION_SCHEMA.TABLES
WHERE TABLE_SCHEMA LIKE 'test%';

For a query that is limited to a specific constant table name, checks need be made only for the named table within the corresponding database directory. Example:

SELECT TABLE_NAME FROM INFORMATION_SCHEMA.TABLES
WHERE TABLE_SCHEMA = 'test' AND TABLE_NAME = 't1';

Use of the literal table name t1 enables the server to check only the files for the t1 table, regardless of how many tables there might be in the test database. By contrast, the following query requires a scan of the test database directory to determine which table names match the pattern 't%':

SELECT TABLE_NAME FROM INFORMATION_SCHEMA.TABLES
WHERE TABLE_SCHEMA = 'test' AND TABLE_NAME LIKE 't%';

The following query requires a scan of the database directory to determine matching database names for the pattern 'test%', and for each matching database, it requires a scan of the database directory to determine matching table names for the pattern 't%':

SELECT TABLE_NAME FROM INFORMATION_SCHEMA.TABLES
WHERE TABLE_SCHEMA = 'test%' AND TABLE_NAME LIKE 't%';

2) Write queries that minimize the number of table files that must be opened

For queries that refer to certain INFORMATION_SCHEMA table columns, several optimizations are available that minimize the number of table files that must be opened. Example:

SELECT TABLE_NAME, ENGINE FROM INFORMATION_SCHEMA.TABLES
WHERE TABLE_SCHEMA = 'test';

In this case, after the server has scanned the database directory to determine the names of the tables in the database, those names become available with no further file system lookups. Thus, TABLE_NAME requires no files to be opened. The ENGINE (storage engine) value can be determined by opening the table's .frm file, without touching other table files such as the .MYD or .MYI file.

Some values, such as INDEX_LENGTH for MyISAM tables, require opening the .MYD or .MYI file as well.

The file-opening optimization types are denoted thus:

  • SKIP_OPEN_TABLE: Table files do not need to be opened. The information has already become available within the query by scanning the database directory.

  • OPEN_FRM_ONLY: Only the table's .frm file need be opened.

  • OPEN_TRIGGER_ONLY: Only the table's .TRG file need be opened.

  • OPEN_FULL_TABLE: The unoptimized information lookup. The .frm, .MYD, and .MYI files must be opened.

The following list indicates how the preceding optimization types apply to INFORMATION_SCHEMA table columns. For tables and columns not named, none of the optimizations apply.

  • COLUMNS: OPEN_FRM_ONLY applies to all columns

  • KEY_COLUMN_USAGE: OPEN_FULL_TABLE applies to all columns

  • PARTITIONS: OPEN_FULL_TABLE applies to all columns

  • REFERENTIAL_CONSTRAINTS: OPEN_FULL_TABLE applies to all columns

  • STATISTICS:

    ColumnOptimization type
    TABLE_CATALOGOPEN_FRM_ONLY
    TABLE_SCHEMAOPEN_FRM_ONLY
    TABLE_NAMEOPEN_FRM_ONLY
    NON_UNIQUEOPEN_FRM_ONLY
    INDEX_SCHEMAOPEN_FRM_ONLY
    INDEX_NAMEOPEN_FRM_ONLY
    SEQ_IN_INDEXOPEN_FRM_ONLY
    COLUMN_NAMEOPEN_FRM_ONLY
    COLLATIONOPEN_FRM_ONLY
    CARDINALITYOPEN_FULL_TABLE
    SUB_PARTOPEN_FRM_ONLY
    PACKEDOPEN_FRM_ONLY
    NULLABLEOPEN_FRM_ONLY
    INDEX_TYPEOPEN_FULL_TABLE
    COMMENTOPEN_FRM_ONLY
  • TABLES:

    ColumnOptimization type
    TABLE_CATALOGSKIP_OPEN_TABLE
    TABLE_SCHEMASKIP_OPEN_TABLE
    TABLE_NAMESKIP_OPEN_TABLE
    TABLE_TYPEOPEN_FRM_ONLY
    ENGINEOPEN_FRM_ONLY
    VERSIONOPEN_FRM_ONLY
    ROW_FORMATOPEN_FULL_TABLE
    TABLE_ROWSOPEN_FULL_TABLE
    AVG_ROW_LENGTHOPEN_FULL_TABLE
    DATA_LENGTHOPEN_FULL_TABLE
    MAX_DATA_LENGTHOPEN_FULL_TABLE
    INDEX_LENGTHOPEN_FULL_TABLE
    DATA_FREEOPEN_FULL_TABLE
    AUTO_INCREMENTOPEN_FULL_TABLE
    CREATE_TIMEOPEN_FULL_TABLE
    UPDATE_TIMEOPEN_FULL_TABLE
    CHECK_TIMEOPEN_FULL_TABLE
    TABLE_COLLATIONOPEN_FRM_ONLY
    CHECKSUMOPEN_FULL_TABLE
    CREATE_OPTIONSOPEN_FRM_ONLY
    TABLE_COMMENTOPEN_FRM_ONLY
  • TABLE_CONSTRAINTS: OPEN_FULL_TABLE applies to all columns

  • TRIGGERS: OPEN_TRIGGER_ONLY applies to all columns

  • VIEWS:

    ColumnOptimization type
    TABLE_CATALOGOPEN_FRM_ONLY
    TABLE_SCHEMAOPEN_FRM_ONLY
    TABLE_NAMEOPEN_FRM_ONLY
    VIEW_DEFINITIONOPEN_FRM_ONLY
    CHECK_OPTIONOPEN_FRM_ONLY
    IS_UPDATABLEOPEN_FULL_TABLE
    DEFINEROPEN_FRM_ONLY
    SECURITY_TYPEOPEN_FRM_ONLY
    CHARACTER_SET_CLIENTOPEN_FRM_ONLY
    COLLATION_CONNECTIONOPEN_FRM_ONLY

3) Use EXPLAIN to determine whether the server can use INFORMATION_SCHEMA optimizations for a query

This applies particularly for INFORMATION_SCHEMA queries that search for information from more than one database, which might take a long time and impact performance. The Extra value in EXPLAIN output indicates which, if any, of the optimizations described earlier the server can use to evaluate INFORMATION_SCHEMA queries. The following examples demonstrate the kinds of information you can expect to see in the Extra value.

mysql> EXPLAIN SELECT TABLE_NAME FROM INFORMATION_SCHEMA.VIEWS WHERE
    -> TABLE_SCHEMA = 'test' AND TABLE_NAME = 'v1'\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: VIEWS
         type: ALL
possible_keys: NULL
          key: TABLE_SCHEMA,TABLE_NAME
      key_len: NULL
          ref: NULL
         rows: NULL
        Extra: Using where; Open_frm_only; Scanned 0 databases

Use of constant database and table lookup values enables the server to avoid directory scans. For references to VIEWS.TABLE_NAME, only the .frm file need be opened.

mysql> EXPLAIN SELECT TABLE_NAME, ROW_FORMAT FROM INFORMATION_SCHEMA.TABLES\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: TABLES
         type: ALL
possible_keys: NULL
          key: NULL
      key_len: NULL
          ref: NULL
         rows: NULL
        Extra: Open_full_table; Scanned all databases

No lookup values are provided (there is no WHERE clause), so the server must scan the data directory and each database directory. For each table thus identified, the table name and row format are selected. TABLE_NAME requires no further table files to be opened (the SKIP_OPEN_TABLE optimization applies). ROW_FORMAT requires all table files to be opened (OPEN_FULL_TABLE applies). EXPLAIN reports OPEN_FULL_TABLE because it is more expensive than SKIP_OPEN_TABLE.

mysql> EXPLAIN SELECT TABLE_NAME, TABLE_TYPE FROM INFORMATION_SCHEMA.TABLES
    -> WHERE TABLE_SCHEMA = 'test'\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: TABLES
         type: ALL
possible_keys: NULL
          key: TABLE_SCHEMA
      key_len: NULL
          ref: NULL
         rows: NULL
        Extra: Using where; Open_frm_only; Scanned 1 database

No table name lookup value is provided, so the server must scan the test database directory. For the TABLE_NAME and TABLE_TYPE columns, the SKIP_OPEN_TABLE and OPEN_FRM_ONLY optimizations apply, respectively. EXPLAIN reports OPEN_FRM_ONLY because it is more expensive.

mysql> EXPLAIN SELECT B.TABLE_NAME
    -> FROM INFORMATION_SCHEMA.TABLES AS A, INFORMATION_SCHEMA.COLUMNS AS B
    -> WHERE A.TABLE_SCHEMA = 'test'
    -> AND A.TABLE_NAME = 't1'
    -> AND B.TABLE_NAME = A.TABLE_NAME\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: A
         type: ALL
possible_keys: NULL
          key: TABLE_SCHEMA,TABLE_NAME
      key_len: NULL
          ref: NULL
         rows: NULL
        Extra: Using where; Skip_open_table; Scanned 0 databases
*************************** 2. row ***************************
           id: 1
  select_type: SIMPLE
        table: B
         type: ALL
possible_keys: NULL
          key: NULL
      key_len: NULL
          ref: NULL
         rows: NULL
        Extra: Using where; Open_frm_only; Scanned all databases;
               Using join buffer

For the first EXPLAIN output row: Constant database and table lookup values enable the server to avoid directory scans for TABLES values. References to TABLES.TABLE_NAME require no further table files.

For the second EXPLAIN output row: All COLUMNS table values are OPEN_FRM_ONLY lookups, so COLUMNS.TABLE_NAME requires the .frm file to be opened.

mysql> EXPLAIN SELECT * FROM INFORMATION_SCHEMA.COLLATIONS\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: COLLATIONS
         type: ALL
possible_keys: NULL
          key: NULL
      key_len: NULL
          ref: NULL
         rows: NULL
        Extra:

In this case, no optimizations apply because COLLATIONS is not one of the INFORMATION_SCHEMA tables for which optimizations are available.

8.2.5 Other Optimization Tips

This section lists a number of miscellaneous tips for improving query processing speed:

  • If your application makes several database requests to perform related updates, combining the statements into a stored routine can help performance. Similarly, if your application computes a single result based on several column values or large volumes of data, combining the computation into a UDF (user-defined function) can help performance. The resulting fast database operations are then available to be reused by other queries, applications, and even code written in different programming languages. See Section 20.2, “Using Stored Routines (Procedures and Functions)” and Section 25.4, “Adding New Functions to MySQL” for more information.

  • To fix any compression issues that occur with ARCHIVE tables, use OPTIMIZE TABLE. See Section 15.5, “The ARCHIVE Storage Engine”.

  • If possible, classify reports as live or as statistical, where data needed for statistical reports is created only from summary tables that are generated periodically from the live data.

  • If you have data that does not conform well to a rows-and-columns table structure, you can pack and store data into a BLOB column. In this case, you must provide code in your application to pack and unpack information, but this might save I/O operations to read and write the sets of related values.

  • With Web servers, store images and other binary assets as files, with the path name stored in the database rather than the file itself. Most Web servers are better at caching files than database contents, so using files is generally faster. (Although you must handle backups and storage issues yourself in this case.)

  • If you need really high speed, look at the low-level MySQL interfaces. For example, by accessing the MySQL InnoDB or MyISAM storage engine directly, you could get a substantial speed increase compared to using the SQL interface.

  • Replication can provide a performance benefit for some operations. You can distribute client retrievals among replication servers to split up the load. To avoid slowing down the master while making backups, you can make backups using a slave server. See Chapter 17, Replication.

8.3 Optimization and Indexes

The best way to improve the performance of SELECT operations is to create indexes on one or more of the columns that are tested in the query. The index entries act like pointers to the table rows, allowing the query to quickly determine which rows match a condition in the WHERE clause, and retrieve the other column values for those rows. All MySQL data types can be indexed.

Although it can be tempting to create an indexes for every possible column used in a query, unnecessary indexes waste space and waste time for MySQL to determine which indexes to use. Indexes also add to the cost of inserts, updates, and deletes because each index must be updated. You must find the right balance to achieve fast queries using the optimal set of indexes.

8.3.1 How MySQL Uses Indexes

Indexes are used to find rows with specific column values quickly. Without an index, MySQL must begin with the first row and then read through the entire table to find the relevant rows. The larger the table, the more this costs. If the table has an index for the columns in question, MySQL can quickly determine the position to seek to in the middle of the data file without having to look at all the data. This is much faster than reading every row sequentially.

Most MySQL indexes (PRIMARY KEY, UNIQUE, INDEX, and FULLTEXT) are stored in B-trees. Exceptions: Indexes on spatial data types use R-trees; MEMORY tables also support hash indexes; InnoDB uses inverted lists for FULLTEXT indexes.

In general, indexes are used as described in the following discussion. Characteristics specific to hash indexes (as used in MEMORY tables) are described in Section 8.3.8, “Comparison of B-Tree and Hash Indexes”.

MySQL uses indexes for these operations:

  • To find the rows matching a WHERE clause quickly.

  • To eliminate rows from consideration. If there is a choice between multiple indexes, MySQL normally uses the index that finds the smallest number of rows (the most selective index).

  • If the table has a multiple-column index, any leftmost prefix of the index can be used by the optimizer to look up rows. For example, if you have a three-column index on (col1, col2, col3), you have indexed search capabilities on (col1), (col1, col2), and (col1, col2, col3). For more information, see Section 8.3.5, “Multiple-Column Indexes”.

  • To retrieve rows from other tables when performing joins. MySQL can use indexes on columns more efficiently if they are declared as the same type and size. In this context, VARCHAR and CHAR are considered the same if they are declared as the same size. For example, VARCHAR(10) and CHAR(10) are the same size, but VARCHAR(10) and CHAR(15) are not.

    For comparisons between nonbinary string columns, both columns should use the same character set. For example, comparing a utf8 column with a latin1 column precludes use of an index.

    Comparison of dissimilar columns (comparing a string column to a temporal or numeric column, for example) may prevent use of indexes if values cannot be compared directly without conversion. For a given value such as 1 in the numeric column, it might compare equal to any number of values in the string column such as '1', ' 1', '00001', or '01.e1'. This rules out use of any indexes for the string column.

  • To find the MIN() or MAX() value for a specific indexed column key_col. This is optimized by a preprocessor that checks whether you are using WHERE key_part_N = constant on all key parts that occur before key_col in the index. In this case, MySQL does a single key lookup for each MIN() or MAX() expression and replaces it with a constant. If all expressions are replaced with constants, the query returns at once. For example:

    SELECT MIN(key_part2),MAX(key_part2)
      FROM tbl_name WHERE key_part1=10;
    
  • To sort or group a table if the sorting or grouping is done on a leftmost prefix of a usable index (for example, ORDER BY key_part1, key_part2). If all key parts are followed by DESC, the key is read in reverse order. See Section 8.2.1.15, “ORDER BY Optimization”, and Section 8.2.1.16, “GROUP BY Optimization”.

  • In some cases, a query can be optimized to retrieve values without consulting the data rows. (An index that provides all the necessary results for a query is called a covering index.) If a query uses from a table only columns that are included in some index, the selected values can be retrieved from the index tree for greater speed:

    SELECT key_part3 FROM tbl_name
      WHERE key_part1=1
    

Indexes are less important for queries on small tables, or big tables where report queries process most or all of the rows. When a query needs to access most of the rows, reading sequentially is faster than working through an index. Sequential reads minimize disk seeks, even if not all the rows are needed for the query. See Section 8.2.1.21, “How to Avoid Full Table Scans” for details.

8.3.2 Using Primary Keys

The primary key for a table represents the column or set of columns that you use in your most vital queries. It has an associated index, for fast query performance. Query performance benefits from the NOT NULL optimization, because it cannot include any NULL values. With the InnoDB storage engine, the table data is physically organized to do ultra-fast lookups and sorts based on the primary key column or columns.

If your table is big and important, but does not have an obvious column or set of columns to use as a primary key, you might create a separate column with auto-increment values to use as the primary key. These unique IDs can serve as pointers to corresponding rows in other tables when you join tables using foreign keys.

8.3.3 Using Foreign Keys

If a table has many columns, and you query many different combinations of columns, it might be efficient to split the less-frequently used data into separate tables with a few columns each, and relate them back to the main table by duplicating the numeric ID column from the main table. That way, each small table can have a primary key for fast lookups of its data, and you can query just the set of columns that you need using a join operation. Depending on how the data is distributed, the queries might perform less I/O and take up less cache memory because the relevant columns are packed together on disk. (To maximize performance, queries try to read as few data blocks as possible from disk; tables with only a few columns can fit more rows in each data block.)

8.3.4 Column Indexes

The most common type of index involves a single column, storing copies of the values from that column in a data structure, allowing fast lookups for the rows with the corresponding column values. The B-tree data structure lets the index quickly find a specific value, a set of values, or a range of values, corresponding to operators such as =, >, , BETWEEN, IN, and so on, in a WHERE clause.

The maximum number of indexes per table and the maximum index length is defined per storage engine. See Chapter 14, The InnoDB Storage Engine, and Chapter 15, Alternative Storage Engines. All storage engines support at least 16 indexes per table and a total index length of at least 256 bytes. Most storage engines have higher limits.

For additional information about column indexes, see Section 13.1.14, “CREATE INDEX Syntax”.

Prefix Indexes

With col_name(N) syntax in an index specification for a string column, you can create an index that uses only the first N characters of the column. Indexing only a prefix of column values in this way can make the index file much smaller. When you index a BLOB or TEXT column, you must specify a prefix length for the index. For example:

CREATE TABLE test (blob_col BLOB, INDEX(blob_col(10)));

Prefixes can be up to 1000 bytes long (767 bytes for InnoDB tables, unless you have innodb_large_prefix set).

Note

Prefix limits are measured in bytes, whereas the prefix length in CREATE TABLE, ALTER TABLE, and CREATE INDEX statements is interpreted as number of characters for nonbinary string types (CHAR, VARCHAR, TEXT) and number of bytes for binary string types (BINARY, VARBINARY, BLOB). Take this into account when specifying a prefix length for a nonbinary string column that uses a multibyte character set.

For additional information about index prefixes, see Section 13.1.14, “CREATE INDEX Syntax”.

FULLTEXT Indexes

FULLTEXT indexes are used for full-text searches. Only the InnoDB and MyISAM storage engines support FULLTEXT indexes and only for CHAR, VARCHAR, and TEXT columns. Indexing always takes place over the entire column and column prefix indexing is not supported. For details, see Section 12.9, “Full-Text Search Functions”.

Optimizations are applied to certain kinds of FULLTEXT queries against single InnoDB tables. Queries with these characteristics are particularly efficient:

  • FULLTEXT queries that only return the document ID, or the document ID and the search rank.

  • FULLTEXT queries that sort the matching rows in descending order of score and apply a LIMIT clause to take the top N matching rows. For this optimization to apply, there must be no WHERE clauses and only a single ORDER BY clause in descending order.

  • FULLTEXT queries that retrieve only the COUNT(*) value of rows matching a search term, with no additional WHERE clauses. Code the WHERE clause as WHERE MATCH(text) AGAINST ('other_text'), without any > 0 comparison operator.

Spatial Indexes

You can create indexes on spatial data types. MyISAM and (as of MySQL 5.7.5) InnoDB support R-tree indexes on spatial types. Other storage engines use B-trees for indexing spatial types (except for ARCHIVE, which does not support spatial type indexing).

Indexes in the MEMORY Storage Engine

The MEMORY storage engine uses HASH indexes by default, but also supports BTREE indexes.

8.3.5 Multiple-Column Indexes

MySQL can create composite indexes (that is, indexes on multiple columns). An index may consist of up to 16 columns. For certain data types, you can index a prefix of the column (see Section 8.3.4, “Column Indexes”).

MySQL can use multiple-column indexes for queries that test all the columns in the index, or queries that test just the first column, the first two columns, the first three columns, and so on. If you specify the columns in the right order in the index definition, a single composite index can speed up several kinds of queries on the same table.

A multiple-column index can be considered a sorted array, the rows of which contain values that are created by concatenating the values of the indexed columns.

Note

As an alternative to a composite index, you can introduce a column that is hashed based on information from other columns. If this column is short, reasonably unique, and indexed, it might be faster than a wide index on many columns. In MySQL, it is very easy to use this extra column:

SELECT * FROM tbl_name
  WHERE hash_col=MD5(CONCAT(val1,val2))
  AND col1=val1 AND col2=val2;

Suppose that a table has the following specification:

CREATE TABLE test (
    id         INT NOT NULL,
    last_name  CHAR(30) NOT NULL,
    first_name CHAR(30) NOT NULL,
    PRIMARY KEY (id),
    INDEX name (last_name,first_name)
);

The name index is an index over the last_name and first_name columns. The index can be used for lookups in queries that specify values in a known range for combinations of last_name and first_name values. It can also be used for queries that specify just a last_name value because that column is a leftmost prefix of the index (as described later in this section). Therefore, the name index is used for lookups in the following queries:

SELECT * FROM test WHERE last_name='Widenius';

SELECT * FROM test
  WHERE last_name='Widenius' AND first_name='Michael';

SELECT * FROM test
  WHERE last_name='Widenius'
  AND (first_name='Michael' OR first_name='Monty');

SELECT * FROM test
  WHERE last_name='Widenius'
  AND first_name >='M' AND first_name < 'N';

However, the name index is not used for lookups in the following queries:

SELECT * FROM test WHERE first_name='Michael';

SELECT * FROM test
  WHERE last_name='Widenius' OR first_name='Michael';

Suppose that you issue the following SELECT statement:

SELECT * FROM tbl_name
  WHERE col1=val1 AND col2=val2;

If a multiple-column index exists on col1 and col2, the appropriate rows can be fetched directly. If separate single-column indexes exist on col1 and col2, the optimizer attempts to use the Index Merge optimization (see Section 8.2.1.4, “Index Merge Optimization”), or attempts to find the most restrictive index by deciding which index excludes more rows and using that index to fetch the rows.

If the table has a multiple-column index, any leftmost prefix of the index can be used by the optimizer to look up rows. For example, if you have a three-column index on (col1, col2, col3), you have indexed search capabilities on (col1), (col1, col2), and (col1, col2, col3).

MySQL cannot use the index to perform lookups if the columns do not form a leftmost prefix of the index. Suppose that you have the SELECT statements shown here:

SELECT * FROM tbl_name WHERE col1=val1;
SELECT * FROM tbl_name WHERE col1=val1 AND col2=val2;

SELECT * FROM tbl_name WHERE col2=val2;
SELECT * FROM tbl_name WHERE col2=val2 AND col3=val3;

If an index exists on (col1, col2, col3), only the first two queries use the index. The third and fourth queries do involve indexed columns, but (col2) and (col2, col3) are not leftmost prefixes of (col1, col2, col3).

8.3.6 Verifying Index Usage

Always check whether all your queries really use the indexes that you have created in the tables. Use the EXPLAIN statement, as described in Section 8.8.1, “Optimizing Queries with EXPLAIN”.

8.3.7 InnoDB and MyISAM Index Statistics Collection

Storage engines collect statistics about tables for use by the optimizer. Table statistics are based on value groups, where a value group is a set of rows with the same key prefix value. For optimizer purposes, an important statistic is the average value group size.

MySQL uses the average value group size in the following ways:

  • To estimate how may rows must be read for each ref access

  • To estimate how many row a partial join will produce; that is, the number of rows that an operation of this form will produce:

    (...) JOIN tbl_name ON tbl_name.key = expr
    

As the average value group size for an index increases, the index is less useful for those two purposes because the average number of rows per lookup increases: For the index to be good for optimization purposes, it is best that each index value target a small number of rows in the table. When a given index value yields a large number of rows, the index is less useful and MySQL is less likely to use it.

The average value group size is related to table cardinality, which is the number of value groups. The SHOW INDEX statement displays a cardinality value based on N/S, where N is the number of rows in the table and S is the average value group size. That ratio yields an approximate number of value groups in the table.

For a join based on the <=> comparison operator, NULL is not treated differently from any other value: NULL <=> NULL, just as N <=> N for any other N.

However, for a join based on the = operator, NULL is different from non-NULL values: expr1 = expr2 is not true when expr1 or expr2 (or both) are NULL. This affects ref accesses for comparisons of the form tbl_name.key = expr: MySQL will not access the table if the current value of expr is NULL, because the comparison cannot be true.

For = comparisons, it does not matter how many NULL values are in the table. For optimization purposes, the relevant value is the average size of the non-NULL value groups. However, MySQL does not currently enable that average size to be collected or used.

For InnoDB and MyISAM tables, you have some control over collection of table statistics by means of the innodb_stats_method and myisam_stats_method system variables, respectively. These variables have three possible values, which differ as follows:

  • When the variable is set to nulls_equal, all NULL values are treated as identical (that is, they all form a single value group).

    If the NULL value group size is much higher than the average non-NULL value group size, this method skews the average value group size upward. This makes index appear to the optimizer to be less useful than it really is for joins that look for non-NULL values. Consequently, the nulls_equal method may cause the optimizer not to use the index for ref accesses when it should.

  • When the variable is set to nulls_unequal, NULL values are not considered the same. Instead, each NULL value forms a separate value group of size 1.

    If you have many NULL values, this method skews the average value group size downward. If the average non-NULL value group size is large, counting NULL values each as a group of size 1 causes the optimizer to overestimate the value of the index for joins that look for non-NULL values. Consequently, the nulls_unequal method may cause the optimizer to use this index for ref lookups when other methods may be better.

  • When the variable is set to nulls_ignored, NULL values are ignored.

If you tend to use many joins that use <=> rather than =, NULL values are not special in comparisons and one NULL is equal to another. In this case, nulls_equal is the appropriate statistics method.

The innodb_stats_method system variable has a global value; the myisam_stats_method system variable has both global and session values. Setting the global value affects statistics collection for tables from the corresponding storage engine. Setting the session value affects statistics collection only for the current client connection. This means that you can force a table's statistics to be regenerated with a given method without affecting other clients by setting the session value of myisam_stats_method.

To regenerate MyISAM table statistics, you can use any of the following methods:

Some caveats regarding the use of innodb_stats_method and myisam_stats_method:

  • You can force table statistics to be collected explicitly, as just described. However, MySQL may also collect statistics automatically. For example, if during the course of executing statements for a table, some of those statements modify the table, MySQL may collect statistics. (This may occur for bulk inserts or deletes, or some ALTER TABLE statements, for example.) If this happens, the statistics are collected using whatever value innodb_stats_method or myisam_stats_method has at the time. Thus, if you collect statistics using one method, but the system variable is set to the other method when a table's statistics are collected automatically later, the other method will be used.

  • There is no way to tell which method was used to generate statistics for a given table.

  • These variables apply only to InnoDB and MyISAM tables. Other storage engines have only one method for collecting table statistics. Usually it is closer to the nulls_equal method.

8.3.8 Comparison of B-Tree and Hash Indexes

Understanding the B-tree and hash data structures can help predict how different queries perform on different storage engines that use these data structures in their indexes, particularly for the MEMORY storage engine that lets you choose B-tree or hash indexes.

B-Tree Index Characteristics

A B-tree index can be used for column comparisons in expressions that use the =, >, >=, <, <=, or BETWEEN operators. The index also can be used for LIKE comparisons if the argument to LIKE is a constant string that does not start with a wildcard character. For example, the following SELECT statements use indexes:

SELECT * FROM tbl_name WHERE key_col LIKE 'Patrick%';
SELECT * FROM tbl_name WHERE key_col LIKE 'Pat%_ck%';

In the first statement, only rows with 'Patrick' <= key_col < 'Patricl' are considered. In the second statement, only rows with 'Pat' <= key_col < 'Pau' are considered.

The following SELECT statements do not use indexes:

SELECT * FROM tbl_name WHERE key_col LIKE '%Patrick%';
SELECT * FROM tbl_name WHERE key_col LIKE other_col;

In the first statement, the LIKE value begins with a wildcard character. In the second statement, the LIKE value is not a constant.

If you use ... LIKE '%string%' and string is longer than three characters, MySQL uses the Turbo Boyer-Moore algorithm to initialize the pattern for the string and then uses this pattern to perform the search more quickly.

A search using col_name IS NULL employs indexes if col_name is indexed.

Any index that does not span all AND levels in the WHERE clause is not used to optimize the query. In other words, to be able to use an index, a prefix of the index must be used in every AND group.

The following WHERE clauses use indexes:

... WHERE index_part1=1 AND index_part2=2 AND other_column=3

    /* index = 1 OR index = 2 */
... WHERE index=1 OR A=10 AND index=2

    /* optimized like "index_part1='hello'" */
... WHERE index_part1='hello' AND index_part3=5

    /* Can use index on index1 but not on index2 or index3 */
... WHERE index1=1 AND index2=2 OR index1=3 AND index3=3;

These WHERE clauses do not use indexes:

    /* index_part1 is not used */
... WHERE index_part2=1 AND index_part3=2

    /*  Index is not used in both parts of the WHERE clause  */
... WHERE index=1 OR A=10

    /* No index spans all rows  */
... WHERE index_part1=1 OR index_part2=10

Sometimes MySQL does not use an index, even if one is available. One circumstance under which this occurs is when the optimizer estimates that using the index would require MySQL to access a very large percentage of the rows in the table. (In this case, a table scan is likely to be much faster because it requires fewer seeks.) However, if such a query uses LIMIT to retrieve only some of the rows, MySQL uses an index anyway, because it can much more quickly find the few rows to return in the result.

Hash Index Characteristics

Hash indexes have somewhat different characteristics from those just discussed:

  • They are used only for equality comparisons that use the = or <=> operators (but are very fast). They are not used for comparison operators such as < that find a range of values. Systems that rely on this type of single-value lookup are known as key-value stores; to use MySQL for such applications, use hash indexes wherever possible.

  • The optimizer cannot use a hash index to speed up ORDER BY operations. (This type of index cannot be used to search for the next entry in order.)

  • MySQL cannot determine approximately how many rows there are between two values (this is used by the range optimizer to decide which index to use). This may affect some queries if you change a MyISAM or InnoDB table to a hash-indexed MEMORY table.

  • Only whole keys can be used to search for a row. (With a B-tree index, any leftmost prefix of the key can be used to find rows.)

8.3.9 Optimizer Use of Generated Column Indexes

MySQL supports indexes on generated columns. For example:

CREATE TABLE t1 (f1 INT, gc INT AS (f1 + 1) STORED, INDEX (gc));

The generated column, gc, is defined as the expression f1 + 1. The column is also indexed and the optimizer can take that index into account during execution plan construction. In the following query, the WHERE clause refers to gc and the optimizer considers whether the index on that column yields a more efficient plan:

SELECT * FROM t1 WHERE gc > 9;

As of MySQL 5.7.8, the optimizer can use indexes on generated columns to generate execution plans, even in the absence of direct references in queries to those columns by name. This occurs if the WHERE, ORDER BY, or GROUP BY clause refers to an expression that matches the definition of some indexed generated column. The following query does not refer directly to gc but does use an expression that matches the definition of gc:

SELECT * FROM t1 WHERE f1 + 1 > 9;

The optimizer recognizes that the expression f1 + 1 matches the definition of gc and that gc is indexed, so it considers that index during execution plan construction. You can see this using EXPLAIN:

mysql> EXPLAIN SELECT * FROM t1 WHERE f1 + 1 > 9\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: t1
   partitions: NULL
         type: range
possible_keys: gc
          key: gc
      key_len: 5
          ref: NULL
         rows: 1
     filtered: 100.00
        Extra: Using index condition

In effect, the optimizer has replaced the expression f1 + 1 with the name of the generated column that matches the expression. That is also apparent in the rewritten query available in the extended EXPLAIN information displayed by SHOW WARNINGS:

mysql> SHOW WARNINGS\G
*************************** 1. row ***************************
  Level: Note
   Code: 1003
Message: /* select#1 */ select `test`.`t1`.`f1` AS `f1`,`test`.`t1`.`gc`
         AS `gc` from `test`.`t1` where (`test`.`t1`.`gc` > 9)

The following restrictions and conditions apply to the optimizer's use of generated column indexes:

  • For a query expression to match a generated column definition, the expression must be identical and it must have the same result type. For example, if the generated column expression is f1 + 1, the optimizer will not recognize a match if the query uses 1 + f1, or if f1 + 1 (an integer expression) is compared with a string.

  • The optimization applies to these operators: =, <, <=, >, >=, BETWEEN, and IN().

    For operators other than BETWEEN and IN(), either operand can be replaced by a matching generated column. For BETWEEN and IN(), only the first argument can be replaced by a matching generated column, and the other arguments must have the same result type. BETWEEN and IN() are not yet supported for comparisons involving JSON values.

  • The generated column must be defined as an expression that contains at least a function call or one of the operators mentioned in the preceding item. The expression cannot consist of a simple reference to another column. For example, gc INT AS (f1) STORED consists only of a column reference, so indexes on gc are not considered.

  • For comparisons of strings to indexed generated columns that compute a value from a JSON function that returns a quoted string, JSON_UNQUOTE() is needed in the column definition to remove the extra quotes from the function value. (For direct comparison of a string to the function result, the JSON comparator handles quote removal, but this does not occur for index lookups.) For example, instead of writing a column definition like this:

    doc_name TEXT AS (JSON_EXTRACT(jdoc, '$.name')) STORED
    

    Write it like this:

    doc_name TEXT AS (JSON_UNQUOTE(JSON_EXTRACT(jdoc, '$.name'))) STORED
    

    With the latter definition, the optimizer can detect a match for both of these comparisons:

    ... WHERE JSON_EXTRACT(jdoc, '$.name') = 'some_string' ...
    ... WHERE JSON_UNQUOTE(JSON_EXTRACT(jdoc, '$.name')) = 'some_string' ...
    

    Without JSON_UNQUOTE() in the column definition, the optimizer detects a match only for the first of those comparisons.

  • If the optimizer fails to choose the desired index, an index hint can be used to force the optimizer to make a different choice.

8.4 Optimizing Database Structure

In your role as a database designer, look for the most efficient way to organize your schemas, tables, and columns. As when tuning application code, you minimize I/O, keep related items together, and plan ahead so that performance stays high as the data volume increases. Starting with an efficient database design makes it easier for team members to write high-performing application code, and makes the database likely to endure as applications evolve and are rewritten.

8.4.1 Optimizing Data Size

Design your tables to minimize their space on the disk. This can result in huge improvements by reducing the amount of data written to and read from disk. Smaller tables normally require less main memory while their contents are being actively processed during query execution. Any space reduction for table data also results in smaller indexes that can be processed faster.

MySQL supports many different storage engines (table types) and row formats. For each table, you can decide which storage and indexing method to use. Choosing the proper table format for your application can give you a big performance gain. See Chapter 14, The InnoDB Storage Engine, and Chapter 15, Alternative Storage Engines.

You can get better performance for a table and minimize storage space by using the techniques listed here:

Table Columns

  • Use the most efficient (smallest) data types possible. MySQL has many specialized types that save disk space and memory. For example, use the smaller integer types if possible to get smaller tables. MEDIUMINT is often a better choice than INT because a MEDIUMINT column uses 25% less space.

  • Declare columns to be NOT NULL if possible. It makes SQL operations faster, by enabling better use of indexes and eliminating overhead for testing whether each value is NULL. You also save some storage space, one bit per column. If you really need NULL values in your tables, use them. Just avoid the default setting that allows NULL values in every column.

Row Format

  • In MySQL 5.7.8 and earlier, InnoDB tables are created in the COMPACT row format by default. As of MySQL 5.7.9, the default row format is DYNAMIC, and the default row format is configurable using the innodb_default_row_format configuration option.

    To request a row format other than the DYNAMIC row format, you can configure innodb_default_row_format or specify the ROW_FORMAT option explicitly in a CREATE TABLE or ALTER TABLE statement.

    The compact row format decreases row storage space by about 20% at the cost of increasing CPU use for some operations. If your workload is a typical one that is limited by cache hit rates and disk speed it is likely to be faster. If it is a rare case that is limited by CPU speed, it might be slower.

    The compact InnoDB format also changes how CHAR columns containing utf8 or utf8mb4 data are stored. With ROW_FORMAT=REDUNDANT, a utf8 or utf8mb4 CHAR(N) column occupies the maximum character byte length × N bytes. Many languages can be written primarily using single-byte utf8 or utf8mb4 characters, so a fixed storage length often wastes space. With ROW_FORMAT=COMPACT, InnoDB allocates a variable amount of storage for these columns by stripping trailing spaces if necessary. The minimum storage length is kept as N bytes to facilitate in-place updates in typical cases. For more information, see Section 14.2.6.7, “Physical Row Structure”.

  • To minimize space even further by storing table data in compressed form, specify ROW_FORMAT=COMPRESSED when creating InnoDB tables, or run the myisampack command on an existing MyISAM table. (InnoDB tables compressed tables are readable and writable, while MyISAM compressed tables are read-only.)

  • For MyISAM tables, if you do not have any variable-length columns (VARCHAR, TEXT, or BLOB columns), a fixed-size row format is used. This is faster but may waste some space. See Section 15.2.3, “MyISAM Table Storage Formats”. You can hint that you want to have fixed length rows even if you have VARCHAR columns with the CREATE TABLE option ROW_FORMAT=FIXED.

Indexes

  • The primary index of a table should be as short as possible. This makes identification of each row easy and efficient. For InnoDB tables, the primary key columns are duplicated in each secondary index entry, so a short primary key saves considerable space if you have many secondary indexes.

  • Create only the indexes that you need to improve query performance. Indexes are good for retrieval, but slow down insert and update operations. If you access a table mostly by searching on a combination of columns, create a single composite index on them rather than a separate index for each column. The first part of the index should be the column most used. If you always use many columns when selecting from the table, the first column in the index should be the one with the most duplicates, to obtain better compression of the index.

  • If it is very likely that a long string column has a unique prefix on the first number of characters, it is better to index only this prefix, using MySQL's support for creating an index on the leftmost part of the column (see Section 13.1.14, “CREATE INDEX Syntax”). Shorter indexes are faster, not only because they require less disk space, but because they also give you more hits in the index cache, and thus fewer disk seeks. See Section 8.12.2, “Tuning Server Parameters”.

Joins

  • In some circumstances, it can be beneficial to split into two a table that is scanned very often. This is especially true if it is a dynamic-format table and it is possible to use a smaller static format table that can be used to find the relevant rows when scanning the table.

  • Declare columns with identical information in different tables with identical data types, to speed up joins based on the corresponding columns.

  • Keep column names simple, so that you can use the same name across different tables and simplify join queries. For example, in a table named customer, use a column name of name instead of customer_name. To make your names portable to other SQL servers, consider keeping them shorter than 18 characters.

Normalization

  • Normally, try to keep all data nonredundant (observing what is referred to in database theory as third normal form). Instead of repeating lengthy values such as names and addresses, assign them unique IDs, repeat these IDs as needed across multiple smaller tables, and join the tables in queries by referencing the IDs in the join clause.

  • If speed is more important than disk space and the maintenance costs of keeping multiple copies of data, for example in a business intelligence scenario where you analyze all the data from large tables, you can relax the normalization rules, duplicating information or creating summary tables to gain more speed.

8.4.2 Optimizing MySQL Data Types

8.4.2.1 Optimizing for Numeric Data

  • For unique IDs or other values that can be represented as either strings or numbers, prefer numeric columns to string columns. Since large numeric values can be stored in fewer bytes than the corresponding strings, it is faster and takes less memory to transfer and compare them.

  • If you are using numeric data, it is faster in many cases to access information from a database (using a live connection) than to access a text file. Information in the database is likely to be stored in a more compact format than in the text file, so accessing it involves fewer disk accesses. You also save code in your application because you can avoid parsing the text file to find line and column boundaries.

8.4.2.2 Optimizing for Character and String Types

For character and string columns, follow these guidelines:

  • Use binary collation order for fast comparison and sort operations, when you do not need language-specific collation features. You can use the BINARY operator to use binary collation within a particular query.

  • When comparing values from different columns, declare those columns with the same character set and collation wherever possible, to avoid string conversions while running the query.

  • For column values less than 8KB in size, use binary VARCHAR instead of BLOB. The GROUP BY and ORDER BY clauses can generate temporary tables, and these temporary tables can use the MEMORY storage engine if the original table does not contain any BLOB columns.

  • If a table contains string columns such as name and address, but many queries do not retrieve those columns, consider splitting the string columns into a separate table and using join queries with a foreign key when necessary. When MySQL retrieves any value from a row, it reads a data block containing all the columns of that row (and possibly other adjacent rows). Keeping each row small, with only the most frequently used columns, allows more rows to fit in each data block. Such compact tables reduce disk I/O and memory usage for common queries.

  • When you use a randomly generated value as a primary key in an InnoDB table, prefix it with an ascending value such as the current date and time if possible. When consecutive primary values are physically stored near each other, InnoDB can insert and retrieve them faster.

  • See Section 8.4.2.1, “Optimizing for Numeric Data” for reasons why a numeric column is usually preferable to an equivalent string column.

8.4.2.3 Optimizing for BLOB Types

  • When storing a large blob containing textual data, consider compressing it first. Do not use this technique when the entire table is compressed by InnoDB or MyISAM.

  • For a table with several columns, to reduce memory requirements for queries that do not use the BLOB column, consider splitting the BLOB column into a separate table and referencing it with a join query when needed.

  • Since the performance requirements to retrieve and display a BLOB value might be very different from other data types, you could put the BLOB-specific table on a different storage device or even a separate database instance. For example, to retrieve a BLOB might require a large sequential disk read that is better suited to a traditional hard drive than to an SSD device.

  • See Section 8.4.2.2, “Optimizing for Character and String Types” for reasons why a binary VARCHAR column is sometimes preferable to an equivalent BLOB column.

  • Rather than testing for equality against a very long text string, you can store a hash of the column value in a separate column, index that column, and test the hashed value in queries. (Use the MD5() or CRC32() function to produce the hash value.) Since hash functions can produce duplicate results for different inputs, you still include a clause AND blob_column = long_string_value in the query to guard against false matches; the performance benefit comes from the smaller, easily scanned index for the hashed values.

8.4.2.4 Using PROCEDURE ANALYSE

ANALYSE([max_elements[,max_memory]])

ANALYSE() examines the result from a query and returns an analysis of the results that suggests optimal data types for each column that may help reduce table sizes. To obtain this analysis, append PROCEDURE ANALYSE to the end of a SELECT statement:

SELECT ... FROM ... WHERE ... PROCEDURE ANALYSE([max_elements,[max_memory]])

For example:

SELECT col1, col2 FROM table1 PROCEDURE ANALYSE(10, 2000);

The results show some statistics for the values returned by the query, and propose an optimal data type for the columns. This can be helpful for checking your existing tables, or after importing new data. You may need to try different settings for the arguments so that PROCEDURE ANALYSE() does not suggest the ENUM data type when it is not appropriate.

The arguments are optional and are used as follows:

  • max_elements (default 256) is the maximum number of distinct values that ANALYSE() notices per column. This is used by ANALYSE() to check whether the optimal data type should be of type ENUM; if there are more than max_elements distinct values, then ENUM is not a suggested type.

  • max_memory (default 8192) is the maximum amount of memory that ANALYSE() should allocate per column while trying to find all distinct values.

A PROCEDURE clause is not permitted in a UNION statement.

8.4.3 Optimizing for Many Tables

Some techniques for keeping individual queries fast involve splitting data across many tables. When the number of tables runs into the thousands or even millions, the overhead of dealing with all these tables becomes a new performance consideration.

8.4.3.1 How MySQL Opens and Closes Tables

When you execute a mysqladmin status command, you should see something like this:

Uptime: 426 Running threads: 1 Questions: 11082
Reloads: 1 Open tables: 12

The Open tables value of 12 can be somewhat puzzling if you have only six tables.

MySQL is multi-threaded, so there may be many clients issuing queries for a given table simultaneously. To minimize the problem with multiple client sessions having different states on the same table, the table is opened independently by each concurrent session. This uses additional memory but normally increases performance. With MyISAM tables, one extra file descriptor is required for the data file for each client that has the table open. (By contrast, the index file descriptor is shared between all sessions.)

The table_open_cache and max_connections system variables affect the maximum number of files the server keeps open. If you increase one or both of these values, you may run up against a limit imposed by your operating system on the per-process number of open file descriptors. Many operating systems permit you to increase the open-files limit, although the method varies widely from system to system. Consult your operating system documentation to determine whether it is possible to increase the limit and how to do so.

table_open_cache is related to max_connections. For example, for 200 concurrent running connections, specify a table cache size of at least 200 * N, where N is the maximum number of tables per join in any of the queries which you execute. You must also reserve some extra file descriptors for temporary tables and files.

Make sure that your operating system can handle the number of open file descriptors implied by the table_open_cache setting. If table_open_cache is set too high, MySQL may run out of file descriptors and refuse connections, fail to perform queries, and be very unreliable.

You should also take into account the fact that the MyISAM storage engine needs two file descriptors for each unique open table. For a partitioned MyISAM table, two file descriptors are required for each partition of the opened table. (Note further that when MyISAM opens a partitioned table, it opens every partition of this table, whether or not a given partition is actually used. See MyISAM and partition file descriptor usage.) You can increase the number of file descriptors available to MySQL using the --open-files-limit startup option to mysqld. See Section B.5.2.18, “'File' Not Found and Similar Errors”.

The cache of open tables is kept at a level of table_open_cache entries. The server autosizes the cache size at startup. To set the size explicitly, set the table_open_cache system variable at startup. Note that MySQL may temporarily open more tables than this to execute queries.

MySQL closes an unused table and removes it from the table cache under the following circumstances:

  • When the cache is full and a thread tries to open a table that is not in the cache.

  • When the cache contains more than table_open_cache entries and a table in the cache is no longer being used by any threads.

  • When a table flushing operation occurs. This happens when someone issues a FLUSH TABLES statement or executes a mysqladmin flush-tables or mysqladmin refresh command.

When the table cache fills up, the server uses the following procedure to locate a cache entry to use:

  • Tables that are not currently in use are released, beginning with the table least recently used.

  • If a new table needs to be opened, but the cache is full and no tables can be released, the cache is temporarily extended as necessary. When the cache is in a temporarily extended state and a table goes from a used to unused state, the table is closed and released from the cache.

A MyISAM table is opened for each concurrent access. This means the table needs to be opened twice if two threads access the same table or if a thread accesses the table twice in the same query (for example, by joining the table to itself). Each concurrent open requires an entry in the table cache. The first open of any MyISAM table takes two file descriptors: one for the data file and one for the index file. Each additional use of the table takes only one file descriptor for the data file. The index file descriptor is shared among all threads.

If you are opening a table with the HANDLER tbl_name OPEN statement, a dedicated table object is allocated for the thread. This table object is not shared by other threads and is not closed until the thread calls HANDLER tbl_name CLOSE or the thread terminates. When this happens, the table is put back in the table cache (if the cache is not full). See Section 13.2.4, “HANDLER Syntax”.

You can determine whether your table cache is too small by checking the mysqld status variable Opened_tables, which indicates the number of table-opening operations since the server started:

mysql> SHOW GLOBAL STATUS LIKE 'Opened_tables';
+---------------+-------+
| Variable_name | Value |
+---------------+-------+
| Opened_tables | 2741  |
+---------------+-------+

If the value is very large or increases rapidly, even when you have not issued many FLUSH TABLES statements, increase the table cache size. See Section 5.1.4, “Server System Variables”, and Section 5.1.6, “Server Status Variables”.

8.4.3.2 Disadvantages of Creating Many Tables in the Same Database

If you have many MyISAM tables in the same database directory, open, close, and create operations are slow. If you execute SELECT statements on many different tables, there is a little overhead when the table cache is full, because for every table that has to be opened, another must be closed. You can reduce this overhead by increasing the number of entries permitted in the table cache.

8.4.4 Internal Temporary Table Use in MySQL

In some cases, the server creates internal temporary tables while processing statements. Users have no direct control over when this occurs.

The server creates temporary tables under conditions such as these:

  • Evaluation of UNION statements, with some exceptions described later.

  • Evaluation of some views, such those that use the TEMPTABLE algorithm, UNION, or aggregation.

  • Evaluation of derived tables (subqueries in the FROM clause).

  • Tables created for subquery or semi-join materialization (see Section 8.2.1.18, “Subquery Optimization”).

  • Evaluation of statements that contain an ORDER BY clause and a different GROUP BY clause, or for which the ORDER BY or GROUP BY contains columns from tables other than the first table in the join queue.

  • Evaluation of DISTINCT combined with ORDER BY may require a temporary table.

  • For queries that use the SQL_SMALL_RESULT option, MySQL uses an in-memory temporary table, unless the query also contains elements (described later) that require on-disk storage.

  • Evaluation of multiple-table UPDATE statements.

  • Evaluation of GROUP_CONCAT() or COUNT(DISTINCT) expressions.

As of MySQL 5.7.3, the server does not use a temporary table for UNION statements that meet certain qualifications. Instead, it retains from temporary table creation only the data structures necessary to perform result column typecasting. The table is not fully instantiated and no rows are written to or read from it; rows are sent directly to the client. The result is reduced memory and disk requirements, and smaller delay before the first row is sent to the client because the server need not wait until the last query block is executed. EXPLAIN and optimizer trace output reflects this execution strategy: The UNION RESULT query block is not present because that block corresponds to the part that reads from the temporary table.

These conditions qualify a UNION for evaluation without a temporary table:

  • The union is UNION ALL, not UNION or UNION DISTINCT.

  • There is no global ORDER BY clause.

  • The union is not the top-level query block of an {INSERT | REPLACE} ... SELECT ... statement.

To determine whether a statement requires a temporary table, use EXPLAIN and check the Extra column to see whether it says Using temporary (see Section 8.8.1, “Optimizing Queries with EXPLAIN”). EXPLAIN will not necessarily say Using temporary for derived or materialized temporary tables.

Storage Engines Used for Temporary Tables

An internal temporary table can be held in memory and processed by the MEMORY storage engine, or stored on disk by the InnoDB or MyISAM storage engine.

If an internal temporary table is created as an in-memory table but becomes too large, MySQL automatically converts it to an on-disk table. The maximum size for in-memory temporary tables is the minimum of the tmp_table_size and max_heap_table_size values. This differs from MEMORY tables explicitly created with CREATE TABLE: For such tables, only the max_heap_table_size system variable determines how large the table is permitted to grow and there is no conversion to on-disk format.

As of MySQL 5.7.5, the internal_tmp_disk_storage_engine system variable determines which storage engine the server uses to manage on-disk internal temporary tables. The value can be INNODB or MYISAM. The default in MySQL 5.7.5 is MYISAM. As of MySQL 5.7.6, the default is INNODB. Before MySQL 5.7.5, the server always uses MyISAM for on-disk internal temporary tables.

Note

Under internal_tmp_disk_storage_engine=INNODB, queries that generate temporary tables that exceed InnoDB row or column limits will return Row size too large or Too many columns errors. The workaround is to set internal_tmp_disk_storage_engine to MYISAM.

Some conditions prevent the use of an in-memory temporary table, in which case the server uses an on-disk table instead:

  • Presence of a BLOB or TEXT column in the table

  • Presence of any string column in a GROUP BY or DISTINCT clause larger than 512 bytes for binary strings or 512 characters for nonbinary strings. (Before MySQL 5.7.3, the limit is 512 bytes regardless of string type.)

  • Presence of any string column with a maximum length larger than 512 (bytes for binary strings, characters for nonbinary strings) in the SELECT list, if UNION or UNION ALL is used

  • The SHOW COLUMNS and DESCRIBE statements use BLOB as the type for some columns, thus the temporary table used for the results is an on-disk table.

When the server creates an internal temporary table (either in memory or on disk), it increments the Created_tmp_tables status variable. If the server creates the table on disk (either initially or by converting an in-memory table) it increments the Created_tmp_disk_tables status variable.

Temporary Table Storage Format

In-memory temporary tables are managed by the MEMORY storage engine, which uses fixed-length row format. VARCHAR and VARBINARY column values are padded to the maximum column length, in effect storing them as CHAR and BINARY columns.

On-disk temporary tables are managed by the InnoDB or MyISAM storage engine (depending on the internal_tmp_disk_storage_engine setting). Both engines store temporary tables using dynamic-width row format. Columns take only as much storage as needed, which reduces disk I/O and space requirements, and processing time compared to on-disk tables that use fixed-length rows.

For statements that initially create an internal temporary table in memory, then convert it to an on-disk table, better performance might be achieved by skipping the conversion step and creating the table on disk to begin with. The big_tables system variable can be used to force disk storage of internal temporary tables.

8.5 Optimizing for InnoDB Tables

InnoDB is the storage engine that MySQL customers typically use in production databases where reliability and concurrency are important. InnoDB is the default storage engine in MySQL. This section explains how to optimize database operations for InnoDB tables.

8.5.1 Optimizing Storage Layout for InnoDB Tables

  • Once your data reaches a stable size, or a growing table has increased by tens or some hundreds of megabytes, consider using the OPTIMIZE TABLE statement to reorganize the table and compact any wasted space. The reorganized tables require less disk I/O to perform full table scans. This is a straightforward technique that can improve performance when other techniques such as improving index usage or tuning application code are not practical.

    OPTIMIZE TABLE copies the data part of the table and rebuilds the indexes. The benefits come from improved packing of data within indexes, and reduced fragmentation within the tablespaces and on disk. The benefits vary depending on the data in each table. You may find that there are significant gains for some and not for others, or that the gains decrease over time until you next optimize the table. This operation can be slow if the table is large or if the indexes being rebuilt do not fit into the buffer pool. The first run after adding a lot of data to a table is often much slower than later runs.

  • In InnoDB, having a long PRIMARY KEY (either a single column with a lengthy value, or several columns that form a long composite value) wastes a lot of disk space. The primary key value for a row is duplicated in all the secondary index records that point to the same row. (See Section 14.2.6, “InnoDB Table and Index Structures”.) Create an AUTO_INCREMENT column as the primary key if your primary key is long, or index a prefix of a long VARCHAR column instead of the entire column.

  • Use the VARCHAR data type instead of CHAR to store variable-length strings or for columns with many NULL values. A CHAR(N) column always takes N characters to store data, even if the string is shorter or its value is NULL. Smaller tables fit better in the buffer pool and reduce disk I/O.

    When using COMPACT row format (the default InnoDB format) and variable-length character sets, such as utf8 or sjis, CHAR(N) columns occupy a variable amount of space, but still at least N bytes.

  • For tables that are big, or contain lots of repetitive text or numeric data, consider using COMPRESSED row format. Less disk I/O is required to bring data into the buffer pool, or to perform full table scans. Before making a permanent decision, measure the amount of compression you can achieve by using COMPRESSED versus COMPACT row format.

8.5.2 Optimizing InnoDB Transaction Management

To optimize InnoDB transaction processing, find the ideal balance between the performance overhead of transactional features and the workload of your server. For example, an application might encounter performance issues if it commits thousands of times per second, and different performance issues if it commits only every 2-3 hours.

  • The default MySQL setting AUTOCOMMIT=1 can impose performance limitations on a busy database server. Where practical, wrap several related DML operations into a single transaction, by issuing SET AUTOCOMMIT=0 or a START TRANSACTION statement, followed by a COMMIT statement after making all the changes.

    InnoDB must flush the log to disk at each transaction commit if that transaction made modifications to the database. When each change is followed by a commit (as with the default autocommit setting), the I/O throughput of the storage device puts a cap on the number of potential operations per second.

  • Alternatively, for transactions that consist only of a single SELECT statement, turning on AUTOCOMMIT helps InnoDB to recognize read-only transactions and optimize them. See Section 8.5.3, “Optimizing InnoDB Read-Only Transactions” for requirements.

  • Avoid performing rollbacks after inserting, updating, or deleting huge numbers of rows. If a big transaction is slowing down server performance, rolling it back can make the problem worse, potentially taking several times as long to perform as the original DML operations. Killing the database process does not help, because the rollback starts again on server startup.

    To minimize the chance of this issue occurring:

    • Increase the size of the buffer pool so that all the DML changes can be cached rather than immediately written to disk.

    • Set innodb_change_buffering=all so that update and delete operations are buffered in addition to inserts.

    • Consider issuing COMMIT statements periodically during the big DML operation, possibly breaking a single delete or update into multiple statements that operate on smaller numbers of rows.

    To get rid of a runaway rollback once it occurs, increase the buffer pool so that the rollback becomes CPU-bound and runs fast, or kill the server and restart with innodb_force_recovery=3, as explained in Section 14.16.1, “The InnoDB Recovery Process”.

    This issue is expected to be less prominent in MySQL 5.5 and higher because the default setting innodb_change_buffering=all allows update and delete operations to be cached in memory, making them faster to perform in the first place, and also faster to roll back if needed. Make sure to use this parameter setting on servers that process long-running transactions with many inserts, updates, or deletes.

  • If you can afford the loss of some of the latest committed transactions if a crash occurs, you can set the innodb_flush_log_at_trx_commit parameter to 0. InnoDB tries to flush the log once per second anyway, although the flush is not guaranteed. Also, set the value of innodb_support_xa to 0, which will reduce the number of disk flushes due to synchronizing on disk data and the binary log.

    Note

    innodb_support_xa is deprecated and will be removed in a future release. As of MySQL 5.7.10, InnoDB support for two-phase commit in XA transactions is always enabled and disabling innodb_support_xa is no longer permitted.

  • When rows are modified or deleted, the rows and associated undo logs are not physically removed immediately, or even immediately after the transaction commits. The old data is preserved until transactions that started earlier or concurrently are finished, so that those transactions can access the previous state of modified or deleted rows. Thus, a long-running transaction can prevent InnoDB from purging data that was changed by a different transaction.

  • When rows are modified or deleted within a long-running transaction, other transactions using the READ COMMITTED and REPEATABLE READ isolation levels have to do more work to reconstruct the older data if they read those same rows.

  • When a long-running transaction modifies a table, queries against that table from other transactions do not make use of the covering index technique. Queries that normally could retrieve all the result columns from a secondary index, instead look up the appropriate values from the table data.

    If secondary index pages are found to have a PAGE_MAX_TRX_ID that is too new, or if records in the secondary index are delete-marked, InnoDB may need to look up records using a clustered index.

8.5.3 Optimizing InnoDB Read-Only Transactions

InnoDB can avoid the overhead associated with setting up the transaction ID (TRX_ID field) for transactions that are known to be read-only. A transaction ID is only needed for a transaction that might perform write operations or locking reads such as SELECT ... FOR UPDATE. Eliminating unnecessary transaction IDs reduces the size of internal data structures that are consulted each time a query or DML statement constructs a read view.

InnoDB detects read-only transactions when:

  • The transaction is started with the START TRANSACTION READ ONLY statement. In this case, attempting to make changes to the database (for InnoDB, MyISAM, or other types of tables) causes an error, and the transaction continues in read-only state:

    ERROR 1792 (25006): Cannot execute statement in a READ ONLY transaction.
    

    You can still make changes to session-specific temporary tables in a read-only transaction, or issue locking queries for them, because those changes and locks are not visible to any other transaction.

  • The autocommit setting is turned on, so that the transaction is guaranteed to be a single statement, and the single statement making up the transaction is a non-locking SELECT statement. That is, a SELECT that does not use a FOR UPDATE or LOCK IN SHARED MODE clause.

  • The transaction is started without the READ ONLY option, but no updates or statements that explicitly lock rows have been executed yet. Until updates or explicit locks are required, a transaction stays in read-only mode.

Thus, for a read-intensive application such as a report generator, you can tune a sequence of InnoDB queries by grouping them inside START TRANSACTION READ ONLY and COMMIT, or by turning on the autocommit setting before running the SELECT statements, or simply by avoiding any DML statements interspersed with the queries.

For information about START TRANSACTION and autocommit, see Section 13.3.1, “START TRANSACTION, COMMIT, and ROLLBACK Syntax”.

Note

Transactions that qualify as auto-commit, non-locking, and read-only (AC-NL-RO) are kept out of certain internal InnoDB data structures and are therefore not listed in SHOW ENGINE INNODB STATUS output.

8.5.4 Optimizing InnoDB Redo Logging

Consider the following guidelines for optimizing redo logging:

  • Make your redo log files big, even as big as the buffer pool. When InnoDB has written the redo log files full, it must write the modified contents of the buffer pool to disk in a checkpoint. Small redo log files cause many unnecessary disk writes. Although historically big redo log files caused lengthy recovery times, recovery is now much faster and you can confidently use large redo log files.

    The size and number of redo log files are configured using the innodb_log_file_size and innodb_log_files_in_group configuration options. For information about modifying an existing redo log file configuration, see Section 14.5.2, “Changing the Number or Size of InnoDB Redo Log Files”.

  • Consider increasing the size of the log_buffer. A large log buffer enables large transactions to run without a need to write the log to disk before the transactions commit. Thus, if you have transactions that update, insert, or delete many rows, making the log buffer larger saves disk I/O. Log buffer size is configured using the innodb_log_buffer_size configuration option.

8.5.5 Bulk Data Loading for InnoDB Tables

These performance tips supplement the general guidelines for fast inserts in Section 8.2.2.1, “Speed of INSERT Statements”.

  • When importing data into InnoDB, turn off autocommit mode, because it performs a log flush to disk for every insert. To disable autocommit during your import operation, surround it with SET autocommit and COMMIT statements:

    SET autocommit=0;
    ... SQL import statements ...
    COMMIT;
    

    The mysqldump option --opt creates dump files that are fast to import into an InnoDB table, even without wrapping them with the SET autocommit and COMMIT statements.

  • If you have UNIQUE constraints on secondary keys, you can speed up table imports by temporarily turning off the uniqueness checks during the import session:

    SET unique_checks=0;
    ... SQL import statements ...
    SET unique_checks=1;
    

    For big tables, this saves a lot of disk I/O because InnoDB can use its change buffer to write secondary index records in a batch. Be certain that the data contains no duplicate keys.

  • If you have FOREIGN KEY constraints in your tables, you can speed up table imports by turning off the foreign key checks for the duration of the import session:

    SET foreign_key_checks=0;
    ... SQL import statements ...
    SET foreign_key_checks=1;
    

    For big tables, this can save a lot of disk I/O.

  • Use the multiple-row INSERT syntax to reduce communication overhead between the client and the server if you need to insert many rows:

    INSERT INTO yourtable VALUES (1,2), (5,5), ...;
    

    This tip is valid for inserts into any table, not just InnoDB tables.

  • When doing bulk inserts into tables with auto-increment columns, set innodb_autoinc_lock_mode to 2 instead of the default value 1. See Section 14.6.5, “AUTO_INCREMENT Handling in InnoDB” for details.

  • For optimal performance when loading data into an InnoDB FULLTEXT index, follow this set of steps:

    • Define a column FTS_DOC_ID at table creation time, of type BIGINT UNSIGNED NOT NULL, with a unique index named FTS_DOC_ID_INDEX. For example:

      CREATE TABLE t1 (
      FTS_DOC_ID BIGINT unsigned NOT NULL AUTO_INCREMENT,
      title varchar(255) NOT NULL DEFAULT ”,
      text mediumtext NOT NULL,
      PRIMARY KEY (`FTS_DOC_ID`)
      ) ENGINE=InnoDB DEFAULT CHARSET=latin1;
      CREATE UNIQUE INDEX FTS_DOC_ID_INDEX on t1(FTS_DOC_ID);
                    
    • Load the data into the table.

    • Create the FULLTEXT index after the data is loaded.

    Note

    When adding FTS_DOC_ID column at table creation time, ensure that the FTS_DOC_ID column is updated when the FULLTEXT indexed column is updated, as the FTS_DOC_ID must increase monotonically with each INSERT or UPDATE. If you choose not to add the FTS_DOC_ID at table creation time and have InnoDB manage DOC IDs for you, InnoDB will add the FTS_DOC_ID as a hidden column with the next CREATE FULLTEXT INDEX call. This approach, however, requires a table rebuild which will impact performance.

8.5.6 Optimizing InnoDB Queries

To tune queries for InnoDB tables, create an appropriate set of indexes on each table. See Section 8.3.1, “How MySQL Uses Indexes” for details. Follow these guidelines for InnoDB indexes:

  • Because each InnoDB table has a primary key (whether you request one or not), specify a set of primary key columns for each table, columns that are used in the most important and time-critical queries.

  • Do not specify too many or too long columns in the primary key, because these column values are duplicated in each secondary index. When an index contains unnecessary data, the I/O to read this data and memory to cache it reduce the performance and scalability of the server.

  • Do not create a separate secondary index for each column, because each query can only make use of one index. Indexes on rarely tested columns or columns with only a few different values might not be helpful for any queries. If you have many queries for the same table, testing different combinations of columns, try to create a small number of concatenated indexes rather than a large number of single-column indexes. If an index contains all the columns needed for the result set (known as a covering index), the query might be able to avoid reading the table data at all.

  • If an indexed column cannot contain any NULL values, declare it as NOT NULL when you create the table. The optimizer can better determine which index is most effective to use for a query, when it knows whether each column contains NULL values.

  • You can optimize single-query transactions for InnoDB tables, using the technique in Section 8.5.3, “Optimizing InnoDB Read-Only Transactions”.

  • If you often have recurring queries for tables that are not updated frequently, enable the query cache:

    [mysqld]
    query_cache_type = 1
    query_cache_size = 10M
    

8.5.7 Optimizing InnoDB DDL Operations

  • For DDL operations on tables and indexes (CREATE, ALTER, and DROP statements), the most significant aspect for InnoDB tables is that creating and dropping secondary indexes is much faster in MySQL 5.5 and higher, than in MySQL 5.1 and before. See InnoDB Fast Index Creation for details.

  • Fast index creation makes it faster in some cases to drop an index before loading data into a table, then re-create the index after loading the data.

  • Use TRUNCATE TABLE to empty a table, not DELETE FROM tbl_name. Foreign key constraints can make a TRUNCATE statement work like a regular DELETE statement, in which case a sequence of commands like DROP TABLE and CREATE TABLE might be fastest.

  • Because the primary key is integral to the storage layout of each InnoDB table, and changing the definition of the primary key involves reorganizing the whole table, always set up the primary key as part of the CREATE TABLE statement, and plan ahead so that you do not need to ALTER or DROP the primary key afterward.

8.5.8 Optimizing InnoDB Disk I/O

If you follow the best practices for database design and the tuning techniques for SQL operations, but your database is still slowed by heavy disk I/O activity, explore these low-level techniques related to disk I/O. If the Unix top tool or the Windows Task Manager shows that the CPU usage percentage with your workload is less than 70%, your workload is probably disk-bound.

  • When table data is cached in the InnoDB buffer pool, it can be accessed repeatedly by queries without requiring any disk I/O. Specify the size of the buffer pool with the innodb_buffer_pool_size option. This memory area is important enough that busy databases often specify a size approximately 80% of the amount of physical memory. For more information, see Section 8.10.1, “The InnoDB Buffer Pool”.

  • In some versions of GNU/Linux and Unix, flushing files to disk with the Unix fsync() call (which InnoDB uses by default) and similar methods is surprisingly slow. If database write performance is an issue, conduct benchmarks with the innodb_flush_method parameter set to O_DSYNC.

  • When using the InnoDB storage engine on Solaris 10 for x86_64 architecture (AMD Opteron), use direct I/O for InnoDB-related files, to avoid degradation of InnoDB performance. To use direct I/O for an entire UFS file system used for storing InnoDB-related files, mount it with the forcedirectio option; see mount_ufs(1M). (The default on Solaris 10/x86_64 is not to use this option.) To apply direct I/O only to InnoDB file operations rather than the whole file system, set innodb_flush_method = O_DIRECT. With this setting, InnoDB calls directio() instead of fcntl() for I/O to data files (not for I/O to log files).

  • When using the InnoDB storage engine with a large innodb_buffer_pool_size value on any release of Solaris 2.6 and up and any platform (sparc/x86/x64/amd64), conduct benchmarks with InnoDB data files and log files on raw devices or on a separate direct I/O UFS file system, using the forcedirectio mount option as described earlier. (It is necessary to use the mount option rather than setting innodb_flush_method if you want direct I/O for the log files.) Users of the Veritas file system VxFS should use the convosync=direct mount option.

    Do not place other MySQL data files, such as those for MyISAM tables, on a direct I/O file system. Executables or libraries must not be placed on a direct I/O file system.

  • If you have additional storage devices available to set up a RAID configuration or symbolic links to different disks, Section 8.12.3, “Optimizing Disk I/O” for additional low-level I/O tips.

  • If throughput drops periodically because of InnoDB checkpoint operations, consider increasing the value of the innodb_io_capacity configuration option. Higher values cause more frequent flushing, avoiding the backlog of work that can cause dips in throughput.

  • If the system is not falling behind with InnoDB flushing operations, consider lowering the value of the innodb_io_capacity configuration option. Typically, you keep this option value as low as practical, but not so low that it causes periodic drops in throughput as mentioned in the preceding bullet. In a typical scenario where you could lower the option value, you might see a combination like this in the output from SHOW ENGINE INNODB STATUS:

    • History list length low, below a few thousand.

    • Insert buffer merges close to rows inserted.

    • Modified pages in buffer pool consistently well below innodb_max_dirty_pages_pct of the buffer pool. (Measure at a time when the server is not doing bulk inserts; it is normal during bulk inserts for the modified pages percentage to rise significantly.)

    • Log sequence number - Last checkpoint is at less than 7/8 or ideally less than 6/8 of the total size of the InnoDB log files.

  • Other InnoDB configuration options to consider when tuning I/O-bound workloads include innodb_adaptive_flushing, innodb_change_buffer_max_size, innodb_change_buffering, innodb_flush_neighbors, innodb_log_buffer_size, innodb_log_file_size, innodb_lru_scan_depth, innodb_max_dirty_pages_pct, innodb_max_purge_lag, innodb_open_files, innodb_page_size, innodb_random_read_ahead, innodb_read_ahead_threshold, innodb_read_io_threads, innodb_rollback_segments, innodb_write_io_threads, and sync_binlog.

8.5.9 Optimizing InnoDB Configuration Variables

Different settings work best for servers with light, predictable loads, versus servers that are running near full capacity all the time, or that experience spikes of high activity.

Because the InnoDB storage engine performs many of its optimizations automatically, many performance-tuning tasks involve monitoring to ensure that the database is performing well, and changing configuration options when performance drops. See Section 14.14, “InnoDB Integration with MySQL Performance Schema” for information about detailed InnoDB performance monitoring.

The main configuration steps you can perform include:

8.5.10 Optimizing InnoDB for Systems with Many Tables

  • InnoDB computes index cardinality values for a table the first time that table is accessed after startup, instead of storing such values in the table. This step can take significant time on systems that partition the data into many tables. Since this overhead only applies to the initial table open operation, to warm up a table for later use, access it immediately after startup by issuing a statement such as SELECT 1 FROM tbl_name LIMIT 1.

8.6 Optimizing for MyISAM Tables

The MyISAM storage engine performs best with read-mostly data or with low-concurrency operations, because table locks limit the ability to perform simultaneous updates. In MySQL, InnoDB is the default storage engine rather than MyISAM.

8.6.1 Optimizing MyISAM Queries

Some general tips for speeding up queries on MyISAM tables:

  • To help MySQL better optimize queries, use ANALYZE TABLE or run myisamchk --analyze on a table after it has been loaded with data. This updates a value for each index part that indicates the average number of rows that have the same value. (For unique indexes, this is always 1.) MySQL uses this to decide which index to choose when you join two tables based on a nonconstant expression. You can check the result from the table analysis by using SHOW INDEX FROM tbl_name and examining the Cardinality value. myisamchk --description --verbose shows index distribution information.

  • To sort an index and data according to an index, use myisamchk --sort-index --sort-records=1 (assuming that you want to sort on index 1). This is a good way to make queries faster if you have a unique index from which you want to read all rows in order according to the index. The first time you sort a large table this way, it may take a long time.

  • Try to avoid complex SELECT queries on MyISAM tables that are updated frequently, to avoid problems with table locking that occur due to contention between readers and writers.

  • MyISAM supports concurrent inserts: If a table has no free blocks in the middle of the data file, you can INSERT new rows into it at the same time that other threads are reading from the table. If it is important to be able to do this, consider using the table in ways that avoid deleting rows. Another possibility is to run OPTIMIZE TABLE to defragment the table after you have deleted a lot of rows from it. This behavior is altered by setting the concurrent_insert variable. You can force new rows to be appended (and therefore permit concurrent inserts), even in tables that have deleted rows. See Section 8.11.3, “Concurrent Inserts”.

  • For MyISAM tables that change frequently, try to avoid all variable-length columns (VARCHAR, BLOB, and TEXT). The table uses dynamic row format if it includes even a single variable-length column. See Chapter 15, Alternative Storage Engines.

  • It is normally not useful to split a table into different tables just because the rows become large. In accessing a row, the biggest performance hit is the disk seek needed to find the first byte of the row. After finding the data, most modern disks can read the entire row fast enough for most applications. The only cases where splitting up a table makes an appreciable difference is if it is a MyISAM table using dynamic row format that you can change to a fixed row size, or if you very often need to scan the table but do not need most of the columns. See Chapter 15, Alternative Storage Engines.

  • Use ALTER TABLE ... ORDER BY expr1, expr2, ... if you usually retrieve rows in expr1, expr2, ... order. By using this option after extensive changes to the table, you may be able to get higher performance.

  • If you often need to calculate results such as counts based on information from a lot of rows, it may be preferable to introduce a new table and update the counter in real time. An update of the following form is very fast:

    UPDATE tbl_name SET count_col=count_col+1 WHERE key_col=constant;
    

    This is very important when you use MySQL storage engines such as MyISAM that has only table-level locking (multiple readers with single writers). This also gives better performance with most database systems, because the row locking manager in this case has less to do.

  • Use OPTIMIZE TABLE periodically to avoid fragmentation with dynamic-format MyISAM tables. See Section 15.2.3, “MyISAM Table Storage Formats”.

  • Declaring a MyISAM table with the DELAY_KEY_WRITE=1 table option makes index updates faster because they are not flushed to disk until the table is closed. The downside is that if something kills the server while such a table is open, you must ensure that the table is okay by running the server with the --myisam-recover-options option, or by running myisamchk before restarting the server. (However, even in this case, you should not lose anything by using DELAY_KEY_WRITE, because the key information can always be generated from the data rows.)

  • Strings are automatically prefix- and end-space compressed in MyISAM indexes. See Section 13.1.14, “CREATE INDEX Syntax”.

  • You can increase performance by caching queries or answers in your application and then executing many inserts or updates together. Locking the table during this operation ensures that the index cache is only flushed once after all updates. You can also take advantage of MySQL's query cache to achieve similar results; see Section 8.10.3, “The MySQL Query Cache”.

8.6.2 Bulk Data Loading for MyISAM Tables

These performance tips supplement the general guidelines for fast inserts in Section 8.2.2.1, “Speed of INSERT Statements”.

  • For a MyISAM table, you can use concurrent inserts to add rows at the same time that SELECT statements are running, if there are no deleted rows in middle of the data file. See Section 8.11.3, “Concurrent Inserts”.

  • With some extra work, it is possible to make LOAD DATA INFILE run even faster for a MyISAM table when the table has many indexes. Use the following procedure:

    1. Execute a FLUSH TABLES statement or a mysqladmin flush-tables command.

    2. Use myisamchk --keys-used=0 -rq /path/to/db/tbl_name to remove all use of indexes for the table.

    3. Insert data into the table with LOAD DATA INFILE. This does not update any indexes and therefore is very fast.

    4. If you intend only to read from the table in the future, use myisampack to compress it. See Section 15.2.3.3, “Compressed Table Characteristics”.

    5. Re-create the indexes with myisamchk -rq /path/to/db/tbl_name. This creates the index tree in memory before writing it to disk, which is much faster than updating the index during LOAD DATA INFILE because it avoids lots of disk seeks. The resulting index tree is also perfectly balanced.

    6. Execute a FLUSH TABLES statement or a mysqladmin flush-tables command.

    LOAD DATA INFILE performs the preceding optimization automatically if the MyISAM table into which you insert data is empty. The main difference between automatic optimization and using the procedure explicitly is that you can let myisamchk allocate much more temporary memory for the index creation than you might want the server to allocate for index re-creation when it executes the LOAD DATA INFILE statement.

    You can also disable or enable the nonunique indexes for a MyISAM table by using the following statements rather than myisamchk. If you use these statements, you can skip the FLUSH TABLE operations:

    ALTER TABLE tbl_name DISABLE KEYS;
    ALTER TABLE tbl_name ENABLE KEYS;
    
  • To speed up INSERT operations that are performed with multiple statements for nontransactional tables, lock your tables:

    LOCK TABLES a WRITE;
    INSERT INTO a VALUES (1,23),(2,34),(4,33);
    INSERT INTO a VALUES (8,26),(6,29);
    ...
    UNLOCK TABLES;
    

    This benefits performance because the index buffer is flushed to disk only once, after all INSERT statements have completed. Normally, there would be as many index buffer flushes as there are INSERT statements. Explicit locking statements are not needed if you can insert all rows with a single INSERT.

    Locking also lowers the total time for multiple-connection tests, although the maximum wait time for individual connections might go up because they wait for locks. Suppose that five clients attempt to perform inserts simultaneously as follows:

    • Connection 1 does 1000 inserts

    • Connections 2, 3, and 4 do 1 insert

    • Connection 5 does 1000 inserts

    If you do not use locking, connections 2, 3, and 4 finish before 1 and 5. If you use locking, connections 2, 3, and 4 probably do not finish before 1 or 5, but the total time should be about 40% faster.

    INSERT, UPDATE, and DELETE operations are very fast in MySQL, but you can obtain better overall performance by adding locks around everything that does more than about five successive inserts or updates. If you do very many successive inserts, you could do a LOCK TABLES followed by an UNLOCK TABLES once in a while (each 1,000 rows or so) to permit other threads to access table. This would still result in a nice performance gain.

    INSERT is still much slower for loading data than LOAD DATA INFILE, even when using the strategies just outlined.

  • To increase performance for MyISAM tables, for both LOAD DATA INFILE and INSERT, enlarge the key cache by increasing the key_buffer_size system variable. See Section 8.12.2, “Tuning Server Parameters”.

8.6.3 Speed of REPAIR TABLE Statements

REPAIR TABLE for MyISAM tables is similar to using myisamchk for repair operations, and some of the same performance optimizations apply:

  • myisamchk has variables that control memory allocation. You may be able to its improve performance by setting these variables, as described in Section 4.6.3.6, “myisamchk Memory Usage”.

  • For REPAIR TABLE, the same principle applies, but because the repair is done by the server, you set server system variables instead of myisamchk variables. Also, in addition to setting memory-allocation variables, increasing the myisam_max_sort_file_size system variable increases the likelihood that the repair will use the faster filesort method and avoid the slower repair by key cache method. Set the variable to the maximum file size for your system, after checking to be sure that there is enough free space to hold a copy of the table files. The free space must be available in the file system containing the original table files.

Suppose that a myisamchk table-repair operation is done using the following options to set its memory-allocation variables:

--key_buffer_size=128M --myisam_sort_buffer_size=256M
--read_buffer_size=64M --write_buffer_size=64M

Some of those myisamchk variables correspond to server system variables:

myisamchk VariableSystem Variable
key_buffer_sizekey_buffer_size
myisam_sort_buffer_sizemyisam_sort_buffer_size
read_buffer_sizeread_buffer_size
write_buffer_sizenone

Each of the server system variables can be set at runtime, and some of them (myisam_sort_buffer_size, read_buffer_size) have a session value in addition to a global value. Setting a session value limits the effect of the change to your current session and does not affect other users. Changing a global-only variable (key_buffer_size, myisam_max_sort_file_size) affects other users as well. For key_buffer_size, you must take into account that the buffer is shared with those users. For example, if you set the myisamchk key_buffer_size variable to 128MB, you could set the corresponding key_buffer_size system variable larger than that (if it is not already set larger), to permit key buffer use by activity in other sessions. However, changing the global key buffer size invalidates the buffer, causing increased disk I/O and slowdown for other sessions. An alternative that avoids this problem is to use a separate key cache, assign to it the indexes from the table to be repaired, and deallocate it when the repair is complete. See Section 8.10.2.2, “Multiple Key Caches”.

Based on the preceding remarks, a REPAIR TABLE operation can be done as follows to use settings similar to the myisamchk command. Here a separate 128MB key buffer is allocated and the file system is assumed to permit a file size of at least 100GB.

SET SESSION myisam_sort_buffer_size = 256*1024*1024;
SET SESSION read_buffer_size = 64*1024*1024;
SET GLOBAL myisam_max_sort_file_size = 100*1024*1024*1024;
SET GLOBAL repair_cache.key_buffer_size = 128*1024*1024;
CACHE INDEX tbl_name IN repair_cache;
LOAD INDEX INTO CACHE tbl_name;
REPAIR TABLE tbl_name ;
SET GLOBAL repair_cache.key_buffer_size = 0;

If you intend to change a global variable but want to do so only for the duration of a REPAIR TABLE operation to minimally affect other users, save its value in a user variable and restore it afterward. For example:

SET @old_myisam_sort_buffer_size = @@global.myisam_max_sort_file_size;
SET GLOBAL myisam_max_sort_file_size = 100*1024*1024*1024;
REPAIR TABLE tbl_name ;
SET GLOBAL myisam_max_sort_file_size = @old_myisam_max_sort_file_size;

The system variables that affect REPAIR TABLE can be set globally at server startup if you want the values to be in effect by default. For example, add these lines to the server my.cnf file:

[mysqld]
myisam_sort_buffer_size=256M
key_buffer_size=1G
myisam_max_sort_file_size=100G

These settings do not include read_buffer_size. Setting read_buffer_size globally to a large value does so for all sessions and can cause performance to suffer due to excessive memory allocation for a server with many simultaneous sessions.

8.7 Optimizing for MEMORY Tables

Consider using MEMORY tables for noncritical data that is accessed often, and is read-only or rarely updated. Benchmark your application against equivalent InnoDB or MyISAM tables under a realistic workload, to confirm that any additional performance is worth the risk of losing data, or the overhead of copying data from a disk-based table at application start.

For best performance with MEMORY tables, examine the kinds of queries against each table, and specify the type to use for each associated index, either a B-tree index or a hash index. On the CREATE INDEX statement, use the clause USING BTREE or USING HASH. B-tree indexes are fast for queries that do greater-than or less-than comparisons through operators such as > or BETWEEN. Hash indexes are only fast for queries that look up single values through the = operator, or a restricted set of values through the IN operator. For why USING BTREE is often a better choice than the default USING HASH, see Section 8.2.1.21, “How to Avoid Full Table Scans”. For implementation details of the different types of MEMORY indexes, see Section 8.3.8, “Comparison of B-Tree and Hash Indexes”.

8.8 Understanding the Query Execution Plan

Depending on the details of your tables, columns, indexes, and the conditions in your WHERE clause, the MySQL optimizer considers many techniques to efficiently perform the lookups involved in an SQL query. A query on a huge table can be performed without reading all the rows; a join involving several tables can be performed without comparing every combination of rows. The set of operations that the optimizer chooses to perform the most efficient query is called the query execution plan, also known as the EXPLAIN plan. Your goals are to recognize the aspects of the EXPLAIN plan that indicate a query is optimized well, and to learn the SQL syntax and indexing techniques to improve the plan if you see some inefficient operations.

8.8.1 Optimizing Queries with EXPLAIN

The EXPLAIN statement can be used to obtain information about how MySQL executes a statement:

With the help of EXPLAIN, you can see where you should add indexes to tables so that the statement executes faster by using indexes to find rows. You can also use EXPLAIN to check whether the optimizer joins the tables in an optimal order. To give a hint to the optimizer to use a join order corresponding to the order in which the tables are named in a SELECT statement, begin the statement with SELECT STRAIGHT_JOIN rather than just SELECT. (See Section 13.2.9, “SELECT Syntax”.) However, STRAIGHT_JOIN may prevent indexes from being used because it disables semi-join transformations. See Section 8.2.1.18.1, “Optimizing Subqueries with Semi-Join Transformations”.

The optimizer trace may sometimes provide information complementary to that of EXPLAIN. However, the optimizer trace format and content are subject to change between versions. For details, see MySQL Internals: Tracing the Optimizer.

If you have a problem with indexes not being used when you believe that they should be, run ANALYZE TABLE to update table statistics, such as cardinality of keys, that can affect the choices the optimizer makes. See Section 13.7.2.1, “ANALYZE TABLE Syntax”.

Note

EXPLAIN can also be used to obtain information about the columns in a table. EXPLAIN tbl_name is synonymous with DESCRIBE tbl_name and SHOW COLUMNS FROM tbl_name. For more information, see Section 13.8.1, “DESCRIBE Syntax”, and Section 13.7.5.5, “SHOW COLUMNS Syntax”.

8.8.2 EXPLAIN Output Format

The EXPLAIN statement provides information about the execution plan for a SELECT statement.

EXPLAIN returns a row of information for each table used in the SELECT statement. It lists the tables in the output in the order that MySQL would read them while processing the statement. MySQL resolves all joins using a nested-loop join method. This means that MySQL reads a row from the first table, and then finds a matching row in the second table, the third table, and so on. When all tables are processed, MySQL outputs the selected columns and backtracks through the table list until a table is found for which there are more matching rows. The next row is read from this table and the process continues with the next table.

Before MySQL 5.7.3, when the EXTENDED keyword is used, EXPLAIN produces extra information that can be viewed by issuing a SHOW WARNINGS statement following the EXPLAIN statement. EXPLAIN EXTENDED also displays the filtered column. See Section 8.8.3, “EXPLAIN EXTENDED Output Format”. As of MySQL 5.7.3, extended output is enabled by default and the EXTENDED keyword is unnecessary.

Note

You cannot use the EXTENDED and PARTITIONS keywords together in the same EXPLAIN statement. In addition, neither of these keywords can be used together with the FORMAT option. (FORMAT=JSON causes EXPLAIN to display extended and partition information automatically; using FORMAT=TRADITIONAL has no effect on EXPLAIN output.)

EXPLAIN Output Columns

This section describes the output columns produced by EXPLAIN. Later sections provide additional information about the type and Extra columns.

Each output row from EXPLAIN provides information about one table. Each row contains the values summarized in Table 8.1, “EXPLAIN Output Columns”, and described in more detail following the table. Column names are shown in the table's first column; the second column provides the equivalent property name shown in the output when FORMAT=JSON is used.

Table 8.1 EXPLAIN Output Columns

ColumnJSON NameMeaning
idselect_idThe SELECT identifier
select_typeNoneThe SELECT type
tabletable_nameThe table for the output row
partitionspartitionsThe matching partitions
typeaccess_typeThe join type
possible_keyspossible_keysThe possible indexes to choose
keykeyThe index actually chosen
key_lenkey_lengthThe length of the chosen key
refrefThe columns compared to the index
rowsrowsEstimate of rows to be examined
filteredfilteredPercentage of rows filtered by table condition
ExtraNoneAdditional information

Note

JSON properties which are NULL are not displayed in JSON-formatted EXPLAIN output.

  • id (JSON name: select_id)

    The SELECT identifier. This is the sequential number of the SELECT within the query. The value can be NULL if the row refers to the union result of other rows. In this case, the table column shows a value like <unionM,N> to indicate that the row refers to the union of the rows with id values of M and N.

  • select_type (JSON name: none)

    The type of SELECT, which can be any of those shown in the following table. A JSON-formatted EXPLAIN exposes the SELECT type as a property of a query_block, unless it is SIMPLE or PRIMARY. The JSON names (where applicable) are also shown in the table.

    select_type ValueJSON NameMeaning
    SIMPLENoneSimple SELECT (not using UNION or subqueries)
    PRIMARYNoneOutermost SELECT
    UNIONNoneSecond or later SELECT statement in a UNION
    DEPENDENT UNIONdependent (true)Second or later SELECT statement in a UNION, dependent on outer query
    UNION RESULTunion_resultResult of a UNION.
    SUBQUERYNoneFirst SELECT in subquery
    DEPENDENT SUBQUERYdependent (true)First SELECT in subquery, dependent on outer query
    DERIVEDNoneDerived table SELECT (subquery in FROM clause)
    MATERIALIZEDmaterialized_from_subqueryMaterialized subquery
    UNCACHEABLE SUBQUERYcacheable (false)A subquery for which the result cannot be cached and must be re-evaluated for each row of the outer query
    UNCACHEABLE UNIONcacheable (false)The second or later select in a UNION that belongs to an uncacheable subquery (see UNCACHEABLE SUBQUERY)

    DEPENDENT typically signifies the use of a correlated subquery. See Section 13.2.10.7, “Correlated Subqueries”.

    DEPENDENT SUBQUERY evaluation differs from UNCACHEABLE SUBQUERY evaluation. For DEPENDENT SUBQUERY, the subquery is re-evaluated only once for each set of different values of the variables from its outer context. For UNCACHEABLE SUBQUERY, the subquery is re-evaluated for each row of the outer context.

    Cacheability of subqueries differs from caching of query results in the query cache (which is described in Section 8.10.3.1, “How the Query Cache Operates”). Subquery caching occurs during query execution, whereas the query cache is used to store results only after query execution finishes.

    When you specify FORMAT=JSON with EXPLAIN, the output has no single property directly equivalent to select_type; the query_block property corresponds to a given SELECT. Properties equivalent to most of the SELECT subquery types just shown are available (an example being materialized_from_subquery for MATERIALIZED), and are displayed when appropriate. There are no JSON equivalents for SIMPLE or PRIMARY.

    As of MySQL 5.7.2, the select_type value for non-SELECT statements displays the statement type for affected tables. For example, select_type is DELETE for DELETE statements.

  • table (JSON name: table_name)

    The name of the table to which the row of output refers. This can also be one of the following values:

    • <unionM,N>: The row refers to the union of the rows with id values of M and N.

    • <derivedN>: The row refers to the derived table result for the row with an id value of N. A derived table may result, for example, from a subquery in the FROM clause.

    • <subqueryN>: The row refers to the result of a materialized subquery for the row with an id value of N. See Section 8.2.1.18.2, “Optimizing Subqueries with Subquery Materialization”.

  • partitions (JSON name: partitions)

    The partitions from which records would be matched by the query. This column is displayed only if the PARTITIONS keyword is used. The value is NULL for nonpartitioned tables. See Section 19.3.5, “Obtaining Information About Partitions”.

  • type (JSON name: access_type)

    The join type. For descriptions of the different types, see EXPLAIN Join Types.

  • possible_keys (JSON name: possible_keys)

    The possible_keys column indicates which indexes MySQL can choose from use to find the rows in this table. Note that this column is totally independent of the order of the tables as displayed in the output from EXPLAIN. That means that some of the keys in possible_keys might not be usable in practice with the generated table order.

    If this column is NULL (or undefined in JSON-formatted output), there are no relevant indexes. In this case, you may be able to improve the performance of your query by examining the WHERE clause to check whether it refers to some column or columns that would be suitable for indexing. If so, create an appropriate index and check the query with EXPLAIN again. See Section 13.1.8, “ALTER TABLE Syntax”.

    To see what indexes a table has, use SHOW INDEX FROM tbl_name.

  • key (JSON name: key)

    The key column indicates the key (index) that MySQL actually decided to use. If MySQL decides to use one of the possible_keys indexes to look up rows, that index is listed as the key value.

    It is possible that key will name an index that is not present in the possible_keys value. This can happen if none of the possible_keys indexes are suitable for looking up rows, but all the columns selected by the query are columns of some other index. That is, the named index covers the selected columns, so although it is not used to determine which rows to retrieve, an index scan is more efficient than a data row scan.

    For InnoDB, a secondary index might cover the selected columns even if the query also selects the primary key because InnoDB stores the primary key value with each secondary index. If key is NULL, MySQL found no index to use for executing the query more efficiently.

    To force MySQL to use or ignore an index listed in the possible_keys column, use FORCE INDEX, USE INDEX, or IGNORE INDEX in your query. See Section 8.9.4, “Index Hints”.

    For MyISAM tables, running ANALYZE TABLE helps the optimizer choose better indexes. For MyISAM tables, myisamchk --analyze does the same. See Section 13.7.2.1, “ANALYZE TABLE Syntax”, and Section 7.6, “MyISAM Table Maintenance and Crash Recovery”.

  • key_len (JSON name: key_length)

    The key_len column indicates the length of the key that MySQL decided to use. The length is NULL if the key column says NULL. Note that the value of key_len enables you to determine how many parts of a multiple-part key MySQL actually uses.

  • ref (JSON name: ref)

    The ref column shows which columns or constants are compared to the index named in the key column to select rows from the table.

    If the value is func, the value used is the result of some function. To see which function, use EXPLAIN EXTENDED followed by SHOW WARNINGS. The function might actually be an operator such as an arithmetic operator.

  • rows (JSON name: rows)

    The rows column indicates the number of rows MySQL believes it must examine to execute the query.

    For InnoDB tables, this number is an estimate, and may not always be exact.

  • filtered (JSON name: filtered)

    The filtered column indicates an estimated percentage of table rows that will be filtered by the table condition. That is, rows shows the estimated number of rows examined and rows × filtered / 100 shows the number of rows that will be joined with previous tables. Before MySQL 5.7.3, this column is displayed if you use EXPLAIN EXTENDED. As of MySQL 5.7.3, extended output is enabled by default and the EXTENDED keyword is unnecessary.

  • Extra (JSON name: none)

    This column contains additional information about how MySQL resolves the query. For descriptions of the different values, see EXPLAIN Extra Information.

    There is no single JSON property corresponding to the Extra column; however, values that can occur in this column are exposed as JSON properties, or as the text of the message property.

EXPLAIN Join Types

The type column of EXPLAIN output describes how tables are joined. In JSON-formatted output, these are found as values of the access_type property. The following list describes the join types, ordered from the best type to the worst:

  • system

    The table has only one row (= system table). This is a special case of the const join type.

  • const

    The table has at most one matching row, which is read at the start of the query. Because there is only one row, values from the column in this row can be regarded as constants by the rest of the optimizer. const tables are very fast because they are read only once.

    const is used when you compare all parts of a PRIMARY KEY or UNIQUE index to constant values. In the following queries, tbl_name can be used as a const table:

    SELECT * FROM tbl_name WHERE primary_key=1;
    
    SELECT * FROM tbl_name
      WHERE primary_key_part1=1 AND primary_key_part2=2;
    
  • eq_ref

    One row is read from this table for each combination of rows from the previous tables. Other than the system and const types, this is the best possible join type. It is used when all parts of an index are used by the join and the index is a PRIMARY KEY or UNIQUE NOT NULL index.

    eq_ref can be used for indexed columns that are compared using the = operator. The comparison value can be a constant or an expression that uses columns from tables that are read before this table. In the following examples, MySQL can use an eq_ref join to process ref_table:

    SELECT * FROM ref_table,other_table
      WHERE ref_table.key_column=other_table.column;
    
    SELECT * FROM ref_table,other_table
      WHERE ref_table.key_column_part1=other_table.column
      AND ref_table.key_column_part2=1;
    
  • ref

    All rows with matching index values are read from this table for each combination of rows from the previous tables. ref is used if the join uses only a leftmost prefix of the key or if the key is not a PRIMARY KEY or UNIQUE index (in other words, if the join cannot select a single row based on the key value). If the key that is used matches only a few rows, this is a good join type.

    ref can be used for indexed columns that are compared using the = or <=> operator. In the following examples, MySQL can use a ref join to process ref_table:

    SELECT * FROM ref_table WHERE key_column=expr;
    
    SELECT * FROM ref_table,other_table
      WHERE ref_table.key_column=other_table.column;
    
    SELECT * FROM ref_table,other_table
      WHERE ref_table.key_column_part1=other_table.column
      AND ref_table.key_column_part2=1;
    
  • fulltext

    The join is performed using a FULLTEXT index.

  • ref_or_null

    This join type is like ref, but with the addition that MySQL does an extra search for rows that contain NULL values. This join type optimization is used most often in resolving subqueries. In the following examples, MySQL can use a ref_or_null join to process ref_table:

    SELECT * FROM ref_table
      WHERE key_column=expr OR key_column IS NULL;
    

    See Section 8.2.1.8, “IS NULL Optimization”.

  • index_merge

    This join type indicates that the Index Merge optimization is used. In this case, the key column in the output row contains a list of indexes used, and key_len contains a list of the longest key parts for the indexes used. For more information, see Section 8.2.1.4, “Index Merge Optimization”.

  • unique_subquery

    This type replaces eq_ref for some IN subqueries of the following form:

    value IN (SELECT primary_key FROM single_table WHERE some_expr)
    

    unique_subquery is just an index lookup function that replaces the subquery completely for better efficiency.

  • index_subquery

    This join type is similar to unique_subquery. It replaces IN subqueries, but it works for nonunique indexes in subqueries of the following form:

    value IN (SELECT key_column FROM single_table WHERE some_expr)
    
  • range

    Only rows that are in a given range are retrieved, using an index to select the rows. The key column in the output row indicates which index is used. The key_len contains the longest key part that was used. The ref column is NULL for this type.

    range can be used when a key column is compared to a constant using any of the =, <>, >, >=, <, <=, IS NULL, <=>, BETWEEN, or IN() operators:

    SELECT * FROM tbl_name
      WHERE key_column = 10;
    
    SELECT * FROM tbl_name
      WHERE key_column BETWEEN 10 and 20;
    
    SELECT * FROM tbl_name
      WHERE key_column IN (10,20,30);
    
    SELECT * FROM tbl_name
      WHERE key_part1 = 10 AND key_part2 IN (10,20,30);
    
  • index

    The index join type is the same as ALL, except that the index tree is scanned. This occurs two ways:

    • If the index is a covering index for the queries and can be used to satisfy all data required from the table, only the index tree is scanned. In this case, the Extra column says Using index. An index-only scan usually is faster than ALL because the size of the index usually is smaller than the table data.

    • A full table scan is performed using reads from the index to look up data rows in index order. Uses index does not appear in the Extra column.

    MySQL can use this join type when the query uses only columns that are part of a single index.

  • ALL

    A full table scan is done for each combination of rows from the previous tables. This is normally not good if the table is the first table not marked const, and usually very bad in all other cases. Normally, you can avoid ALL by adding indexes that enable row retrieval from the table based on constant values or column values from earlier tables.

EXPLAIN Extra Information

The Extra column of EXPLAIN output contains additional information about how MySQL resolves the query. The following list explains the values that can appear in this column. Each item also indicates for JSON-formatted output which property displays the Extra value. For some of these, there is a specific property. The others display as the text of the message property.

If you want to make your queries as fast as possible, look out for Extra column values of Using filesort and Using temporary, or, in JSON-formatted EXPLAIN output, for using_filesort and using_temporary_table properties equal to true.

  • Child of 'table' pushed join@1 (JSON: message text)

    This table is referenced as the child of table in a join that can be pushed down to the NDB kernel. Applies only in MySQL Cluster, when pushed-down joins are enabled. See the description of the ndb_join_pushdown server system variable for more information and examples.

  • const row not found (JSON property: const_row_not_found)

    For a query such as SELECT ... FROM tbl_name, the table was empty.

  • Deleting all rows (JSON property: message)

    For DELETE, some storage engines (such as MyISAM) support a handler method that removes all table rows in a simple and fast way. This Extra value is displayed if the engine uses this optimization.

  • Distinct (JSON property: distinct)

    MySQL is looking for distinct values, so it stops searching for more rows for the current row combination after it has found the first matching row.

  • FirstMatch(tbl_name) (JSON property: first_match)

    The semi-join FirstMatch join shortcutting strategy is used for tbl_name.

  • Full scan on NULL key (JSON property: message)

    This occurs for subquery optimization as a fallback strategy when the optimizer cannot use an index-lookup access method.

  • Impossible HAVING (JSON property: message)

    The HAVING clause is always false and cannot select any rows.

  • Impossible WHERE (JSON property: message)

    The WHERE clause is always false and cannot select any rows.

  • Impossible WHERE noticed after reading const tables (JSON property: message)

    MySQL has read all const (and system) tables and notice that the WHERE clause is always false.

  • LooseScan(m..n) (JSON property: message)

    The semi-join LooseScan strategy is used. m and n are key part numbers.

  • No matching min/max row (JSON property: message)

    No row satisfies the condition for a query such as SELECT MIN(...) FROM ... WHERE condition.

  • no matching row in const table (JSON property: message)

    For a query with a join, there was an empty table or a table with no rows satisfying a unique index condition.

  • No matching rows after partition pruning (JSON property: message)

    For DELETE or UPDATE, the optimizer found nothing to delete or update after partition pruning. It is similar in meaning to Impossible WHERE for SELECT statements.

  • No tables used (JSON property: message)

    The query has no FROM clause, or has a FROM DUAL clause.

    For INSERT or REPLACE statements, EXPLAIN displays this value when there is no SELECT part. For example, it appears for EXPLAIN INSERT INTO t VALUES(10) because that is equivalent to EXPLAIN INSERT INTO t SELECT 10 FROM DUAL.

  • Not exists (JSON property: message)

    MySQL was able to do a LEFT JOIN optimization on the query and does not examine more rows in this table for the previous row combination after it finds one row that matches the LEFT JOIN criteria. Here is an example of the type of query that can be optimized this way:

    SELECT * FROM t1 LEFT JOIN t2 ON t1.id=t2.id
      WHERE t2.id IS NULL;
    

    Assume that t2.id is defined as NOT NULL. In this case, MySQL scans t1 and looks up the rows in t2 using the values of t1.id. If MySQL finds a matching row in t2, it knows that t2.id can never be NULL, and does not scan through the rest of the rows in t2 that have the same id value. In other words, for each row in t1, MySQL needs to do only a single lookup in t2, regardless of how many rows actually match in t2.

  • Plan isn't ready yet (JSON property: none)

    This value occurs with EXPLAIN FOR CONNECTION when the optimizer has not finished creating the execution plan for the statement executing in the named connection. If execution plan output comprises multiple lines, any or all of them could have this Extra value, depending on the progress of the optimizer in determining the full execution plan.

  • Range checked for each record (index map: N) (JSON property: message)

    MySQL found no good index to use, but found that some of indexes might be used after column values from preceding tables are known. For each row combination in the preceding tables, MySQL checks whether it is possible to use a range or index_merge access method to retrieve rows. This is not very fast, but is faster than performing a join with no index at all. The applicability criteria are as described in Section 8.2.1.3, “Range Optimization”, and Section 8.2.1.4, “Index Merge Optimization”, with the exception that all column values for the preceding table are known and considered to be constants.

    Indexes are numbered beginning with 1, in the same order as shown by SHOW INDEX for the table. The index map value N is a bitmask value that indicates which indexes are candidates. For example, a value of 0x19 (binary 11001) means that indexes 1, 4, and 5 will be considered.

  • Scanned N databases (JSON property: message)

    This indicates how many directory scans the server performs when processing a query for INFORMATION_SCHEMA tables, as described in Section 8.2.4, “Optimizing INFORMATION_SCHEMA Queries”. The value of N can be 0, 1, or all.

  • Select tables optimized away (JSON property: message)

    The optimizer determined 1) that at most one row should be returned, and 2) that to produce this row, a deterministic set of rows must be read. When the rows to be read can be read during the optimization phase (for example, by reading index rows), there is no need to read any tables during query execution.

    The first condition is fulfilled when the query is implicitly grouped (contains an aggregate function but no GROUP BY clause). The second condition is fulfilled when one row lookup is performed per index used. The number of indexes read determines the number of rows to read.

    Consider the following implicitly grouped query:

    SELECT MIN(c1), MIN(c2) FROM t1;
    

    Suppose that MIN(c1) can be retrieved by reading one index row and MIN(c2) can be retrieved by reading one row from a different index. That is, for each column c1 and c2, there exists an index where the column is the first column of the index. In this case, one row is returned, produced by reading two deterministic rows.

    This Extra value does not occur if the rows to read are not deterministic. Consider this query:

    SELECT MIN(c2) FROM t1 WHERE c1 <= 10;
    

    Suppose that (c1, c2) is a covering index. Using this index, all rows with c1 <= 10 must be scanned to find the minimum c2 value. By contrast, consider this query:

    SELECT MIN(c2) FROM t1 WHERE c1 = 10;
    

    In this case, the first index row with c1 = 10 contains the minimum c2 value. Only one row must be read to produce the returned row.

    For storage engines that maintain an exact row count per table (such as MyISAM, but not InnoDB), this Extra value can occur for COUNT(*) queries for which the WHERE clause is missing or always true and there is no GROUP BY clause. (This is an instance of an implicitly grouped query where the storage engine influences whether a deterministic number of rows can be read.)

  • Skip_open_table, Open_frm_only, Open_trigger_only, Open_full_table (JSON property: message)

    These values indicate file-opening optimizations that apply to queries for INFORMATION_SCHEMA tables, as described in Section 8.2.4, “Optimizing INFORMATION_SCHEMA Queries”.

    • Skip_open_table: Table files do not need to be opened. The information has already become available within the query by scanning the database directory.

    • Open_frm_only: Only the table's .frm file need be opened.

    • Open_trigger_only: Only the table's .TRG file need be opened.

    • Open_full_table: The unoptimized information lookup. The .frm, .MYD, and .MYI files must be opened.

  • Start temporary, End temporary (JSON property: message)

    This indicates temporary table use for the semi-join Duplicate Weedout strategy.

  • unique row not found (JSON property: message)

    For a query such as SELECT ... FROM tbl_name, no rows satisfy the condition for a UNIQUE index or PRIMARY KEY on the table.

  • Using filesort (JSON property: using_filesort)

    MySQL must do an extra pass to find out how to retrieve the rows in sorted order. The sort is done by going through all rows according to the join type and storing the sort key and pointer to the row for all rows that match the WHERE clause. The keys then are sorted and the rows are retrieved in sorted order. See Section 8.2.1.15, “ORDER BY Optimization”.

  • Using index (JSON property: using_index)

    The column information is retrieved from the table using only information in the index tree without having to do an additional seek to read the actual row. This strategy can be used when the query uses only columns that are part of a single index.

    For InnoDB tables that have a user-defined clustered index, that index can be used even when Using index is absent from the Extra column. This is the case if type is index and key is PRIMARY.

  • Using index condition (JSON property: using_index_condition)

    Tables are read by accessing index tuples and testing them first to determine whether to read full table rows. In this way, index information is used to defer (push down) reading full table rows unless it is necessary. See Section 8.2.1.6, “Index Condition Pushdown Optimization”.

  • Using index for group-by (JSON property: using_index_for_group_by)

    Similar to the Using index table access method, Using index for group-by indicates that MySQL found an index that can be used to retrieve all columns of a GROUP BY or DISTINCT query without any extra disk access to the actual table. Additionally, the index is used in the most efficient way so that for each group, only a few index entries are read. For details, see Section 8.2.1.16, “GROUP BY Optimization”.

  • Using join buffer (Block Nested Loop), Using join buffer (Batched Key Access) (JSON property: using_join_buffer)

    Tables from earlier joins are read in portions into the join buffer, and then their rows are used from the buffer to perform the join with the current table. (Block Nested Loop) indicates use of the Block Nested-Loop algorithm and (Batched Key Access) indicates use of the Batched Key Access algorithm. That is, the keys from the table on the preceding line of the EXPLAIN output will be buffered, and the matching rows will be fetched in batches from the table represented by the line in which Using join buffer appears.

    In JSON-formatted output, the value of using_join_buffer is always either one of Block Nested Loop or Batched Key Access.

  • Using MRR (JSON property: message)

    Tables are read using the Multi-Range Read optimization strategy. See Section 8.2.1.13, “Multi-Range Read Optimization”.

  • Using sort_union(...), Using union(...), Using intersect(...) (JSON property: message)

    These indicate how index scans are merged for the index_merge join type. See Section 8.2.1.4, “Index Merge Optimization”.

  • Using temporary (JSON property: using_temporary_table)

    To resolve the query, MySQL needs to create a temporary table to hold the result. This typically happens if the query contains GROUP BY and ORDER BY clauses that list columns differently.

  • Using where (JSON property: attached_condition)

    A WHERE clause is used to restrict which rows to match against the next table or send to the client. Unless you specifically intend to fetch or examine all rows from the table, you may have something wrong in your query if the Extra value is not Using where and the table join type is ALL or index.

    Using where has no direct counterpart in JSON-formatted output; the attached_condition property contains any WHERE condition used.

  • Using where with pushed condition (JSON property: message)

    This item applies to NDB tables only. It means that MySQL Cluster is using the Condition Pushdown optimization to improve the efficiency of a direct comparison between a nonindexed column and a constant. In such cases, the condition is pushed down to the cluster's data nodes and is evaluated on all data nodes simultaneously. This eliminates the need to send nonmatching rows over the network, and can speed up such queries by a factor of 5 to 10 times over cases where Condition Pushdown could be but is not used. For more information, see Section 8.2.1.5, “Engine Condition Pushdown Optimization”.

  • Zero limit (JSON property: message)

    The query had a LIMIT 0 clause and cannot select any rows.

EXPLAIN Output Interpretation

You can get a good indication of how good a join is by taking the product of the values in the rows column of the EXPLAIN output. This should tell you roughly how many rows MySQL must examine to execute the query. If you restrict queries with the max_join_size system variable, this row product also is used to determine which multiple-table SELECT statements to execute and which to abort. See Section 8.12.2, “Tuning Server Parameters”.

The following example shows how a multiple-table join can be optimized progressively based on the information provided by EXPLAIN.

Suppose that you have the SELECT statement shown here and that you plan to examine it using EXPLAIN:

EXPLAIN SELECT tt.TicketNumber, tt.TimeIn,
               tt.ProjectReference, tt.EstimatedShipDate,
               tt.ActualShipDate, tt.ClientID,
               tt.ServiceCodes, tt.RepetitiveID,
               tt.CurrentProcess, tt.CurrentDPPerson,
               tt.RecordVolume, tt.DPPrinted, et.COUNTRY,
               et_1.COUNTRY, do.CUSTNAME
        FROM tt, et, et AS et_1, do
        WHERE tt.SubmitTime IS NULL
          AND tt.ActualPC = et.EMPLOYID
          AND tt.AssignedPC = et_1.EMPLOYID
          AND tt.ClientID = do.CUSTNMBR;

For this example, make the following assumptions:

  • The columns being compared have been declared as follows.

    TableColumnData Type
    ttActualPCCHAR(10)
    ttAssignedPCCHAR(10)
    ttClientIDCHAR(10)
    etEMPLOYIDCHAR(15)
    doCUSTNMBRCHAR(15)
  • The tables have the following indexes.

    TableIndex
    ttActualPC
    ttAssignedPC
    ttClientID
    etEMPLOYID (primary key)
    doCUSTNMBR (primary key)
  • The tt.ActualPC values are not evenly distributed.

Initially, before any optimizations have been performed, the EXPLAIN statement produces the following information:

table type possible_keys key  key_len ref  rows  Extra
et    ALL  PRIMARY       NULL NULL    NULL 74
do    ALL  PRIMARY       NULL NULL    NULL 2135
et_1  ALL  PRIMARY       NULL NULL    NULL 74
tt    ALL  AssignedPC,   NULL NULL    NULL 3872
           ClientID,
           ActualPC
      Range checked for each record (index map: 0x23)

Because type is ALL for each table, this output indicates that MySQL is generating a Cartesian product of all the tables; that is, every combination of rows. This takes quite a long time, because the product of the number of rows in each table must be examined. For the case at hand, this product is 74 × 2135 × 74 × 3872 = 45,268,558,720 rows. If the tables were bigger, you can only imagine how long it would take.

One problem here is that MySQL can use indexes on columns more efficiently if they are declared as the same type and size. In this context, VARCHAR and CHAR are considered the same if they are declared as the same size. tt.ActualPC is declared as CHAR(10) and et.EMPLOYID is CHAR(15), so there is a length mismatch.

To fix this disparity between column lengths, use ALTER TABLE to lengthen ActualPC from 10 characters to 15 characters:

mysql> ALTER TABLE tt MODIFY ActualPC VARCHAR(15);

Now tt.ActualPC and et.EMPLOYID are both VARCHAR(15). Executing the EXPLAIN statement again produces this result:

table type   possible_keys key     key_len ref         rows    Extra
tt    ALL    AssignedPC,   NULL    NULL    NULL        3872    Using
             ClientID,                                         where
             ActualPC
do    ALL    PRIMARY       NULL    NULL    NULL        2135
      Range checked for each record (index map: 0x1)
et_1  ALL    PRIMARY       NULL    NULL    NULL        74
      Range checked for each record (index map: 0x1)
et    eq_ref PRIMARY       PRIMARY 15      tt.ActualPC 1

This is not perfect, but is much better: The product of the rows values is less by a factor of 74. This version executes in a couple of seconds.

A second alteration can be made to eliminate the column length mismatches for the tt.AssignedPC = et_1.EMPLOYID and tt.ClientID = do.CUSTNMBR comparisons:

mysql> ALTER TABLE tt MODIFY AssignedPC VARCHAR(15),
    ->                MODIFY ClientID   VARCHAR(15);

After that modification, EXPLAIN produces the output shown here:

table type   possible_keys key      key_len ref           rows Extra
et    ALL    PRIMARY       NULL     NULL    NULL          74
tt    ref    AssignedPC,   ActualPC 15      et.EMPLOYID   52   Using
             ClientID,                                         where
             ActualPC
et_1  eq_ref PRIMARY       PRIMARY  15      tt.AssignedPC 1
do    eq_ref PRIMARY       PRIMARY  15      tt.ClientID   1

At this point, the query is optimized almost as well as possible. The remaining problem is that, by default, MySQL assumes that values in the tt.ActualPC column are evenly distributed, and that is not the case for the tt table. Fortunately, it is easy to tell MySQL to analyze the key distribution:

mysql> ANALYZE TABLE tt;

With the additional index information, the join is perfect and EXPLAIN produces this result:

table type   possible_keys key     key_len ref           rows Extra
tt    ALL    AssignedPC    NULL    NULL    NULL          3872 Using
             ClientID,                                        where
             ActualPC
et    eq_ref PRIMARY       PRIMARY 15      tt.ActualPC   1
et_1  eq_ref PRIMARY       PRIMARY 15      tt.AssignedPC 1
do    eq_ref PRIMARY       PRIMARY 15      tt.ClientID   1

The rows column in the output from EXPLAIN is an educated guess from the MySQL join optimizer. Check whether the numbers are even close to the truth by comparing the rows product with the actual number of rows that the query returns. If the numbers are quite different, you might get better performance by using STRAIGHT_JOIN in your SELECT statement and trying to list the tables in a different order in the FROM clause. (However, STRAIGHT_JOIN may prevent indexes from being used because it disables semi-join transformations. See Section 8.2.1.18.1, “Optimizing Subqueries with Semi-Join Transformations”.)

It is possible in some cases to execute statements that modify data when EXPLAIN SELECT is used with a subquery; for more information, see Section 13.2.10.8, “Subqueries in the FROM Clause”.

8.8.3 EXPLAIN EXTENDED Output Format

When EXPLAIN is used with the EXTENDED keyword, the output includes a filtered column not otherwise displayed. This column indicates the estimated percentage of table rows that will be filtered by the table condition. In addition, the statement produces extra information that can be viewed by issuing a SHOW WARNINGS statement following the EXPLAIN statement. The Message value in SHOW WARNINGS output displays how the optimizer qualifies table and column names in the SELECT statement, what the SELECT looks like after the application of rewriting and optimization rules, and possibly other notes about the optimization process.

Note

As of MySQL 5.7.3, the EXPLAIN statement is changed so that the effect of the EXTENDED keyword is always enabled. EXTENDED is still recognized, but is superfluous and is deprecated. It will be removed from EXPLAIN syntax in a future MySQL release.

Here is an example of extended output:

mysql> EXPLAIN EXTENDED
    -> SELECT t1.a, t1.a IN (SELECT t2.a FROM t2) FROM t1\G
*************************** 1. row ***************************
           id: 1
  select_type: PRIMARY
        table: t1
         type: index
possible_keys: NULL
          key: PRIMARY
      key_len: 4
          ref: NULL
         rows: 4
     filtered: 100.00
        Extra: Using index
*************************** 2. row ***************************
           id: 2
  select_type: SUBQUERY
        table: t2
         type: index
possible_keys: a
          key: a
      key_len: 5
          ref: NULL
         rows: 3
     filtered: 100.00
        Extra: Using index
2 rows in set, 1 warning (0.00 sec)

mysql> SHOW WARNINGS\G
*************************** 1. row ***************************
  Level: Note
   Code: 1003
Message: /* select#1 */ select `test`.`t1`.`a` AS `a`,
         <in_optimizer>(`test`.`t1`.`a`,`test`.`t1`.`a` in
         ( <materialize> (/* select#2 */ select `test`.`t2`.`a`
         from `test`.`t2` where 1 having 1 ),
         <primary_index_lookup>(`test`.`t1`.`a` in
         <temporary table> on <auto_key>
         where ((`test`.`t1`.`a` = `materialized-subquery`.`a`))))) AS `t1.a
         IN (SELECT t2.a FROM t2)` from `test`.`t1`
1 row in set (0.00 sec)

EXPLAIN EXTENDED can be used with SELECT, DELETE, INSERT, REPLACE, and UPDATE statements. However, the following SHOW WARNINGS statement displays a nonempty result only for SELECT statements.

Because the statement displayed by SHOW WARNINGS may contain special markers to provide information about query rewriting or optimizer actions, the statement is not necessarily valid SQL and is not intended to be executed. The output may also include rows with Message values that provide additional non-SQL explanatory notes about actions taken by the optimizer.

The following list describes special markers that can appear in EXTENDED output displayed by SHOW WARNINGS:

  • <auto_key>

    An automatically generated key for a temporary table.

  • <cache>(expr)

    The expression (such as a scalar subquery) is executed once and the resulting value is saved in memory for later use. For results consisting of multiple values, a temporary table may be created and you will see <temporary table> instead.

  • <exists>(query fragment)

    The subquery predicate is converted to an EXISTS predicate and the subquery is transformed so that it can be used together with the EXISTS predicate.

  • <in_optimizer>(query fragment)

    This is an internal optimizer object with no user significance.

  • <index_lookup>(query fragment)

    The query fragment is processed using an index lookup to find qualifying rows.

  • <if>(condition, expr1, expr2)

    If the condition is true, evaluate to expr1, otherwise expr2.

  • <is_not_null_test>(expr)

    A test to verify that the expression does not evaluate to NULL.

  • <materialize>(query fragment)

    Subquery materialization is used.

  • `materialized-subquery`.col_name

    A reference to the column col_name in an internal temporary table materialized to hold the result from evaluating a subquery.

  • <primary_index_lookup>(query fragment)

    The query fragment is processed using a primary key lookup to find qualifying rows.

  • <ref_null_helper>(expr)

    This is an internal optimizer object with no user significance.

  • /* select#N */ select_stmt

    The SELECT is associated with the row in non-EXTENDED EXPLAIN output that has an id value of N.

  • outer_tables semi join (inner_tables)

    A semi-join operation. inner_tables shows the tables that were not pulled out. See Section 8.2.1.18.1, “Optimizing Subqueries with Semi-Join Transformations”.

  • <temporary table>

    This represents an internal temporary table created to cache an intermediate result.

When some tables are of const or system type, expressions involving columns from these tables are evaluated early by the optimizer and are not part of the displayed statement. However, with FORMAT=JSON, some const table accesses are displayed as a ref access that uses a const value.

8.8.4 Obtaining Execution Plan Information for a Named Connection

To obtain the execution plan for an explainable statement executing in a named connection, use this statement:

EXPLAIN [options] FOR CONNECTION connection_id;

EXPLAIN FOR CONNECTION returns the EXPLAIN information that is currently being used to execute a query in a given connection. Because of changes to data (and supporting statistics) it may produce a different result from running EXPLAIN on the equivalent query text. This difference in behavior can be useful in diagnosing more transient performance problems. For example, if you are running a statement in one session that is taking a long time to complete, using EXPLAIN FOR CONNECTION in another session may yield useful information about the cause of the delay.

connection_id is the connection identifier, as obtained from the INFORMATION_SCHEMA PROCESSLIST table or the SHOW PROCESSLIST statement. If you have the PROCESS privilege, you can specify the identifier for any connection. Otherwise, you can specify the identifier only for your own connections.

If the named connection is not executing a statement, the result is empty. Otherwise, EXPLAIN FOR CONNECTION applies only if the statement being executed in the named connection is explainable. This includes SELECT, DELETE, INSERT, REPLACE, and UPDATE. (However, EXPLAIN FOR CONNECTION does not work for prepared statements, even prepared statements of those types.)

If the named connection is executing an explainable statement, the output is what you would obtain by using EXPLAIN on the statement itself.

If the named connection is executing a statement that is not explainable, an error occurs. For example, you cannot name the connection identifier for your current session because EXPLAIN is not explainable:

mysql> SELECT CONNECTION_ID();
+-----------------+
| CONNECTION_ID() |
+-----------------+
|             373 |
+-----------------+
1 row in set (0.00 sec)

mysql> EXPLAIN FOR CONNECTION 373;
ERROR 1889 (HY000): EXPLAIN FOR CONNECTION command is supported
only for SELECT/UPDATE/INSERT/DELETE/REPLACE

The Com_explain_other status variable indicates the number of EXPLAIN FOR CONNECTION statements executed.

8.8.5 Estimating Query Performance

In most cases, you can estimate query performance by counting disk seeks. For small tables, you can usually find a row in one disk seek (because the index is probably cached). For bigger tables, you can estimate that, using B-tree indexes, you need this many seeks to find a row: log(row_count) / log(index_block_length / 3 * 2 / (index_length + data_pointer_length)) + 1.

In MySQL, an index block is usually 1,024 bytes and the data pointer is usually four bytes. For a 500,000-row table with a key value length of three bytes (the size of MEDIUMINT), the formula indicates log(500,000)/log(1024/3*2/(3+4)) + 1 = 4 seeks.

This index would require storage of about 500,000 * 7 * 3/2 = 5.2MB (assuming a typical index buffer fill ratio of 2/3), so you probably have much of the index in memory and so need only one or two calls to read data to find the row.

For writes, however, you need four seek requests to find where to place a new index value and normally two seeks to update the index and write the row.

The preceding discussion does not mean that your application performance slowly degenerates by log N. As long as everything is cached by the OS or the MySQL server, things become only marginally slower as the table gets bigger. After the data gets too big to be cached, things start to go much slower until your applications are bound only by disk seeks (which increase by log N). To avoid this, increase the key cache size as the data grows. For MyISAM tables, the key cache size is controlled by the key_buffer_size system variable. See Section 8.12.2, “Tuning Server Parameters”.

8.9 Controlling the Query Optimizer

MySQL provides optimizer control through system variables that affect how query plans are evaluated, switchable optimizations, optimizer and index hints, and the optimizer cost model.

8.9.1 Controlling Query Plan Evaluation

The task of the query optimizer is to find an optimal plan for executing an SQL query. Because the difference in performance between good and bad plans can be orders of magnitude (that is, seconds versus hours or even days), most query optimizers, including that of MySQL, perform a more or less exhaustive search for an optimal plan among all possible query evaluation plans. For join queries, the number of possible plans investigated by the MySQL optimizer grows exponentially with the number of tables referenced in a query. For small numbers of tables (typically less than 7 to 10) this is not a problem. However, when larger queries are submitted, the time spent in query optimization may easily become the major bottleneck in the server's performance.

A more flexible method for query optimization enables the user to control how exhaustive the optimizer is in its search for an optimal query evaluation plan. The general idea is that the fewer plans that are investigated by the optimizer, the less time it spends in compiling a query. On the other hand, because the optimizer skips some plans, it may miss finding an optimal plan.

The behavior of the optimizer with respect to the number of plans it evaluates can be controlled using two system variables:

  • The optimizer_prune_level variable tells the optimizer to skip certain plans based on estimates of the number of rows accessed for each table. Our experience shows that this kind of educated guess rarely misses optimal plans, and may dramatically reduce query compilation times. That is why this option is on (optimizer_prune_level=1) by default. However, if you believe that the optimizer missed a better query plan, this option can be switched off (optimizer_prune_level=0) with the risk that query compilation may take much longer. Note that, even with the use of this heuristic, the optimizer still explores a roughly exponential number of plans.

  • The optimizer_search_depth variable tells how far into the future of each incomplete plan the optimizer should look to evaluate whether it should be expanded further. Smaller values of optimizer_search_depth may result in orders of magnitude smaller query compilation times. For example, queries with 12, 13, or more tables may easily require hours and even days to compile if optimizer_search_depth is close to the number of tables in the query. At the same time, if compiled with optimizer_search_depth equal to 3 or 4, the optimizer may compile in less than a minute for the same query. If you are unsure of what a reasonable value is for optimizer_search_depth, this variable can be set to 0 to tell the optimizer to determine the value automatically.

8.9.2 Controlling Switchable Optimizations

The optimizer_switch system variable enables control over optimizer behavior. Its value is a set of flags, each of which has a value of on or off to indicate whether the corresponding optimizer behavior is enabled or disabled. This variable has global and session values and can be changed at runtime. The global default can be set at server startup.

To see the current set of optimizer flags, select the variable value:

mysql> SELECT @@optimizer_switch\G
*************************** 1. row ***************************
@@optimizer_switch: index_merge=on,index_merge_union=on,
                    index_merge_sort_union=on,
                    index_merge_intersection=on,
                    engine_condition_pushdown=on,
                    index_condition_pushdown=on,
                    mrr=on,mrr_cost_based=on,
                    block_nested_loop=on,batched_key_access=off,
                    materialization=on,semijoin=on,loosescan=on,
                    firstmatch=on,duplicateweedout=on,
                    subquery_materialization_cost_based=on,
                    use_index_extensions=on,
                    condition_fanout_filter=on,derived_merge=on

To change the value of optimizer_switch, assign a value consisting of a comma-separated list of one or more commands:

SET [GLOBAL|SESSION] optimizer_switch='command[,command]...';

Each command value should have one of the forms shown in the following table.

Command SyntaxMeaning
defaultReset every optimization to its default value
opt_name=defaultSet the named optimization to its default value
opt_name=offDisable the named optimization
opt_name=onEnable the named optimization

The order of the commands in the value does not matter, although the default command is executed first if present. Setting an opt_name flag to default sets it to whichever of on or off is its default value. Specifying any given opt_name more than once in the value is not permitted and causes an error. Any errors in the value cause the assignment to fail with an error, leaving the value of optimizer_switch unchanged.

The following table lists the permissible opt_name flag names, grouped by optimization strategy.

OptimizationFlag NameMeaningDefault
Batched Key Accessbatched_key_accessControls use of BKA join algorithmOFF
Block Nested-Loopblock_nested_loopControls use of BNL join algorithmON
Condition Filteringcondition_fanout_filterControls use of condition filteringON
Engine Condition Pushdownengine_condition_pushdownControls engine condition pushdownON
Index Condition Pushdownindex_condition_pushdownControls index condition pushdownON
Index Extensionsuse_index_extensionsControls use of index extensionsON
Index Mergeindex_mergeControls all Index Merge optimizationsON
 index_merge_intersectionControls the Index Merge Intersection Access optimizationON
 index_merge_sort_unionControls the Index Merge Sort-Union Access optimizationON
 index_merge_unionControls the Index Merge Union Access optimizationON
Multi-Range ReadmrrControls the Multi-Range Read strategyON
 mrr_cost_basedControls use of cost-based MRR if mrr=onON
Semi-joinsemijoinControls all semi-join strategiesON
 firstmatchControls the semi-join FirstMatch strategyON
 loosescanControls the semi-join LooseScan strategy (not to be confused with LooseScan for GROUP BY)ON
 duplicateweedoutControls the semi-join Duplicate Weedout strategyON
Subquery materializationmaterializationControls materialization (including semi-join materialization)ON
 subquery_materialization_cost_basedUsed cost-based materialization choiceON
Derived table mergingderived_mergeControls merging of derived tables and views into outer query blockON

For batched_key_access to have any effect when set to on, the mrr flag must also be on. Currently, the cost estimation for MRR is too pessimistic. Hence, it is also necessary for mrr_cost_based to be off for BKA to be used.

The semijoin, firstmatch, loosescan, duplicateweedout (added in MySQL 5.7.8), and materialization flags enable control over semi-join and subquery materialization strategies. The semijoin flag controls whether semi-joins are used. If it is set to on, the firstmatch and loosescan flags enable finer control over the permitted semi-join strategies. The materialization flag controls whether subquery materialization is used. If semijoin and materialization are both on, semi-joins also use materialization where applicable. These flags are on by default.

If the duplicateweedout semi-join strategy is disabled, it is not used unless all other applicable strategies are also disabled.

The subquery_materialization_cost_based flag enables control over the choice between subquery materialization and IN-to-EXISTS subquery transformation. If the flag is on (the default), the optimizer performs a cost-based choice between subquery materialization and IN-to-EXISTS subquery transformation if either method could be used. If the flag is off, the optimizer chooses subquery materialization over IN -> EXISTS subquery transformation.

The derived_merge flag controls whether the optimizer attempts to merge derived tables and view references into the outer query block, assuming that no other rule prevents merging; for example, an ALGORITHM directive for a view takes precedence over the derived_merge setting. By default, the flag is on to enable merging. For more information, see Section 8.2.1.18.3, “Optimizing Derived Tables and View References”.

For more information about individual optimization strategies, see the following sections:

When you assign a value to optimizer_switch, flags that are not mentioned keep their current values. This makes it possible to enable or disable specific optimizer behaviors in a single statement without affecting other behaviors. The statement does not depend on what other optimizer flags exist and what their values are. Suppose that all Index Merge optimizations are enabled:

mysql> SELECT @@optimizer_switch\G
*************************** 1. row ***************************
@@optimizer_switch: index_merge=on,index_merge_union=on,
                    index_merge_sort_union=on,
                    index_merge_intersection=on,
                    engine_condition_pushdown=on,
                    index_condition_pushdown=on,
                    mrr=on,mrr_cost_based=on,
                    block_nested_loop=on,batched_key_access=off,
                    materialization=on,semijoin=on,loosescan=on,
                    firstmatch=on,
                    subquery_materialization_cost_based=on,
                    use_index_extensions=on,
                    condition_fanout_filter=on

If the server is using the Index Merge Union or Index Merge Sort-Union access methods for certain queries and you want to check whether the optimizer will perform better without them, set the variable value like this:

mysql> SET optimizer_switch='index_merge_union=off,index_merge_sort_union=off';

mysql> SELECT @@optimizer_switch\G
*************************** 1. row ***************************
@@optimizer_switch: index_merge=on,index_merge_union=off,
                    index_merge_sort_union=off,
                    index_merge_intersection=on,
                    engine_condition_pushdown=on,
                    index_condition_pushdown=on,
                    mrr=on,mrr_cost_based=on,
                    block_nested_loop=on,batched_key_access=off,
                    materialization=on,semijoin=on,loosescan=on,
                    firstmatch=on,
                    subquery_materialization_cost_based=on,
                    use_index_extensions=on,
                    condition_fanout_filter=on

8.9.3 Optimizer Hints

One means of control over optimizer strategies is to set the optimizer_switch system variable (see Section 8.9.2, “Controlling Switchable Optimizations”). Changes to this variable affect execution of all subsequent queries; to affect one query differently from another, it's necessary to change optimizer_switch before each one.

As of MySQL 5.7.7, another way to control the optimizer is by using optimizer hints, which can be specified within individual statements. Because optimizer hints apply on a per-statement basis, they provide finer control over statement execution plans than can be achieved using optimizer_switch. For example, you can enable an optimization for one table in a statement and disable the optimization for a different table. Hints within a statement take precedence over optimizer_switch flags.

Examples:

SELECT /*+ NO_RANGE_OPTIMIZATION(t3 PRIMARY, f2_idx) */ f1
  FROM t3 WHERE f1 > 30 AND f1 < 33;
SELECT /*+ BKA(t1) NO_BKA(t2) */ * FROM t1 INNER JOIN t2 WHERE ...;
SELECT /*+ NO_ICP(t1, t2) */ * FROM t1 INNER JOIN t2 WHERE ...;
SELECT /*+ SEMIJOIN(FIRSTMATCH, LOOSESCAN) */ * FROM t1 ...;
EXPLAIN SELECT /*+ NO_ICP(t1) */ * FROM t1 WHERE ...;

Optimizer hints, described here, differ from index hints, described in Section 8.9.4, “Index Hints”. Optimizer and index hints may be used separately or together.

Optimizer hints apply at different scope levels:

  • Global: The hint affects the entire statement

  • Query block: The hint affects a particular query block within a statement

  • Table-level: The hint affects a particular table within a query block

  • Index-level: The hint affects a particular index within a table

The following table summarizes the available optimizer hints, the optimizer strategies they affect, and the scope or scopes at which they apply. More details are given later.

Table 8.2 Optimizer Hints Available

Hint Name Description Applicable Scopes
BKA, NO_BKA Affects Batched Key Access join processing Query block, table
BNL, NO_BNL Affects Block Nested-Loop join processing Query block, table
MAX_EXECUTION_TIME Limits statement execution time Global
MRR, NO_MRR Affects Multi-Range Read optimization Table, index
NO_ICP Affects Index Condition Pushdown optimization Table, index
NO_RANGE_OPTIMIZATION Affects range optimization Table, index
QB_NAME Assigns name to query block Query block
SEMIJOIN, NO_SEMIJOIN Affects semi-join strategies Query block
SUBQUERY Affects materialization, IN-to-EXISTS subquery stratgies Query block

Disabling an optimization prevents the optimizer from using it. Enabling an optimization means the optimizer is free to use the strategy if it applies to statement execution, not that the optimizer necessarily will use it.

Optimizer Hint Overview

MySQL supports comments in SQL statements as described in Section 9.6, “Comment Syntax”. Optimizer hints use a variant of /* ... */ C-style comment syntax that includes a + character following the /* comment opening sequence. Examples:

/*+ BKA(t1) */
/*+ BNL(t1, t2) */
/*+ NO_RANGE_OPTIMIZATION(t4 PRIMARY) */
/*+ QB_NAME(qb2) */

Whitespace is permitted after the + character.

The parser recognizes optimizer hint comments after the initial keyword of SELECT, UPDATE, INSERT, REPLACE, and DELETE statements. Hints are permitted in these contexts:

  • At the beginning of DML statements:

    SELECT /*+ ... */ ...
    INSERT /*+ ... */ ...
    REPLACE /*+ ... */ ...
    UPDATE /*+ ... */ ...
    DELETE /*+ ... */ ...
    
  • At the beginning of query blocks:

    (SELECT /*+ ... */ ... )
    (SELECT ... ) UNION (SELECT /*+ ... */ ... )
    (SELECT /*+ ... */ ... ) UNION (SELECT /*+ ... */ ... )
    UPDATE ... WHERE x IN (SELECT /*+ ... */ ...)
    INSERT ... SELECT /*+ ... */ ...
    
  • In hintable statements prefaced by EXPLAIN. For example:

    EXPLAIN SELECT /*+ ... */ ...
    EXPLAIN UPDATE ... WHERE x IN (SELECT /*+ ... */ ...)
    

    The implication is that you can use EXPLAIN to see how optimizer hints affect execution plans.

A hint comment may contain multiple hints, but a query block cannot contain multiple hint comments. This is valid:

SELECT /*+ BNL(t1) BKA(t2) */ ...

But this is invalid:

SELECT /*+ BNL(t1) */ /* BKA(t2) */ ...

When a hint comment contains multiple hints, the possibility of duplicates and conflicts exists:

  • Duplicate hints: For a hint such as /*+ MRR(idx1) MRR(idx1) */, MySQL uses the first hint and issues a warning about the duplicate hint.

  • Conflicting hints: For a hint such as /*+ MRR(idx1) NO_MRR(idx1) */, MySQL uses the first hint and issues a warning about the second conflicting hint.

Query block names are identifiers and follow the usual rules about what names are valid and how to quote them (see Section 9.2, “Schema Object Names”).

Hint names, query block names, and strategy names are not case sensitive. References to table and index names follow the usual identifier case sensitivity rules (see Section 9.2.2, “Identifier Case Sensitivity”), except that table name comparisons do not use the lower_case_table_names value until MySQL 5.7.8.

Table-Level Optimizer Hints

Table-level hints affect use of the Block Nested-Loop (BNL) and Batched Key Access (BKA) join-processing algorithms (see Section 8.2.1.14, “Block Nested-Loop and Batched Key Access Joins”). These hint types apply to specific tables, or all tables in a query block.

Syntax of table-level hints:

hint_name([@query_block_name] [tbl_name [, tbl_name] ...])
hint_name([tbl_name@query_block_name [, tbl_name@query_block_name] ...])

The syntax refers to these terms:

  • hint_name: These hint names are permitted:

    • BNL, NO_BNL: Enable or disable BNL for the specified tables.

    • BKA, NO_BKA: Enable or disable BKA for the specified tables.

  • tbl_name: The name of a table used in the statement. The hint applies to all tables that it names. If the hint names no tables, it applies to all tables of the query block in which it occurs.

    If a table has an alias, hints must refer to the alias, not the table name hints.

    Table names in hints cannot be qualified with schema names.

  • query_block_name: The query block to which the hint applies. If the hint includes no leading @query_block_name, the hint applies to the query block in which it occurs. For tbl_name@query_block_name syntax, the hint applies to the named table in the named query block. To assign a name to a query block, see Optimizer Hints for Naming Query Blocks.

Examples:

SELECT /*+ NO_BNL() BKA(t1) */ t1.* FROM t1 INNER JOIN t2 INNER JOIN t3;
SELECT /*+ NO_BKA(t1, t2) */ t1.* FROM t1 INNER JOIN t2 INNER JOIN t3;

A table-level hint applies to tables that receive records from previous tables, not sender tables. Consider this statement:

SELECT /*+ BNL(t2) */ FROM t1, t2;

If the optimizer chooses to process t1 first, it applies a Block Nested-Loop join to t2 by buffering the rows from t1 before starting to read from t2. If the optimizer instead chooses to process t2 first, the hint has no effect because t2 is a sender table.

Index-Level Optimizer Hints

Index-level hints affect which index-processing strategies the optimizer uses for particular tables or indexes. These hint types affect use of Index Condition Pushdown (ICP), Multi-Range Read (MRR), and range optimizations (see Section 8.2.1, “Optimizing SELECT Statements”).

Syntax of index-level hints:

hint_name([@query_block_name] tbl_name [index_name [, index_name] ...])
hint_name(tbl_name@query_block_name [index_name [, index_name] ...])

The syntax refers to these terms:

  • hint_name: These hint names are permitted:

    • MRR, NO_MRR: Enable or disable MRR for the specified tables or indexes. MRR hints apply only to InnoDB and MyISAM tables.

    • NO_ICP: Disable ICP for the specified tables or indexes. By default, ICP is a candidate optimization strategy, so there is no hint for enabling it.

    • NO_RANGE_OPTIMIZATION: Disable index range access for the specified tables or indexes. This hint also disables Index Merge and Loose Index Scan for the tables or indexes. By default, range access is a candidate optimization strategy, so there is no hint for enabling it.

      This hint may be useful when the number of ranges may be high and range optimization would require many resources.

  • tbl_name: The table to which the hint applies.`

  • index_name: The name of an index in the named table. The hint applies to all indexes that it names. If the hint names no indexes, it applies to all indexes in the table.

    To refer to a primary key, use the name PRIMARY. To see the index names for a table, use SHOW INDEX.

  • query_block_name: The query block to which the hint applies. If the hint includes no leading @query_block_name, the hint applies to the query block in which it occurs. For tbl_name@query_block_name syntax, the hint applies to the named table in the named query block. To assign a name to a query block, see Optimizer Hints for Naming Query Blocks.

Examples:

SELECT /*+ MRR(t1) */ * FROM t1 WHERE f2 <= 3 AND 3 <= f3;
SELECT /*+ NO_RANGE_OPTIMIZATION(t3 PRIMARY, f2_idx) */ f1
  FROM t3 WHERE f1 > 30 AND f1 < 33;
INSERT INTO t3(f1, f2, f3)
  (SELECT /*+ NO_ICP(t2) */ t2.f1, t2.f2, t2.f3 FROM t1,t2
   WHERE t1.f1=t2.f1 AND t2.f2 BETWEEN t1.f1
   AND t1.f2 AND t2.f2 + 1 >= t1.f1 + 1);

Subquery Optimizer Hints

Subquery hints (added in MySQL 5.7.8) affect whether to use semi-join transformations and which semi-join strategies to permit, and, when semi-joins are not used, whether to use subquery materialization or IN-to-EXISTS transformations. For more information about these optimizations, see Section 8.2.1.18, “Subquery Optimization”.

Syntax of hints that affect semi-join strategies:

hint_name([@query_block_name] [strategy [, strategy] ...])

The syntax refers to these terms:

  • hint_name: These hint names are permitted:

    • SEMIJOIN, NO_SEMIJOIN: Enable or disable the named semi-join strategies.

  • strategy: A semi-join strategy to be enabled or disabled. These strategy names are permitted: DUPSWEEDOUT, FIRSTMATCH, LOOSESCAN, MATERIALIZATION.

    For SEMIJOIN() hints, if no strategies are named, semi-join is used if possible based on the strategies enabled according to the optimizer_switch system variable. If strategies are named but inapplicable for the statement, DUPSWEEDOUT is used.

    For NO_SEMIJOIN() hints, if no strategies are named, semi-join is not used. If strategies are named that rule out all applicable strategies for the statement, DUPSWEEDOUT is used.

If one subquery is nested within another and both are merged into a semi-join of an outer query, any specification of semi-join strategies for the innermost query are ignored. SEMIJOIN() and NO_SEMIJOIN() hints can still be used to enable or disable semi-join transformations for such nested subqueries.

If DUPSWEEDOUT is disabled, on occasion the optimizer may generate a query plan that is far from optimal. This occurs due to heuristic pruning during greedy search, which can be avoided by setting optimizer_prune_level=0.

Examples:

SELECT /*+ NO_SEMIJOIN(@subq1 FIRSTMATCH, LOOSESCAN) */ * FROM t2
  WHERE t2.a IN (SELECT /*+ QB_NAME(subq1) */ a FROM t3);
SELECT /*+ SEMIJOIN(@subq1 MATERIALIZATION, DUPSWEEDOUT) */ * FROM t2
  WHERE t2.a IN (SELECT /*+ QB_NAME(subq1) */ a FROM t3);

Syntax of hints that affect whether to use subquery materialization or IN-to-EXISTS transformations:

SUBQUERY([@query_block_name] strategy)

The hint name is always SUBQUERY.

For SUBQUERY() hints, these strategy values are permitted: INTOEXISTS, MATERIALIZATION.

Examples:

SELECT id, a IN (SELECT /*+ SUBQUERY(MATERIALIZATION) */ a FROM t1) FROM t2;
SELECT * FROM t2 WHERE t2.a IN (SELECT /*+ SUBQUERY(INTOEXISTS) */ a FROM t1);

For semi-join and SUBQUERY() hints, a leading @query_block_name specifies the query block to which the hint applies. If the hint includes no leading @query_block_name, the hint applies to the query block in which it occurs. To assign a name to a query block, see Optimizer Hints for Naming Query Blocks.

If a hint comment contains multiple subquery hints, the first is used. If there are other following hints of that type, they produce a warning. Following hints of other types are silently ignored.

Statement Execution Time Optimizer Hints

The MAX_EXECUTION_TIME() hint is permitted only for SELECT statements. It places a limit N (a timeout value in milliseconds) on how long a statement is permitted to execute before the server terminates it:

MAX_EXECUTION_TIME(N)

Example with a timeout of 1 second (1000 milliseconds):

SELECT /*+ MAX_EXECUTION_TIME(1000) */ * FROM t1 INNER JOIN t2 WHERE ...

The MAX_EXECUTION_TIME(N) hint MAX_EXECUTION_TIME(N) sets a statement execution timeout of N milliseconds. If this option is absent or N is 0, the statement timeout established by the max_execution_time system variable applies. (Prior to MySQL 5.7.8, this variable was named max_statement_time.)

The MAX_EXECUTION_TIME() hint is applicable as follows:

  • For statements with multiple SELECT keywords, such as unions or statements with subqueries, MAX_EXECUTION_TIME() applies to the entire statement and must appear after the first SELECT.

  • It applies to read-only SELECT statements. Statements that are not read only are those that invoke a stored function that modifies data as a side effect.

  • It does not apply to SELECT statements in stored programs and is ignored.

Optimizer Hints for Naming Query Blocks

Table-level, index-level, and subquery optimizer hints permit specific query blocks to be named as part of their argument syntax. To create these names, use the QB_NAME() hint, which assigns a name to the query block in which it occurs:

QB_NAME(name)

QB_NAME() hints can be used to make explicit in a clear way which query blocks other hints apply to. They also permit all non-query block name hints to be specified within a single hint comment for easier understanding of complex statements. Consider the following statement:

SELECT ...
  FROM (SELECT ...
  FROM (SELECT ... FROM ...)) ...

QB_NAME() hints assign names to query blocks in the statement:

SELECT /*+ QB_NAME(qb1) */ ...
  FROM (SELECT /*+ QB_NAME(qb2) */ ...
  FROM (SELECT /*+ QB_NAME(qb3) */ ... FROM ...)) ...

Then other hints can use those names to refer to the appropriate query blocks:

SELECT /*+ QB_NAME(qb1) MRR(@qb1 t1) BKA(@qb2) NO_MRR(@qb3t1 idx1, id2) */ ...
  FROM (SELECT /*+ QB_NAME(qb2) */ ...
  FROM (SELECT /*+ QB_NAME(qb3) */ ... FROM ...)) ...

The resulting effect is as follows:

  • MRR(@qb1 t1) applies to table t1 in query block qb1.

  • BKA(@qb2) applies to query block qb2.

  • NO_MRR(@qb3 t1 idx1, id2) applies to indexes idx1 and idx2 in table t1 in query block qb3.

Query block names are identifiers and follow the usual rules about what names are valid and how to quote them (see Section 9.2, “Schema Object Names”). For example, a query block name that contains spaces must be quoted, which can be done using backticks:

SELECT /*+ BKA(@`my hint name`) */ ...
  FROM (SELECT /*+ QB_NAME(`my hint name`) */ ...) ...

If the ANSI_QUOTES SQL mode is enabled, it is also possible to quote query block names within double quotation marks:

SELECT /*+ BKA(@"my hint name") */ ...
  FROM (SELECT /*+ QB_NAME("my hint name") */ ...) ...

8.9.4 Index Hints

Index hints give the optimizer information about how to choose indexes during query processing. Index hints, described here, differ from optimizer hints, described in Section 8.9.3, “Optimizer Hints”. Index and optimizer hints may be used separately or together.

Index hints are specified following a table name. (For the general syntax for specifying tables in a SELECT statement, see Section 13.2.9.2, “JOIN Syntax”.) The syntax for referring to an individual table, including index hints, looks like this:

tbl_name [[AS] alias] [index_hint_list]

index_hint_list:
    index_hint [, index_hint] ...

index_hint:
    USE {INDEX|KEY}
      [FOR {JOIN|ORDER BY|GROUP BY}] ([index_list])
  | IGNORE {INDEX|KEY}
      [FOR {JOIN|ORDER BY|GROUP BY}] (index_list)
  | FORCE {INDEX|KEY}
      [FOR {JOIN|ORDER BY|GROUP BY}] (index_list)

index_list:
    index_name [, index_name] ...

The USE INDEX (index_list) hint tells MySQL to use only one of the named indexes to find rows in the table. The alternative syntax IGNORE INDEX (index_list) tells MySQL to not use some particular index or indexes. These hints are useful if EXPLAIN shows that MySQL is using the wrong index from the list of possible indexes.

The FORCE INDEX hint acts like USE INDEX (index_list), with the addition that a table scan is assumed to be very expensive. In other words, a table scan is used only if there is no way to use one of the named indexes to find rows in the table.

Each hint requires the names of indexes, not the names of columns. To refer to a primary key, use the name PRIMARY. To see the index names for a table, use SHOW INDEX.

An index_name value need not be a full index name. It can be an unambiguous prefix of an index name. If a prefix is ambiguous, an error occurs.

Examples:

SELECT * FROM table1 USE INDEX (col1_index,col2_index)
  WHERE col1=1 AND col2=2 AND col3=3;

SELECT * FROM table1 IGNORE INDEX (col3_index)
  WHERE col1=1 AND col2=2 AND col3=3;

The syntax for index hints has the following characteristics:

  • It is syntactically valid to omit index_list for USE INDEX, which means use no indexes. Omitting index_list for FORCE INDEX or IGNORE INDEX is a syntax error.

  • You can specify the scope of an index hint by adding a FOR clause to the hint. This provides more fine-grained control over the optimizer's selection of an execution plan for various phases of query processing. To affect only the indexes used when MySQL decides how to find rows in the table and how to process joins, use FOR JOIN. To influence index usage for sorting or grouping rows, use FOR ORDER BY or FOR GROUP BY.

  • You can specify multiple index hints:

    SELECT * FROM t1 USE INDEX (i1) IGNORE INDEX FOR ORDER BY (i2) ORDER BY a;
    

    It is not an error to name the same index in several hints (even within the same hint):

    SELECT * FROM t1 USE INDEX (i1) USE INDEX (i1,i1);
    

    However, it is an error to mix USE INDEX and FORCE INDEX for the same table:

    SELECT * FROM t1 USE INDEX FOR JOIN (i1) FORCE INDEX FOR JOIN (i2);
    

If an index hint includes no FOR clause, the scope of the hint is to apply to all parts of the statement. For example, this hint:

IGNORE INDEX (i1)

is equivalent to this combination of hints:

IGNORE INDEX FOR JOIN (i1)
IGNORE INDEX FOR ORDER BY (i1)
IGNORE INDEX FOR GROUP BY (i1)

In MySQL 5.0, hint scope with no FOR clause was to apply only to row retrieval. To cause the server to use this older behavior when no FOR clause is present, enable the old system variable at server startup. Take care about enabling this variable in a replication setup. With statement-based binary logging, having different modes for the master and slaves might lead to replication errors.

When index hints are processed, they are collected in a single list by type (USE, FORCE, IGNORE) and by scope (FOR JOIN, FOR ORDER BY, FOR GROUP BY). For example:

SELECT * FROM t1
  USE INDEX () IGNORE INDEX (i2) USE INDEX (i1) USE INDEX (i2);

is equivalent to:

SELECT * FROM t1
   USE INDEX (i1,i2) IGNORE INDEX (i2);

The index hints then are applied for each scope in the following order:

  1. {USE|FORCE} INDEX is applied if present. (If not, the optimizer-determined set of indexes is used.)

  2. IGNORE INDEX is applied over the result of the previous step. For example, the following two queries are equivalent:

    SELECT * FROM t1 USE INDEX (i1) IGNORE INDEX (i2) USE INDEX (i2);
    
    SELECT * FROM t1 USE INDEX (i1);
    

For FULLTEXT searches, index hints work as follows:

  • For natural language mode searches, index hints are silently ignored. For example, IGNORE INDEX(i1) is ignored with no warning and the index is still used.

  • For boolean mode searches, index hints with FOR ORDER BY or FOR GROUP BY are silently ignored. Index hints with FOR JOIN or no FOR modifier are honored. In contrast to how hints apply for non-FULLTEXT searches, the hint is used for all phases of query execution (finding rows and retrieval, grouping, and ordering). This is true even if the hint is given for a non-FULLTEXT index.

    For example, the following two queries are equivalent:

    SELECT * FROM t
      USE INDEX (index1)
      IGNORE INDEX (index1) FOR ORDER BY
      IGNORE INDEX (index1) FOR GROUP BY
      WHERE ... IN BOOLEAN MODE ... ;
    
    SELECT * FROM t
      USE INDEX (index1)
      WHERE ... IN BOOLEAN MODE ... ;
    

8.9.5 The Optimizer Cost Model

To generate execution plans, the optimizer uses a cost model that is based on estimates of the cost of various operations that occur during query execution. The optimizer has a set of compiled-in default cost constants available to it to make decisions regarding execution plans.

As of MySQL 5.7.5, the optimizer has in addition a database of cost estimates to use during execution plan construction. These estimates are stored in the server_cost and engine_cost tables in the mysql system database and are configurable at any time. The intent of these tables is to make it possible to easily adjust the cost estimates that the optimizer uses when it attempts to arrive at query execution plans.

Cost Model General Operation

The configurable optimizer cost model works like this:

  • The server reads the cost model tables into memory at startup and uses the in-memory values at runtime. Any non-NULL cost estimate specified in the tables takes precedence over the corresponding compiled-in default cost constant. Any NULL estimate indicates to the optimizer to use the compiled-in default.

  • At runtime, the server may reread the cost tables. This occurs when a storage engine is dynamically loaded or when a FLUSH OPTIMIZER_COSTS statement is executed.

  • Cost tables enable server administrators to easily adjust cost estimates by changing entries in the tables. It is also easy to revert to a default by setting an entry's cost to NULL. The optimizer uses the in-memory cost values, so changes to the tables should be followed by FLUSH OPTIMIZER_COSTS to take effect.

  • The in-memory cost estimates that are current when a client session begins apply throughout that session until it ends. In particular, if the server rereads the cost tables, any changed estimates apply only to subsequently started sessions. Existing sessions are unaffected.

  • Cost tables are specific to a given server instance. The server does not replicate cost table changes to replication slaves.

The Cost Model Database

The optimizer cost model database consists of two tables in the mysql system database that contain cost estimate information for operations that occur during query execution:

  • server_cost: Optimizer cost estimates for general server operations

  • engine_cost: Optimizer cost estimates for operations specific to particular storage engines

The server_cost table contains these columns:

  • cost_name

    The name of a cost estimate used in the cost model. The name is not case sensitive. If the server does not recognize the cost name when it reads this table, it writes a warning to the error log.

  • cost_value

    The cost estimate value. If the value is non-NULL, the server uses it as the cost. Otherwise, it uses the default estimate (the compiled-in value). DBAs can change a cost estimate by updating this column. If the server finds that the cost value is invalid (nonpositive) when it reads this table, it writes a warning to the error log.

    To override a default cost estimate (for an entry that specifies NULL), set the cost to a non-NULL value. To revert to the default, set the value to NULL. Then execute FLUSH OPTIMIZER_COSTS to tell the server to reread the cost tables.

  • last_update

    The time of the last row update.

  • comment

    A descriptive comment associated with the cost estimate. DBAs can use this column to provide information about why a cost estimate row stores a particular value.

The primary key for the server_cost table is the cost_name column, so it is not possible to create multiple entries for any cost estimate.

The server recognizes these cost_name values for the server_cost table:

  • disk_temptable_create_cost (default 40.0), disk_temptable_row_cost (default 1.0)

    The cost estimates for internally created temporary tables stored in a disk-based storage engine (either InnoDB or MyISAM). Increasing these values increases the cost estimate of using internal temporary tables and makes the optimizer prefer query plans with less use of them. For information about such tables, see Section 8.4.4, “Internal Temporary Table Use in MySQL”.

    The larger default values for these disk parameters compared to the default values for the corresponding memory parameters (memory_temptable_create_cost, memory_temptable_row_cost) reflects the greater cost of processing disk-based tables.

  • key_compare_cost (default 0.1)

    The cost of comparing record keys. Increasing this value causes a query plan that compares many keys to become more expensive. For example, a query plan that performs a filesort becomes relatively more expensive compared to a query plan that avoids sorting by using an index.

  • memory_temptable_create_cost (default 2.0), memory_temptable_row_cost (default 0.2)

    The cost estimates for internally created temporary tables stored in the MEMORY storage engine. Increasing these values increases the cost estimate of using internal temporary tables and makes the optimizer prefer query plans with less use of them. For information about such tables, see Section 8.4.4, “Internal Temporary Table Use in MySQL”.

    The smaller default values for these memory parameters compared to the default values for the corresponding disk parameters (disk_temptable_create_cost, disk_temptable_row_cost) reflects the lesser cost of processing memory-based tables.

  • row_evaluate_cost (default 0.2)

    The cost of evaluating record conditions. Increasing this value causes a query plan that examines many rows to become more expensive compared to a query plan that examines fewer rows. For example, a table scan becomes relatively more expensive compared to a range scan that reads fewer rows.

The engine_cost table contains these columns:

  • engine_name

    The name of the storage engine to which this cost estimate applies. The name is not case sensitive. If the value is default, it applies to all storage engines that have no named entry of their own. If the server does not recognize the engine name when it reads this table, it writes a warning to the error log.

  • device_type

    The device type to which this cost estimate applies. The column is intended for specifying different cost estimates for different storage device types, such as hard disk drives versus solid state drives. Currently, this information is not used and 0 is the only permitted value.

  • cost_name

    Same as in the server_cost table.

  • cost_value

    Same as in the server_cost table.

  • last_update

    Same as in the server_cost table.

  • comment

    Same as in the server_cost table.

The primary key for the engine_cost table is a tuple comprising the (cost_name, engine_name, device_type) columns, so it is not possible to create multiple entries for any combination of values in those columns.

The server recognizes these cost_name values for the engine_cost table:

  • io_block_read_cost (default 1.0)

    The cost of reading an index or data block from disk. Increasing this value causes a query plan that reads many disk blocks to become more expensive compared to a query plan that reads fewer disk blocks. For example, a table scan becomes relatively more expensive compared to a range scan that reads fewer blocks.

  • memory_block_read_cost (default 1.0)

    Similar to io_block_read_cost, but represents the cost of reading an index or data block from an in-memory database buffer. This cost parameter was added in MySQL 5.7.8.

Making Changes to the Cost Model Database

For DBAs who wish to change the cost model parameters from their defaults, try doubling or halving the value and measuring the effect.

Changes to the io_block_read_cost and memory_block_read_cost parameters are most likely to yield worthwhile results. These parameter values enable cost models for data access methods to take into account the costs of reading information from different sources; that is, the cost of reading information from disk versus reading information already in a memory buffer. For example, all other things being equal, setting io_block_read_cost to a value larger than memory_block_read_cost causes the optimizer to prefer query plans that read information already held in memory to plans that must read from disk.

This example shows how to change the default value for io_block_read_cost:

UPDATE mysql.engine_cost
  SET cost_value = 2.0
  WHERE cost_name = 'io_block_read_cost';
FLUSH OPTIMIZER_COSTS;

This example shows how to change the value of io_block_read_cost only for the InnoDB storage engine:

INSERT INTO mysql.engine_cost
  VALUES ('InnoDB', 0, 'io_block_read_cost', 3.0,
  CURRENT_TIMESTAMP, 'Using a slower disk for InnoDB');
FLUSH OPTIMIZER_COSTS;

8.10 Buffering and Caching

MySQL uses several strategies that cache information in memory buffers to increase performance.

8.10.1 The InnoDB Buffer Pool

InnoDB maintains a storage area called the buffer pool for caching data and indexes in memory. Knowing how the InnoDB buffer pool works, and taking advantage of it to keep frequently accessed data in memory, is an important aspect of MySQL tuning.

Note

For additional information about the InnoDB buffer pool, see Section 14.4.3, “InnoDB Buffer Pool Configuration”.

Guidelines

Ideally, you set the size of the buffer pool to as large a value as practical, leaving enough memory for other processes on the server to run without excessive paging. The larger the buffer pool, the more InnoDB acts like an in-memory database, reading data from disk once and then accessing the data from memory during subsequent reads. The buffer pool even caches data changed by insert and update operations, so that disk writes can be grouped together for better performance.

Depending on the typical workload on your system, you might adjust the proportions of the parts within the buffer pool. You can tune the way the buffer pool chooses which blocks to cache once it fills up, to keep frequently accessed data in memory despite sudden spikes of activity for operations such as backups or reporting.

With 64-bit systems with large memory sizes, you can split the buffer pool into multiple parts, to minimize contention for the memory structures among concurrent operations. For details, see Section 14.4.3.4, “Using Multiple Buffer Pool Instances”.

Internal Details

InnoDB manages the pool as a list, using a variation of the least recently used (LRU) algorithm. When room is needed to add a new block to the pool, InnoDB evicts the least recently used block and adds the new block to the middle of the list. This midpoint insertion strategy treats the list as two sublists:

  • At the head, a sublist of new (or young) blocks that were accessed recently.

  • At the tail, a sublist of old blocks that were accessed less recently.

This algorithm keeps blocks that are heavily used by queries in the new sublist. The old sublist contains less-used blocks; these blocks are candidates for eviction.

The LRU algorithm operates as follows by default:

  • 3/8 of the buffer pool is devoted to the old sublist.

  • The midpoint of the list is the boundary where the tail of the new sublist meets the head of the old sublist.

  • When InnoDB reads a block into the buffer pool, it initially inserts it at the midpoint (the head of the old sublist). A block can be read in because it is required for a user-specified operation such as an SQL query, or as part of a read-ahead operation performed automatically by InnoDB.

  • Accessing a block in the old sublist makes it young, moving it to the head of the buffer pool (the head of the new sublist). If the block was read in because it was required, the first access occurs immediately and the block is made young. If the block was read in due to read-ahead, the first access does not occur immediately (and might not occur at all before the block is evicted).

  • As the database operates, blocks in the buffer pool that are not accessed age by moving toward the tail of the list. Blocks in both the new and old sublists age as other blocks are made new. Blocks in the old sublist also age as blocks are inserted at the midpoint. Eventually, a block that remains unused for long enough reaches the tail of the old sublist and is evicted.

By default, blocks read by queries immediately move into the new sublist, meaning they will stay in the buffer pool for a long time. A table scan (such as performed for a mysqldump operation, or a SELECT statement with no WHERE clause) can bring a large amount of data into the buffer pool and evict an equivalent amount of older data, even if the new data is never used again. Similarly, blocks that are loaded by the read-ahead background thread and then accessed only once move to the head of the new list. These situations can push frequently used blocks to the old sublist, where they become subject to eviction.

Configuration Options

Several InnoDB system variables control the size of the buffer pool and let you tune the LRU algorithm:

  • innodb_buffer_pool_size

    Specifies the size of the buffer pool. If your buffer pool is small and you have sufficient memory, making the pool larger can improve performance by reducing the amount of disk I/O needed as queries access InnoDB tables. As of MySQL 5.7.5, the innodb_buffer_pool_size option is dynamic, which allows you to configure buffer pool size without restarting the server. See Configuring InnoDB Buffer Pool Size Online for more information.

  • innodb_buffer_pool_instances

    Divides the buffer pool into a user-specified number of separate regions, each with its own LRU list and related data structures, to reduce contention during concurrent memory read and write operations. This option takes effect only when you set the innodb_buffer_pool_size to a size of 1 gigabyte or more. The total size you specify is divided among all the buffer pools. For best efficiency, specify a combination of innodb_buffer_pool_instances and innodb_buffer_pool_size so that each buffer pool instance is at least 1 gigabyte.

  • innodb_old_blocks_pct

    Specifies the approximate percentage of the buffer pool that InnoDB uses for the old block sublist. The range of values is 5 to 95. The default value is 37 (that is, 3/8 of the pool).

  • innodb_old_blocks_time

    Specifies how long in milliseconds (ms) a block inserted into the old sublist must stay there after its first access before it can be moved to the new sublist. The default value is 0: A block inserted into the old sublist moves to the new sublist when Innodb has evicted 1/4 of the inserted block's pages from the buffer pool, no matter how soon after insertion the access occurs. If the value is greater than 0, blocks remain in the old sublist until an access occurs at least that many ms after the first access. For example, a value of 1000 causes blocks to stay in the old sublist for 1 second after the first access before they become eligible to move to the new sublist.

Setting innodb_old_blocks_time greater than 0 prevents one-time table scans from flooding the new sublist with blocks used only for the scan. Rows in a block read in for a scan are accessed many times in rapid succession, but the block is unused after that. If innodb_old_blocks_time is set to a value greater than time to process the block, the block remains in the old sublist and ages to the tail of the list to be evicted quickly. This way, blocks used only for a one-time scan do not act to the detriment of heavily used blocks in the new sublist.

innodb_old_blocks_time can be set at runtime, so you can change it temporarily while performing operations such as table scans and dumps:

SET GLOBAL innodb_old_blocks_time = 1000;
... perform queries that scan tables ...
SET GLOBAL innodb_old_blocks_time = 0;

This strategy does not apply if your intent is to warm up the buffer pool by filling it with a table's content. For example, benchmark tests often perform a table or index scan at server startup, because that data would normally be in the buffer pool after a period of normal use. In this case, leave innodb_old_blocks_time set to 0, at least until the warmup phase is complete.

Monitoring the Buffer Pool Using the InnoDB Standard Monitor

The output from the InnoDB Standard Monitor contains several fields in the BUFFER POOL AND MEMORY section that pertain to operation of the buffer pool LRU algorithm:

  • Old database pages: The number of pages in the old sublist of the buffer pool.

  • Pages made young, not young: The number of old pages that were moved to the head of the buffer pool (the new sublist), and the number of pages that have remained in the old sublist without being made new.

  • youngs/s non-youngs/s: The number of accesses to old pages that have resulted in making them young or not. This metric differs from that of the previous item in two ways. First, it relates only to old pages. Second, it is based on number of accesses to pages and not the number of pages. (There can be multiple accesses to a given page, all of which are counted.)

  • young-making rate: Hits that cause blocks to move to the head of the buffer pool.

  • not: Hits that do not cause blocks to move to the head of the buffer pool (due to the delay not being met).

The young-making rate and not rate will not normally add up to the overall buffer pool hit rate. Hits for blocks in the old sublist cause them to move to the new sublist, but hits to blocks in the new sublist cause them to move to the head of the list only if they are a certain distance from the head.

The preceding information from the Monitor can help you make LRU tuning decisions:

  • If you see very low youngs/s values when you do not have large scans going on, that indicates that you might need to either reduce the delay time, or increase the percentage of the buffer pool used for the old sublist. Increasing the percentage makes the old sublist larger, so blocks in that sublist take longer to move to the tail and be evicted. This increases the likelihood that they will be accessed again and be made young.

  • If you do not see a lot of non-youngs/s when you are doing large table scans (and lots of youngs/s), to tune your delay value to be larger.

Note

Per second averages provided in InnoDB Monitor output are based on the elapsed time between the current time and the last time InnoDB Monitor output was printed.

For more information about InnoDB Monitors, see Section 14.15, “InnoDB Monitors”.

The INNODB_BUFFER_POOL_STATS table and InnoDB buffer pool server status variables provide much of the same buffer pool information that is provided by SHOW ENGINE INNODB STATUS output.

8.10.2 The MyISAM Key Cache

To minimize disk I/O, the MyISAM storage engine exploits a strategy that is used by many database management systems. It employs a cache mechanism to keep the most frequently accessed table blocks in memory:

  • For index blocks, a special structure called the key cache (or key buffer) is maintained. The structure contains a number of block buffers where the most-used index blocks are placed.

  • For data blocks, MySQL uses no special cache. Instead it relies on the native operating system file system cache.

This section first describes the basic operation of the MyISAM key cache. Then it discusses features that improve key cache performance and that enable you to better control cache operation:

  • Multiple sessions can access the cache concurrently.

  • You can set up multiple key caches and assign table indexes to specific caches.

To control the size of the key cache, use the key_buffer_size system variable. If this variable is set equal to zero, no key cache is used. The key cache also is not used if the key_buffer_size value is too small to allocate the minimal number of block buffers (8).

When the key cache is not operational, index files are accessed using only the native file system buffering provided by the operating system. (In other words, table index blocks are accessed using the same strategy as that employed for table data blocks.)

An index block is a contiguous unit of access to the MyISAM index files. Usually the size of an index block is equal to the size of nodes of the index B-tree. (Indexes are represented on disk using a B-tree data structure. Nodes at the bottom of the tree are leaf nodes. Nodes above the leaf nodes are nonleaf nodes.)

All block buffers in a key cache structure are the same size. This size can be equal to, greater than, or less than the size of a table index block. Usually one these two values is a multiple of the other.

When data from any table index block must be accessed, the server first checks whether it is available in some block buffer of the key cache. If it is, the server accesses data in the key cache rather than on disk. That is, it reads from the cache or writes into it rather than reading from or writing to disk. Otherwise, the server chooses a cache block buffer containing a different table index block (or blocks) and replaces the data there by a copy of required table index block. As soon as the new index block is in the cache, the index data can be accessed.

If it happens that a block selected for replacement has been modified, the block is considered dirty. In this case, prior to being replaced, its contents are flushed to the table index from which it came.

Usually the server follows an LRU (Least Recently Used) strategy: When choosing a block for replacement, it selects the least recently used index block. To make this choice easier, the key cache module maintains all used blocks in a special list (LRU chain) ordered by time of use. When a block is accessed, it is the most recently used and is placed at the end of the list. When blocks need to be replaced, blocks at the beginning of the list are the least recently used and become the first candidates for eviction.

The InnoDB storage engine also uses an LRU algorithm, to manage its buffer pool. See Section 8.10.1, “The InnoDB Buffer Pool”.

8.10.2.1 Shared Key Cache Access

Threads can access key cache buffers simultaneously, subject to the following conditions:

  • A buffer that is not being updated can be accessed by multiple sessions.

  • A buffer that is being updated causes sessions that need to use it to wait until the update is complete.

  • Multiple sessions can initiate requests that result in cache block replacements, as long as they do not interfere with each other (that is, as long as they need different index blocks, and thus cause different cache blocks to be replaced).

Shared access to the key cache enables the server to improve throughput significantly.

8.10.2.2 Multiple Key Caches

Shared access to the key cache improves performance but does not eliminate contention among sessions entirely. They still compete for control structures that manage access to the key cache buffers. To reduce key cache access contention further, MySQL also provides multiple key caches. This feature enables you to assign different table indexes to different key caches.

Where there are multiple key caches, the server must know which cache to use when processing queries for a given MyISAM table. By default, all MyISAM table indexes are cached in the default key cache. To assign table indexes to a specific key cache, use the CACHE INDEX statement (see Section 13.7.6.2, “CACHE INDEX Syntax”). For example, the following statement assigns indexes from the tables t1, t2, and t3 to the key cache named hot_cache:

mysql> CACHE INDEX t1, t2, t3 IN hot_cache;
+---------+--------------------+----------+----------+
| Table   | Op                 | Msg_type | Msg_text |
+---------+--------------------+----------+----------+
| test.t1 | assign_to_keycache | status   | OK       |
| test.t2 | assign_to_keycache | status   | OK       |
| test.t3 | assign_to_keycache | status   | OK       |
+---------+--------------------+----------+----------+

The key cache referred to in a CACHE INDEX statement can be created by setting its size with a SET GLOBAL parameter setting statement or by using server startup options. For example:

mysql> SET GLOBAL keycache1.key_buffer_size=128*1024;

To destroy a key cache, set its size to zero:

mysql> SET GLOBAL keycache1.key_buffer_size=0;

You cannot destroy the default key cache. Any attempt to do this is ignored:

mysql> SET GLOBAL key_buffer_size = 0;

mysql> SHOW VARIABLES LIKE 'key_buffer_size';
+-----------------+---------+
| Variable_name   | Value   |
+-----------------+---------+
| key_buffer_size | 8384512 |
+-----------------+---------+

Key cache variables are structured system variables that have a name and components. For keycache1.key_buffer_size, keycache1 is the cache variable name and key_buffer_size is the cache component. See Section 5.1.5.1, “Structured System Variables”, for a description of the syntax used for referring to structured key cache system variables.

By default, table indexes are assigned to the main (default) key cache created at the server startup. When a key cache is destroyed, all indexes assigned to it are reassigned to the default key cache.

For a busy server, you can use a strategy that involves three key caches:

  • A hot key cache that takes up 20% of the space allocated for all key caches. Use this for tables that are heavily used for searches but that are not updated.

  • A cold key cache that takes up 20% of the space allocated for all key caches. Use this cache for medium-sized, intensively modified tables, such as temporary tables.

  • A warm key cache that takes up 60% of the key cache space. Employ this as the default key cache, to be used by default for all other tables.

One reason the use of three key caches is beneficial is that access to one key cache structure does not block access to the others. Statements that access tables assigned to one cache do not compete with statements that access tables assigned to another cache. Performance gains occur for other reasons as well:

  • The hot cache is used only for retrieval queries, so its contents are never modified. Consequently, whenever an index block needs to be pulled in from disk, the contents of the cache block chosen for replacement need not be flushed first.

  • For an index assigned to the hot cache, if there are no queries requiring an index scan, there is a high probability that the index blocks corresponding to nonleaf nodes of the index B-tree remain in the cache.

  • An update operation most frequently executed for temporary tables is performed much faster when the updated node is in the cache and need not be read in from disk first. If the size of the indexes of the temporary tables are comparable with the size of cold key cache, the probability is very high that the updated node is in the cache.

The CACHE INDEX statement sets up an association between a table and a key cache, but the association is lost each time the server restarts. If you want the association to take effect each time the server starts, one way to accomplish this is to use an option file: Include variable settings that configure your key caches, and an init-file option that names a file containing CACHE INDEX statements to be executed. For example:

key_buffer_size = 4G
hot_cache.key_buffer_size = 2G
cold_cache.key_buffer_size = 2G
init_file=/path/to/data-directory/mysqld_init.sql

The statements in mysqld_init.sql are executed each time the server starts. The file should contain one SQL statement per line. The following example assigns several tables each to hot_cache and cold_cache:

CACHE INDEX db1.t1, db1.t2, db2.t3 IN hot_cache
CACHE INDEX db1.t4, db2.t5, db2.t6 IN cold_cache

8.10.2.3 Midpoint Insertion Strategy

By default, the key cache management system uses a simple LRU strategy for choosing key cache blocks to be evicted, but it also supports a more sophisticated method called the midpoint insertion strategy.

When using the midpoint insertion strategy, the LRU chain is divided into two parts: a hot sublist and a warm sublist. The division point between two parts is not fixed, but the key cache management system takes care that the warm part is not too short, always containing at least key_cache_division_limit percent of the key cache blocks. key_cache_division_limit is a component of structured key cache variables, so its value is a parameter that can be set per cache.

When an index block is read from a table into the key cache, it is placed at the end of the warm sublist. After a certain number of hits (accesses of the block), it is promoted to the hot sublist. At present, the number of hits required to promote a block (3) is the same for all index blocks.

A block promoted into the hot sublist is placed at the end of the list. The block then circulates within this sublist. If the block stays at the beginning of the sublist for a long enough time, it is demoted to the warm sublist. This time is determined by the value of the key_cache_age_threshold component of the key cache.

The threshold value prescribes that, for a key cache containing N blocks, the block at the beginning of the hot sublist not accessed within the last N * key_cache_age_threshold / 100 hits is to be moved to the beginning of the warm sublist. It then becomes the first candidate for eviction, because blocks for replacement always are taken from the beginning of the warm sublist.

The midpoint insertion strategy enables you to keep more-valued blocks always in the cache. If you prefer to use the plain LRU strategy, leave the key_cache_division_limit value set to its default of 100.

The midpoint insertion strategy helps to improve performance when execution of a query that requires an index scan effectively pushes out of the cache all the index blocks corresponding to valuable high-level B-tree nodes. To avoid this, you must use a midpoint insertion strategy with the key_cache_division_limit set to much less than 100. Then valuable frequently hit nodes are preserved in the hot sublist during an index scan operation as well.

8.10.2.4 Index Preloading

If there are enough blocks in a key cache to hold blocks of an entire index, or at least the blocks corresponding to its nonleaf nodes, it makes sense to preload the key cache with index blocks before starting to use it. Preloading enables you to put the table index blocks into a key cache buffer in the most efficient way: by reading the index blocks from disk sequentially.

Without preloading, the blocks are still placed into the key cache as needed by queries. Although the blocks will stay in the cache, because there are enough buffers for all of them, they are fetched from disk in random order, and not sequentially.

To preload an index into a cache, use the LOAD INDEX INTO CACHE statement. For example, the following statement preloads nodes (index blocks) of indexes of the tables t1 and t2:

mysql> LOAD INDEX INTO CACHE t1, t2 IGNORE LEAVES;
+---------+--------------+----------+----------+
| Table   | Op           | Msg_type | Msg_text |
+---------+--------------+----------+----------+
| test.t1 | preload_keys | status   | OK       |
| test.t2 | preload_keys | status   | OK       |
+---------+--------------+----------+----------+

The IGNORE LEAVES modifier causes only blocks for the nonleaf nodes of the index to be preloaded. Thus, the statement shown preloads all index blocks from t1, but only blocks for the nonleaf nodes from t2.

If an index has been assigned to a key cache using a CACHE INDEX statement, preloading places index blocks into that cache. Otherwise, the index is loaded into the default key cache.

8.10.2.5 Key Cache Block Size

It is possible to specify the size of the block buffers for an individual key cache using the key_cache_block_size variable. This permits tuning of the performance of I/O operations for index files.

The best performance for I/O operations is achieved when the size of read buffers is equal to the size of the native operating system I/O buffers. But setting the size of key nodes equal to the size of the I/O buffer does not always ensure the best overall performance. When reading the big leaf nodes, the server pulls in a lot of unnecessary data, effectively preventing reading other leaf nodes.

To control the size of blocks in the .MYI index file of MyISAM tables, use the --myisam-block-size option at server startup.

8.10.2.6 Restructuring a Key Cache

A key cache can be restructured at any time by updating its parameter values. For example:

mysql> SET GLOBAL cold_cache.key_buffer_size=4*1024*1024;

If you assign to either the key_buffer_size or key_cache_block_size key cache component a value that differs from the component's current value, the server destroys the cache's old structure and creates a new one based on the new values. If the cache contains any dirty blocks, the server saves them to disk before destroying and re-creating the cache. Restructuring does not occur if you change other key cache parameters.

When restructuring a key cache, the server first flushes the contents of any dirty buffers to disk. After that, the cache contents become unavailable. However, restructuring does not block queries that need to use indexes assigned to the cache. Instead, the server directly accesses the table indexes using native file system caching. File system caching is not as efficient as using a key cache, so although queries execute, a slowdown can be anticipated. After the cache has been restructured, it becomes available again for caching indexes assigned to it, and the use of file system caching for the indexes ceases.

8.10.3 The MySQL Query Cache

The query cache stores the text of a SELECT statement together with the corresponding result that was sent to the client. If an identical statement is received later, the server retrieves the results from the query cache rather than parsing and executing the statement again. The query cache is shared among sessions, so a result set generated by one client can be sent in response to the same query issued by another client.

The query cache can be useful in an environment where you have tables that do not change very often and for which the server receives many identical queries. This is a typical situation for many Web servers that generate many dynamic pages based on database content.

The query cache does not return stale data. When tables are modified, any relevant entries in the query cache are flushed.

Note

The query cache does not work in an environment where you have multiple mysqld servers updating the same MyISAM tables.

The query cache is used for prepared statements under the conditions described in Section 8.10.3.1, “How the Query Cache Operates”.

Note

The query cache is not supported for partitioned tables, and is automatically disabled for queries involving partitioned tables. The query cache cannot be enabled for such queries.

Some performance data for the query cache follows. These results were generated by running the MySQL benchmark suite on a Linux Alpha 2×500MHz system with 2GB RAM and a 64MB query cache.

  • If all the queries you are performing are simple (such as selecting a row from a table with one row), but still differ so that the queries cannot be cached, the overhead for having the query cache active is 13%. This could be regarded as the worst case scenario. In real life, queries tend to be much more complicated, so the overhead normally is significantly lower.

  • Searches for a single row in a single-row table are 238% faster with the query cache than without it. This can be regarded as close to the minimum speedup to be expected for a query that is cached.

To disable the query cache at server startup, set the query_cache_size system variable to 0. By disabling the query cache code, there is no noticeable overhead.

The query cache offers the potential for substantial performance improvement, but do not assume that it will do so under all circumstances. With some query cache configurations or server workloads, you might actually see a performance decrease:

  • Be cautious about sizing the query cache excessively large, which increases the overhead required to maintain the cache, possibly beyond the benefit of enabling it. Sizes in tens of megabytes are usually beneficial. Sizes in the hundreds of megabytes might not be.

  • Server workload has a significant effect on query cache efficiency. A query mix consisting almost entirely of a fixed set of SELECT statements is much more likely to benefit from enabling the cache than a mix in which frequent INSERT statements cause continual invalidation of results in the cache. In some cases, a workaround is to use the SQL_NO_CACHE option to prevent results from even entering the cache for SELECT statements that use frequently modified tables. (See Section 8.10.3.2, “Query Cache SELECT Options”.)

To verify that enabling the query cache is beneficial, test the operation of your MySQL server with the cache enabled and disabled. Then retest periodically because query cache efficiency may change as server workload changes.

8.10.3.1 How the Query Cache Operates

This section describes how the query cache works when it is operational. Section 8.10.3.3, “Query Cache Configuration”, describes how to control whether it is operational.

Incoming queries are compared to those in the query cache before parsing, so the following two queries are regarded as different by the query cache:

SELECT * FROM tbl_name
Select * from tbl_name

Queries must be exactly the same (byte for byte) to be seen as identical. In addition, query strings that are identical may be treated as different for other reasons. Queries that use different databases, different protocol versions, or different default character sets are considered different queries and are cached separately.

The cache is not used for queries of the following types:

  • Queries that are a subquery of an outer query

  • Queries executed within the body of a stored function, trigger, or event

Before a query result is fetched from the query cache, MySQL checks whether the user has SELECT privilege for all databases and tables involved. If this is not the case, the cached result is not used.

If a query result is returned from query cache, the server increments the Qcache_hits status variable, not Com_select. See Section 8.10.3.4, “Query Cache Status and Maintenance”.

If a table changes, all cached queries that use the table become invalid and are removed from the cache. This includes queries that use MERGE tables that map to the changed table. A table can be changed by many types of statements, such as INSERT, UPDATE, DELETE, TRUNCATE TABLE, ALTER TABLE, DROP TABLE, or DROP DATABASE.

The query cache also works within transactions when using InnoDB tables.

The result from a SELECT query on a view is cached.

The query cache works for SELECT SQL_CALC_FOUND_ROWS ... queries and stores a value that is returned by a following SELECT FOUND_ROWS() query. FOUND_ROWS() returns the correct value even if the preceding query was fetched from the cache because the number of found rows is also stored in the cache. The SELECT FOUND_ROWS() query itself cannot be cached.

Prepared statements that are issued using the binary protocol using mysql_stmt_prepare() and mysql_stmt_execute() (see Section 24.8.8, “C API Prepared Statements”), are subject to limitations on caching. Comparison with statements in the query cache is based on the text of the statement after expansion of ? parameter markers. The statement is compared only with other cached statements that were executed using the binary protocol. That is, for query cache purposes, prepared statements issued using the binary protocol are distinct from prepared statements issued using the text protocol (see Section 13.5, “SQL Syntax for Prepared Statements”).

A query cannot be cached if it contains any of the functions shown in the following table.

A query also is not cached under these conditions:

  • It refers to user-defined functions (UDFs) or stored functions.

  • It refers to user variables or local stored program variables.

  • It refers to tables in the mysql, INFORMATION_SCHEMA, or performance_schema database.

  • It refers to any partitioned tables.

  • It is of any of the following forms:

    SELECT ... LOCK IN SHARE MODE
    SELECT ... FOR UPDATE
    SELECT ... INTO OUTFILE ...
    SELECT ... INTO DUMPFILE ...
    SELECT * FROM ... WHERE autoincrement_col IS NULL
    

    The last form is not cached because it is used as the ODBC workaround for obtaining the last insert ID value. See the Connector/ODBC section of Chapter 24, Connectors and APIs.

    Statements within transactions that use SERIALIZABLE isolation level also cannot be cached because they use LOCK IN SHARE MODE locking.

  • It uses TEMPORARY tables.

  • It does not use any tables.

  • It generates warnings.

  • The user has a column-level privilege for any of the involved tables.

8.10.3.2 Query Cache SELECT Options

Two query cache-related options may be specified in SELECT statements:

  • SQL_CACHE

    The query result is cached if it is cacheable and the value of the query_cache_type system variable is ON or DEMAND.

  • SQL_NO_CACHE

    The server does not use the query cache. It neither checks the query cache to see whether the result is already cached, nor does it cache the query result.

Examples:

SELECT SQL_CACHE id, name FROM customer;
SELECT SQL_NO_CACHE id, name FROM customer;

8.10.3.3 Query Cache Configuration

The have_query_cache server system variable indicates whether the query cache is available:

mysql> SHOW VARIABLES LIKE 'have_query_cache';
+------------------+-------+
| Variable_name    | Value |
+------------------+-------+
| have_query_cache | YES   |
+------------------+-------+

When using a standard MySQL binary, this value is always YES, even if query caching is disabled.

Several other system variables control query cache operation. These can be set in an option file or on the command line when starting mysqld. The query cache system variables all have names that begin with query_cache_. They are described briefly in Section 5.1.4, “Server System Variables”, with additional configuration information given here.

To set the size of the query cache, set the query_cache_size system variable. Setting it to 0 disables the query cache, as does setting query_cache_type=0. By default, the query cache is disabled. This is achieved using a default size of 1M, with a default for query_cache_type of 0.

To reduce overhead significantly, also start the server with query_cache_type=0 if you will not be using the query cache.

Note

When using the Windows Configuration Wizard to install or configure MySQL, the default value for query_cache_size will be configured automatically for you based on the different configuration types available. When using the Windows Configuration Wizard, the query cache may be enabled (that is, set to a nonzero value) due to the selected configuration. The query cache is also controlled by the setting of the query_cache_type variable. Check the values of these variables as set in your my.ini file after configuration has taken place.

When you set query_cache_size to a nonzero value, keep in mind that the query cache needs a minimum size of about 40KB to allocate its structures. (The exact size depends on system architecture.) If you set the value too small, you'll get a warning, as in this example:

mysql> SET GLOBAL query_cache_size = 40000;
Query OK, 0 rows affected, 1 warning (0.00 sec)

mysql> SHOW WARNINGS\G
*************************** 1. row ***************************
  Level: Warning
   Code: 1282
Message: Query cache failed to set size 39936;
         new query cache size is 0

mysql> SET GLOBAL query_cache_size = 41984;
Query OK, 0 rows affected (0.00 sec)

mysql> SHOW VARIABLES LIKE 'query_cache_size';
+------------------+-------+
| Variable_name    | Value |
+------------------+-------+
| query_cache_size | 41984 |
+------------------+-------+

For the query cache to actually be able to hold any query results, its size must be set larger:

mysql> SET GLOBAL query_cache_size = 1000000;
Query OK, 0 rows affected (0.04 sec)

mysql> SHOW VARIABLES LIKE 'query_cache_size';
+------------------+--------+
| Variable_name    | Value  |
+------------------+--------+
| query_cache_size | 999424 |
+------------------+--------+
1 row in set (0.00 sec)

The query_cache_size value is aligned to the nearest 1024 byte block. The value reported may therefore be different from the value that you assign.

If the query cache size is greater than 0, the query_cache_type variable influences how it works. This variable can be set to the following values:

  • A value of 0 or OFF prevents caching or retrieval of cached results.

  • A value of 1 or ON enables caching except of those statements that begin with SELECT SQL_NO_CACHE.

  • A value of 2 or DEMAND causes caching of only those statements that begin with SELECT SQL_CACHE.

If query_cache_size is 0, you should also set query_cache_type variable to 0. In this case, the server does not acquire the query cache mutex at all, which means that the query cache cannot be enabled at runtime and there is reduced overhead in query execution.

Setting the GLOBAL query_cache_type value determines query cache behavior for all clients that connect after the change is made. Individual clients can control cache behavior for their own connection by setting the SESSION query_cache_type value. For example, a client can disable use of the query cache for its own queries like this:

mysql> SET SESSION query_cache_type = OFF;

If you set query_cache_type at server startup (rather than at runtime with a SET statement), only the numeric values are permitted.

To control the maximum size of individual query results that can be cached, set the query_cache_limit system variable. The default value is 1MB.

Be careful not to set the size of the cache too large. Due to the need for threads to lock the cache during updates, you may see lock contention issues with a very large cache.

Note

You can set the maximum size that can be specified for the query cache at runtime with the SET statement by using the --maximum-query_cache_size=32M option on the command line or in the configuration file.

When a query is to be cached, its result (the data sent to the client) is stored in the query cache during result retrieval. Therefore the data usually is not handled in one big chunk. The query cache allocates blocks for storing this data on demand, so when one block is filled, a new block is allocated. Because memory allocation operation is costly (timewise), the query cache allocates blocks with a minimum size given by the query_cache_min_res_unit system variable. When a query is executed, the last result block is trimmed to the actual data size so that unused memory is freed. Depending on the types of queries your server executes, you might find it helpful to tune the value of query_cache_min_res_unit:

  • The default value of query_cache_min_res_unit is 4KB. This should be adequate for most cases.

  • If you have a lot of queries with small results, the default block size may lead to memory fragmentation, as indicated by a large number of free blocks. Fragmentation can force the query cache to prune (delete) queries from the cache due to lack of memory. In this case, decrease the value of query_cache_min_res_unit. The number of free blocks and queries removed due to pruning are given by the values of the Qcache_free_blocks and Qcache_lowmem_prunes status variables.

  • If most of your queries have large results (check the Qcache_total_blocks and Qcache_queries_in_cache status variables), you can increase performance by increasing query_cache_min_res_unit. However, be careful to not make it too large (see the previous item).

8.10.3.4 Query Cache Status and Maintenance

To check whether the query cache is present in your MySQL server, use the following statement:

mysql> SHOW VARIABLES LIKE 'have_query_cache';
+------------------+-------+
| Variable_name    | Value |
+------------------+-------+
| have_query_cache | YES   |
+------------------+-------+

You can defragment the query cache to better utilize its memory with the FLUSH QUERY CACHE statement. The statement does not remove any queries from the cache.

The RESET QUERY CACHE statement removes all query results from the query cache. The FLUSH TABLES statement also does this.

To monitor query cache performance, use SHOW STATUS to view the cache status variables:

mysql> SHOW STATUS LIKE 'Qcache%';
+-------------------------+--------+
| Variable_name           | Value  |
+-------------------------+--------+
| Qcache_free_blocks      | 36     |
| Qcache_free_memory      | 138488 |
| Qcache_hits             | 79570  |
| Qcache_inserts          | 27087  |
| Qcache_lowmem_prunes    | 3114   |
| Qcache_not_cached       | 22989  |
| Qcache_queries_in_cache | 415    |
| Qcache_total_blocks     | 912    |
+-------------------------+--------+

Descriptions of each of these variables are given in Section 5.1.6, “Server Status Variables”. Some uses for them are described here.

The total number of SELECT queries is given by this formula:

  Com_select
+ Qcache_hits
+ queries with errors found by parser

The Com_select value is given by this formula:

  Qcache_inserts
+ Qcache_not_cached
+ queries with errors found during the column-privileges check

The query cache uses variable-length blocks, so Qcache_total_blocks and Qcache_free_blocks may indicate query cache memory fragmentation. After FLUSH QUERY CACHE, only a single free block remains.

Every cached query requires a minimum of two blocks (one for the query text and one or more for the query results). Also, every table that is used by a query requires one block. However, if two or more queries use the same table, only one table block needs to be allocated.

The information provided by the Qcache_lowmem_prunes status variable can help you tune the query cache size. It counts the number of queries that have been removed from the cache to free up memory for caching new queries. The query cache uses a least recently used (LRU) strategy to decide which queries to remove from the cache. Tuning information is given in Section 8.10.3.3, “Query Cache Configuration”.

8.10.4 Caching of Prepared Statements and Stored Programs

For certain statements that a client might execute multiple times during a session, the server converts the statement to an internal structure and caches that structure to be used during execution. Caching enables the server to perform more efficiently because it avoids the overhead of reconverting the statement should it be needed again during the session. Conversion and caching occurs for these statements:

  • Prepared statements, both those processed at the SQL level (using the PREPARE statement) and those processed using the binary client/server protocol (using the mysql_stmt_prepare() C API function). The max_prepared_stmt_count system variable controls the total number of statements the server caches. (The sum of the number of prepared statements across all sessions.)

  • Stored programs (stored procedures and functions, triggers, and events). In this case, the server converts and caches the entire program body. The stored_program_cache system variable indicates the approximate number of stored programs the server caches per session.

The server maintains caches for prepared statements and stored programs on a per-session basis. Statements cached for one session are not accessible to other sessions. When a session ends, the server discards any statements cached for it.

When the server uses a cached internal statement structure, it must take care that the structure does not go out of date. Metadata changes can occur for an object used by the statement, causing a mismatch between the current object definition and the definition as represented in the internal statement structure. Metadata changes occur for DDL statements such as those that create, drop, alter, rename, or truncate tables, or that analyze, optimize, or repair tables. Table content changes (for example, with INSERT or UPDATE) do not change metadata, nor do SELECT statements.

Here is an illustration of the problem. Suppose that a client prepares this statement:

PREPARE s1 FROM 'SELECT * FROM t1';

The SELECT * expands in the internal structure to the list of columns in the table. If the set of columns in the table is modified with ALTER TABLE, the prepared statement goes out of date. If the server does not detect this change the next time the client executes s1, the prepared statement will return incorrect results.

To avoid problems caused by metadata changes to tables or views referred to by the prepared statement, the server detects these changes and automatically reprepares the statement when it is next executed. That is, the server reparses the statement and rebuilds the internal structure. Reparsing also occurs after referenced tables or views are flushed from the table definition cache, either implicitly to make room for new entries in the cache, or explicitly due to FLUSH TABLES.

Similarly, if changes occur to objects used by a stored program, the server reparses affected statements within the program.

The server also detects metadata changes for objects in expressions. These might be used in statements specific to stored programs, such as DECLARE CURSOR or flow-control statements such as IF, CASE, and RETURN.

To avoid reparsing entire stored programs, the server reparses affected statements or expressions within a program only as needed. Examples:

  • Suppose that metadata for a table or view is changed. Reparsing occurs for a SELECT * within the program that accesses the table or view, but not for a SELECT * that does not access the table or view.

  • When a statement is affected, the server reparses it only partially if possible. Consider this CASE statement:

    CASE case_expr
      WHEN when_expr1 ...
      WHEN when_expr2 ...
      WHEN when_expr3 ...
      ...
    END CASE
    

    If a metadata change affects only WHEN when_expr3, that expression is reparsed. case_expr and the other WHEN expressions are not reparsed.

Reparsing uses the default database and SQL mode that were in effect for the original conversion to internal form.

The server attempts reparsing up to three times. An error occurs if all attempts fail.

Reparsing is automatic, but to the extent that it occurs, diminishes prepared statement and stored program performance.

For prepared statements, the Com_stmt_reprepare status variable tracks the number of repreparations.

8.11 Optimizing Locking Operations

MySQL manages contention for table contents using locking:

  • Internal locking is performed within the MySQL server itself to manage contention for table contents by multiple threads. This type of locking is internal because it is performed entirely by the server and involves no other programs. See Section 8.11.1, “Internal Locking Methods”.

  • External locking occurs when the server and other programs lock MyISAM table files to coordinate among themselves which program can access the tables at which time. See Section 8.11.5, “External Locking”.

8.11.1 Internal Locking Methods

This section discusses internal locking; that is, locking performed within the MySQL server itself to manage contention for table contents by multiple sessions. This type of locking is internal because it is performed entirely by the server and involves no other programs. For locking performed on MySQL files by other programs, see Section 8.11.5, “External Locking”.

Row-Level Locking

MySQL uses row-level locking for InnoDB tables to support simultaneous write access by multiple sessions, making them suitable for multi-user, highly concurrent, and OLTP applications.

To avoid deadlocks when performing multiple concurrent write operations on a single InnoDB table, acquire necessary locks at the start of the transaction by issuing a SELECT ... FOR UPDATE statement for each group of rows expected to be modified, even if the DML statements come later in the transaction. If transactions modify or lock more than one table, issue the applicable statements in the same order within each transaction. Deadlocks affect performance rather than representing a serious error, because InnoDB automatically detects deadlock conditions and rolls back one of the affected transactions.

Advantages of row-level locking:

  • Fewer lock conflicts when different sessions access different rows.

  • Fewer changes for rollbacks.

  • Possible to lock a single row for a long time.

Table-Level Locking

MySQL uses table-level locking for MyISAM, MEMORY, and MERGE tables, permitting only one session to update those tables at a time. This locking level makes these storage engines more suitable for read-only, read-mostly, or single-user applications.

These storage engines avoid deadlocks by always requesting all needed locks at once at the beginning of a query and always locking the tables in the same order. The tradeoff is that this strategy reduces concurrency; other sessions that want to modify the table must wait until the current DML statement finishes.

Advantages of table-level locking:

  • Relatively little memory required (row locking requires memory per row or group of rows locked)

  • Fast when used on a large part of the table because only a single lock is involved.

  • Fast if you often do GROUP BY operations on a large part of the data or must scan the entire table frequently.

MySQL grants table write locks as follows:

  1. If there are no locks on the table, put a write lock on it.

  2. Otherwise, put the lock request in the write lock queue.

MySQL grants table read locks as follows:

  1. If there are no write locks on the table, put a read lock on it.

  2. Otherwise, put the lock request in the read lock queue.

Table updates are given higher priority than table retrievals. Therefore, when a lock is released, the lock is made available to the requests in the write lock queue and then to the requests in the read lock queue. This ensures that updates to a table are not starved even when there is heavy SELECT activity for the table. However, if there are many updates for a table, SELECT statements wait until there are no more updates.

For information on altering the priority of reads and writes, see Section 8.11.2, “Table Locking Issues”.

You can analyze the table lock contention on your system by checking the Table_locks_immediate and Table_locks_waited status variables, which indicate the number of times that requests for table locks could be granted immediately and the number that had to wait, respectively:

mysql> SHOW STATUS LIKE 'Table%';
+-----------------------+---------+
| Variable_name         | Value   |
+-----------------------+---------+
| Table_locks_immediate | 1151552 |
| Table_locks_waited    | 15324   |
+-----------------------+---------+

The Performance Schema lock tables also provide locking information. See Section 22.9.11, “Performance Schema Lock Tables”.

The MyISAM storage engine supports concurrent inserts to reduce contention between readers and writers for a given table: If a MyISAM table has no free blocks in the middle of the data file, rows are always inserted at the end of the data file. In this case, you can freely mix concurrent INSERT and SELECT statements for a MyISAM table without locks. That is, you can insert rows into a MyISAM table at the same time other clients are reading from it. Holes can result from rows having been deleted from or updated in the middle of the table. If there are holes, concurrent inserts are disabled but are enabled again automatically when all holes have been filled with new data. To control this behavior, use the concurrent_insert system variable. See Section 8.11.3, “Concurrent Inserts”.

If you acquire a table lock explicitly with LOCK TABLES, you can request a READ LOCAL lock rather than a READ lock to enable other sessions to perform concurrent inserts while you have the table locked.

To perform many INSERT and SELECT operations on a table t1 when concurrent inserts are not possible, you can insert rows into a temporary table temp_t1 and update the real table with the rows from the temporary table:

mysql> LOCK TABLES t1 WRITE, temp_t1 WRITE;
mysql> INSERT INTO t1 SELECT * FROM temp_t1;
mysql> DELETE FROM temp_t1;
mysql> UNLOCK TABLES;

Choosing the Type of Locking

Generally, table locks are superior to row-level locks in the following cases:

  • Most statements for the table are reads.

  • Statements for the table are a mix of reads and writes, where writes are updates or deletes for a single row that can be fetched with one key read:

    UPDATE tbl_name SET column=value WHERE unique_key_col=key_value;
    DELETE FROM tbl_name WHERE unique_key_col=key_value;
    
  • SELECT combined with concurrent INSERT statements, and very few UPDATE or DELETE statements.

  • Many scans or GROUP BY operations on the entire table without any writers.

With higher-level locks, you can more easily tune applications by supporting locks of different types, because the lock overhead is less than for row-level locks.

Options other than row-level locking:

  • Versioning (such as that used in MySQL for concurrent inserts) where it is possible to have one writer at the same time as many readers. This means that the database or table supports different views for the data depending on when access begins. Other common terms for this are time travel, copy on write, or copy on demand.

  • Copy on demand is in many cases superior to row-level locking. However, in the worst case, it can use much more memory than using normal locks.

  • Instead of using row-level locks, you can employ application-level locks, such as those provided by GET_LOCK() and RELEASE_LOCK() in MySQL. These are advisory locks, so they work only with applications that cooperate with each other. See Section 12.19, “Miscellaneous Functions”.

8.11.2 Table Locking Issues

InnoDB tables use row-level locking so that multiple sessions and applications can read from and write to the same table simultaneously, without making each other wait or producing inconsistent results. For this storage engine, avoid using the LOCK TABLES statement, because it does not offer any extra protection, but instead reduces concurrency. The automatic row-level locking makes these tables suitable for your busiest databases with your most important data, while also simplifying application logic since you do not need to lock and unlock tables. Consequently, the InnoDB storage engine is the default in MySQL.

MySQL uses table locking (instead of page, row, or column locking) for all storage engines except InnoDB. The locking operations themselves do not have much overhead. But because only one session can write to a table at any one time, for best performance with these other storage engines, use them primarily for tables that are queried often and rarely inserted into or updated.

Performance Considerations Favoring InnoDB

When choosing whether to create a table using InnoDB or a different storage engine, keep in mind the following disadvantages of table locking:

  • Table locking enables many sessions to read from a table at the same time, but if a session wants to write to a table, it must first get exclusive access, meaning it might have to wait for other sessions to finish with the table first. During the update, all other sessions that want to access this particular table must wait until the update is done.

  • Table locking causes problems when a session is waiting because the disk is full and free space needs to become available before the session can proceed. In this case, all sessions that want to access the problem table are also put in a waiting state until more disk space is made available.

  • A SELECT statement that takes a long time to run prevents other sessions from updating the table in the meantime, making the other sessions appear slow or unresponsive. While a session is waiting to get exclusive access to the table for updates, other sessions that issue SELECT statements will queue up behind it, reducing concurrency even for read-only sessions.

Workarounds for Locking Performance Issues

The following items describe some ways to avoid or reduce contention caused by table locking:

  • Consider switching the table to the InnoDB storage engine, either using CREATE TABLE ... ENGINE=INNODB during setup, or using ALTER TABLE ... ENGINE=INNODB for an existing table. See Chapter 14, The InnoDB Storage Engine for more details about this storage engine.

  • Optimize SELECT statements to run faster so that they lock tables for a shorter time. You might have to create some summary tables to do this.

  • Start mysqld with --low-priority-updates. For storage engines that use only table-level locking (such as MyISAM, MEMORY, and MERGE), this gives all statements that update (modify) a table lower priority than SELECT statements. In this case, the second SELECT statement in the preceding scenario would execute before the UPDATE statement, and would not wait for the first SELECT to finish.

  • To specify that all updates issued in a specific connection should be done with low priority, set the low_priority_updates server system variable equal to 1.

  • To give a specific INSERT, UPDATE, or DELETE statement lower priority, use the LOW_PRIORITY attribute.

  • To give a specific SELECT statement higher priority, use the HIGH_PRIORITY attribute. See Section 13.2.9, “SELECT Syntax”.

  • Start mysqld with a low value for the max_write_lock_count system variable to force MySQL to temporarily elevate the priority of all SELECT statements that are waiting for a table after a specific number of inserts to the table occur. This permits READ locks after a certain number of WRITE locks.

  • If you have problems with INSERT combined with SELECT, consider switching to MyISAM tables, which support concurrent SELECT and INSERT statements. (See Section 8.11.3, “Concurrent Inserts”.)

  • If you have problems with mixed SELECT and DELETE statements, the LIMIT option to DELETE may help. See Section 13.2.2, “DELETE Syntax”.

  • Using SQL_BUFFER_RESULT with SELECT statements can help to make the duration of table locks shorter. See Section 13.2.9, “SELECT Syntax”.

  • Splitting table contents into separate tables may help, by allowing queries to run against columns in one table, while updates are confined to columns in a different table.

  • You could change the locking code in mysys/thr_lock.c to use a single queue. In this case, write locks and read locks would have the same priority, which might help some applications.

8.11.3 Concurrent Inserts

The MyISAM storage engine supports concurrent inserts to reduce contention between readers and writers for a given table: If a MyISAM table has no holes in the data file (deleted rows in the middle), an INSERT statement can be executed to add rows to the end of the table at the same time that SELECT statements are reading rows from the table. If there are multiple INSERT statements, they are queued and performed in sequence, concurrently with the SELECT statements. The results of a concurrent INSERT may not be visible immediately.

The concurrent_insert system variable can be set to modify the concurrent-insert processing. By default, the variable is set to AUTO (or 1) and concurrent inserts are handled as just described. If concurrent_insert is set to NEVER (or 0), concurrent inserts are disabled. If the variable is set to ALWAYS (or 2), concurrent inserts at the end of the table are permitted even for tables that have deleted rows. See also the description of the concurrent_insert system variable.

If you are using the binary log, concurrent inserts are converted to normal inserts for CREATE ... SELECT or INSERT ... SELECT statements. This is done to ensure that you can re-create an exact copy of your tables by applying the log during a backup operation. See Section 5.2.4, “The Binary Log”. In addition, for those statements a read lock is placed on the selected-from table such that inserts into that table are blocked. The effect is that concurrent inserts for that table must wait as well.

With LOAD DATA INFILE, if you specify CONCURRENT with a MyISAM table that satisfies the condition for concurrent inserts (that is, it contains no free blocks in the middle), other sessions can retrieve data from the table while LOAD DATA is executing. Use of the CONCURRENT option affects the performance of LOAD DATA a bit, even if no other session is using the table at the same time.

If you specify HIGH_PRIORITY, it overrides the effect of the --low-priority-updates option if the server was started with that option. It also causes concurrent inserts not to be used.

For LOCK TABLE, the difference between READ LOCAL and READ is that READ LOCAL permits nonconflicting INSERT statements (concurrent inserts) to execute while the lock is held. However, this cannot be used if you are going to manipulate the database using processes external to the server while you hold the lock.

8.11.4 Metadata Locking

MySQL uses metadata locking to manage concurrent access to database objects and to ensure data consistency. Metadata locking applies not just to tables, but also to schemas, stored programs (procedures, functions, triggers, and scheduled events), and (as of MySQL 5.7.6) tablespaces.

Metadata locking does involve some overhead, which increases as query volume increases. Metadata contention increases the more that multiple queries attempt to access the same objects.

Metadata locking is not a replacement for the table definition cache, and its mutexes and locks differ from the LOCK_open mutex. The following discussion provides some information about how metadata locking works.

To ensure transaction serializability, the server must not permit one session to perform a data definition language (DDL) statement on a table that is used in an uncompleted explicitly or implicitly started transaction in another session. The server achieves this by acquiring metadata locks on tables used within a transaction and deferring release of those locks until the transaction ends. A metadata lock on a table prevents changes to the table's structure. This locking approach has the implication that a table that is being used by a transaction within one session cannot be used in DDL statements by other sessions until the transaction ends.

This principle applies not only to transactional tables, but also to nontransactional tables. Suppose that a session begins a transaction that uses transactional table t and nontransactional table nt as follows:

START TRANSACTION;
SELECT * FROM t;
SELECT * FROM nt;

The server holds metadata locks on both t and nt until the transaction ends. If another session attempts a DDL or write lock operation on either table, it blocks until metadata lock release at transaction end. For example, a second session blocks if it attempts any of these operations:

DROP TABLE t;
ALTER TABLE t ...;
DROP TABLE nt;
ALTER TABLE nt ...;
LOCK TABLE t ... WRITE;

As of MySQL 5.7.5, the same behavior applies for The LOCK TABLES ... READ. That is, explicitly or implicitly started transactions that update any table (transactional or nontransactional) will block and be blocked by LOCK TABLES ... READ for that table.

If the server acquires metadata locks for a statement that is syntactically valid but fails during execution, it does not release the locks early. Lock release is still deferred to the end of the transaction because the failed statement is written to the binary log and the locks protect log consistency.

In autocommit mode, each statement is in effect a complete transaction, so metadata locks acquired for the statement are held only to the end of the statement.

Metadata locks acquired during a PREPARE statement are released once the statement has been prepared, even if preparation occurs within a multiple-statement transaction.

8.11.5 External Locking

External locking is the use of file system locking to manage contention for MyISAM database tables by multiple processes. External locking is used in situations where a single process such as the MySQL server cannot be assumed to be the only process that requires access to tables. Here are some examples:

  • If you run multiple servers that use the same database directory (not recommended), each server must have external locking enabled.

  • If you use myisamchk to perform table maintenance operations on MyISAM tables, you must either ensure that the server is not running, or that the server has external locking enabled so that it locks table files as necessary to coordinate with myisamchk for access to the tables. The same is true for use of myisampack to pack MyISAM tables.

    If the server is run with external locking enabled, you can use myisamchk at any time for read operations such a checking tables. In this case, if the server tries to update a table that myisamchk is using, the server will wait for myisamchk to finish before it continues.

    If you use myisamchk for write operations such as repairing or optimizing tables, or if you use myisampack to pack tables, you must always ensure that the mysqld server is not using the table. If you do not stop mysqld, at least do a mysqladmin flush-tables before you run myisamchk. Your tables may become corrupted if the server and myisamchk access the tables simultaneously.

With external locking in effect, each process that requires access to a table acquires a file system lock for the table files before proceeding to access the table. If all necessary locks cannot be acquired, the process is blocked from accessing the table until the locks can be obtained (after the process that currently holds the locks releases them).

External locking affects server performance because the server must sometimes wait for other processes before it can access tables.

External locking is unnecessary if you run a single server to access a given data directory (which is the usual case) and if no other programs such as myisamchk need to modify tables while the server is running. If you only read tables with other programs, external locking is not required, although myisamchk might report warnings if the server changes tables while myisamchk is reading them.

With external locking disabled, to use myisamchk, you must either stop the server while myisamchk executes or else lock and flush the tables before running myisamchk. (See Section 8.12.1, “System Factors and Startup Parameter Tuning”.) To avoid this requirement, use the CHECK TABLE and REPAIR TABLE statements to check and repair MyISAM tables.

For mysqld, external locking is controlled by the value of the skip_external_locking system variable. When this variable is enabled, external locking is disabled, and vice versa. External locking is disabled by default.

Use of external locking can be controlled at server startup by using the --external-locking or --skip-external-locking option.

If you do use external locking option to enable updates to MyISAM tables from many MySQL processes, you must ensure that the following conditions are satisfied:

  • Do not use the query cache for queries that use tables that are updated by another process.

  • Do not start the server with the --delay-key-write=ALL option or use the DELAY_KEY_WRITE=1 table option for any shared tables. Otherwise, index corruption can occur.

The easiest way to satisfy these conditions is to always use --external-locking together with --delay-key-write=OFF and --query-cache-size=0. (This is not done by default because in many setups it is useful to have a mixture of the preceding options.)

8.12 Optimizing the MySQL Server

This section discusses optimization techniques for the database server, primarily dealing with system configuration rather than tuning SQL statements. The information in this section is appropriate for DBAs who want to ensure performance and scalability across the servers they manage; for developers constructing installation scripts that include setting up the database; and people running MySQL themselves for development, testing, and so on who want to maximize their own productivity.

8.12.1 System Factors and Startup Parameter Tuning

We start with system-level factors, because some of these decisions must be made very early to achieve large performance gains. In other cases, a quick look at this section may suffice. However, it is always nice to have a sense of how much can be gained by changing factors that apply at this level.

Before using MySQL in production, we advise you to test it on your intended platform.

Other tips:

  • If you have enough RAM, you could remove all swap devices. Some operating systems use a swap device in some contexts even if you have free memory.

  • Avoid external locking for MyISAM tables. The default is for external locking to be disabled. The --external-locking and --skip-external-locking options explicitly enable and disable external locking.

    Disabling external locking does not affect MySQL's functionality as long as you run only one server. Just remember to take down the server (or lock and flush the relevant tables) before you run myisamchk. On some systems it is mandatory to disable external locking because it does not work, anyway.

    The only case in which you cannot disable external locking is when you run multiple MySQL servers (not clients) on the same data, or if you run myisamchk to check (not repair) a table without telling the server to flush and lock the tables first. Note that using multiple MySQL servers to access the same data concurrently is generally not recommended, except when using MySQL Cluster.

    The LOCK TABLES and UNLOCK TABLES statements use internal locking, so you can use them even if external locking is disabled.

8.12.2 Tuning Server Parameters

You can determine the default buffer sizes used by the mysqld server using this command:

shell> mysqld --verbose --help

This command produces a list of all mysqld options and configurable system variables. The output includes the default variable values and looks something like this:

abort-slave-event-count                                      0
allow-suspicious-udfs                                        FALSE
archive                                                      ON
auto-increment-increment                                     1
auto-increment-offset                                        1
autocommit                                                   TRUE
automatic-sp-privileges                                      TRUE
avoid-temporal-upgrade                                       FALSE
back-log                                                     80
basedir                                                      /home/jon/bin/mysql-5.7/
...
tmpdir                                                       /tmp
transaction-alloc-block-size                                 8192
transaction-isolation                                        REPEATABLE-READ
transaction-prealloc-size                                    4096
transaction-read-only                                        FALSE
transaction-write-set-extraction                             OFF
updatable-views-with-limit                                   YES
validate-user-plugins                                        TRUE
verbose                                                      TRUE
wait-timeout

For a mysqld server that is currently running, you can see the current values of its system variables by connecting to it and issuing this statement:

mysql> SHOW VARIABLES;

You can also see some statistical and status indicators for a running server by issuing this statement:

mysql> SHOW STATUS;

System variable and status information also can be obtained using mysqladmin:

shell> mysqladmin variables
shell> mysqladmin extended-status

For a full description of all system and status variables, see Section 5.1.4, “Server System Variables”, and Section 5.1.6, “Server Status Variables”.

MySQL uses algorithms that are very scalable, so you can usually run with very little memory. However, normally better performance results from giving MySQL more memory.

When tuning a MySQL server, the two most important variables to configure are key_buffer_size and table_open_cache. You should first feel confident that you have these set appropriately before trying to change any other variables.

The following examples indicate some typical variable values for different runtime configurations.

  • If you have at least 1-2GB of memory and many tables and want maximum performance with a moderate number of clients, use something like this:

    shell> mysqld_safe --key_buffer_size=384M --table_open_cache=4000 \
               --sort_buffer_size=4M --read_buffer_size=1M &
    
  • If you have only 256MB of memory and only a few tables, but you still do a lot of sorting, you can use something like this:

    shell> mysqld_safe --key_buffer_size=64M --sort_buffer_size=1M
    

    If there are very many simultaneous connections, swapping problems may occur unless mysqld has been configured to use very little memory for each connection. mysqld performs better if you have enough memory for all connections.

  • With little memory and lots of connections, use something like this:

    shell> mysqld_safe --key_buffer_size=512K --sort_buffer_size=100K \
               --read_buffer_size=100K &
    

    Or even this:

    shell> mysqld_safe --key_buffer_size=512K --sort_buffer_size=16K \
               --table_open_cache=32 --read_buffer_size=8K \
               --net_buffer_length=1K &
    

If you are performing GROUP BY or ORDER BY operations on tables that are much larger than your available memory, increase the value of read_rnd_buffer_size to speed up the reading of rows following sorting operations.

You can make use of the example option files included with your MySQL distribution; see Section 5.1.2, “Server Configuration Defaults”.

If you specify an option on the command line for mysqld or mysqld_safe, it remains in effect only for that invocation of the server. To use the option every time the server runs, put it in an option file.

To see the effects of a parameter change, do something like this:

shell> mysqld --key_buffer_size=128M --verbose --help

The variable values are listed near the end of the output. Make sure that the --verbose and --help options are last. Otherwise, the effect of any options listed after them on the command line are not reflected in the output.

For information on optimizing the InnoDB storage engine performance, see Section 8.5, “Optimizing for InnoDB Tables”.

8.12.3 Optimizing Disk I/O

This section describes ways to configure storage devices when you can devote more and faster storage hardware to the database server. For information about optimizing an InnoDB configuration to improve I/O performance, see Section 8.5.8, “Optimizing InnoDB Disk I/O”.

  • Disk seeks are a huge performance bottleneck. This problem becomes more apparent when the amount of data starts to grow so large that effective caching becomes impossible. For large databases where you access data more or less randomly, you can be sure that you need at least one disk seek to read and a couple of disk seeks to write things. To minimize this problem, use disks with low seek times.

  • Increase the number of available disk spindles (and thereby reduce the seek overhead) by either symlinking files to different disks or striping the disks:

    • Using symbolic links

      This means that, for MyISAM tables, you symlink the index file and data files from their usual location in the data directory to another disk (that may also be striped). This makes both the seek and read times better, assuming that the disk is not used for other purposes as well. See Section 8.12.4, “Using Symbolic Links”.

      Symbolic links are not supported for use with InnoDB tables. However, you can create an InnoDB file-per-table tablespace in a location outside of the MySQL data directory using the DATA DIRECTORY = absolute_path_to_directory clause of the CREATE TABLE statement. For more information, see Section 14.5.5, “Creating a File-Per-Table Tablespace Outside the Data Directory”. General tablespaces can also be created in a location outside of the MySQL data directory. For more information, see Section 14.5.9, “InnoDB General Tablespaces”.

    • Striping

      Striping means that you have many disks and put the first block on the first disk, the second block on the second disk, and the N-th block on the (N MOD number_of_disks) disk, and so on. This means if your normal data size is less than the stripe size (or perfectly aligned), you get much better performance. Striping is very dependent on the operating system and the stripe size, so benchmark your application with different stripe sizes. See Section 8.13.2, “Using Your Own Benchmarks”.

      The speed difference for striping is very dependent on the parameters. Depending on how you set the striping parameters and number of disks, you may get differences measured in orders of magnitude. You have to choose to optimize for random or sequential access.

  • For reliability, you may want to use RAID 0+1 (striping plus mirroring), but in this case, you need 2 × N drives to hold N drives of data. This is probably the best option if you have the money for it. However, you may also have to invest in some volume-management software to handle it efficiently.

  • A good option is to vary the RAID level according to how critical a type of data is. For example, store semi-important data that can be regenerated on a RAID 0 disk, but store really important data such as host information and logs on a RAID 0+1 or RAID N disk. RAID N can be a problem if you have many writes, due to the time required to update the parity bits.

  • On Linux, you can get much better performance by using hdparm to configure your disk's interface. (Up to 100% under load is not uncommon.) The following hdparm options should be quite good for MySQL, and probably for many other applications:

    hdparm -m 16 -d 1
    

    Performance and reliability when using this command depend on your hardware, so we strongly suggest that you test your system thoroughly after using hdparm. Please consult the hdparm manual page for more information. If hdparm is not used wisely, file system corruption may result, so back up everything before experimenting!

  • You can also set the parameters for the file system that the database uses:

    If you do not need to know when files were last accessed (which is not really useful on a database server), you can mount your file systems with the -o noatime option. That skips updates to the last access time in inodes on the file system, which avoids some disk seeks.

    On many operating systems, you can set a file system to be updated asynchronously by mounting it with the -o async option. If your computer is reasonably stable, this should give you better performance without sacrificing too much reliability. (This flag is on by default on Linux.)

8.12.4 Using Symbolic Links

You can move databases or tables from the database directory to other locations and replace them with symbolic links to the new locations. You might want to do this, for example, to move a database to a file system with more free space or increase the speed of your system by spreading your tables to different disks.

For InnoDB tables, use the DATA DIRECTORY clause on the CREATE TABLE statement instead of symbolic links, as explained in Section 14.5.5, “Creating a File-Per-Table Tablespace Outside the Data Directory”. This new feature is a supported, cross-platform technique.

The recommended way to do this is to symlink entire database directories to a different disk. Symlink MyISAM tables only as a last resort.

To determine the location of your data directory, use this statement:

SHOW VARIABLES LIKE 'datadir';

8.12.4.1 Using Symbolic Links for Databases on Unix

On Unix, the way to symlink a database is first to create a directory on some disk where you have free space and then to create a soft link to it from the MySQL data directory.

shell> mkdir /dr1/databases/test
shell> ln -s /dr1/databases/test /path/to/datadir

MySQL does not support linking one directory to multiple databases. Replacing a database directory with a symbolic link works as long as you do not make a symbolic link between databases. Suppose that you have a database db1 under the MySQL data directory, and then make a symlink db2 that points to db1:

shell> cd /path/to/datadir
shell> ln -s db1 db2

The result is that, or any table tbl_a in db1, there also appears to be a table tbl_a in db2. If one client updates db1.tbl_a and another client updates db2.tbl_a, problems are likely to occur.

8.12.4.2 Using Symbolic Links for MyISAM Tables on Unix

Symlinks are fully supported only for MyISAM tables. For files used by tables for other storage engines, you may get strange problems if you try to use symbolic links. For InnoDB tables, use the alternative technique explained in Section 14.5.5, “Creating a File-Per-Table Tablespace Outside the Data Directory” instead.

Do not symlink tables on systems that do not have a fully operational realpath() call. (Linux and Solaris support realpath()). To determine whether your system supports symbolic links, check the value of the have_symlink system variable using this statement:

SHOW VARIABLES LIKE 'have_symlink';

The handling of symbolic links for MyISAM tables works as follows:

  • In the data directory, you always have the table format (.frm) file, the data (.MYD) file, and the index (.MYI) file. The data file and index file can be moved elsewhere and replaced in the data directory by symlinks. The format file cannot.

  • You can symlink the data file and the index file independently to different directories.

  • To instruct a running MySQL server to perform the symlinking, use the DATA DIRECTORY and INDEX DIRECTORY options to CREATE TABLE. See Section 13.1.18, “CREATE TABLE Syntax”. Alternatively, if mysqld is not running, symlinking can be accomplished manually using ln -s from the command line.

    Note

    The path used with either or both of the DATA DIRECTORY and INDEX DIRECTORY options may not include the MySQL data directory. (Bug #32167)

  • myisamchk does not replace a symlink with the data file or index file. It works directly on the file to which the symlink points. Any temporary files are created in the directory where the data file or index file is located. The same is true for the ALTER TABLE, OPTIMIZE TABLE, and REPAIR TABLE statements.

  • Note

    When you drop a table that is using symlinks, both the symlink and the file to which the symlink points are dropped. This is an extremely good reason not to run mysqld as the system root or permit system users to have write access to MySQL database directories.

  • If you rename a table with ALTER TABLE ... RENAME or RENAME TABLE and you do not move the table to another database, the symlinks in the database directory are renamed to the new names and the data file and index file are renamed accordingly.

  • If you use ALTER TABLE ... RENAME or RENAME TABLE to move a table to another database, the table is moved to the other database directory. If the table name changed, the symlinks in the new database directory are renamed to the new names and the data file and index file are renamed accordingly.

  • If you are not using symlinks, start mysqld with the --skip-symbolic-links option to ensure that no one can use mysqld to drop or rename a file outside of the data directory.

These table symlink operations are not supported:

  • ALTER TABLE ignores the DATA DIRECTORY and INDEX DIRECTORY table options.

  • As indicated previously, only the data and index files can be symbolic links. The .frm file must never be a symbolic link. Attempting to do this (for example, to make one table name a synonym for another) produces incorrect results. Suppose that you have a database db1 under the MySQL data directory, a table tbl1 in this database, and in the db1 directory you make a symlink tbl2 that points to tbl1:

    shell> cd /path/to/datadir/db1
    shell> ln -s tbl1.frm tbl2.frm
    shell> ln -s tbl1.MYD tbl2.MYD
    shell> ln -s tbl1.MYI tbl2.MYI
    

    Problems result if one thread reads db1.tbl1 and another thread updates db1.tbl2:

    • The query cache is fooled (it has no way of knowing that tbl1 has not been updated, so it returns outdated results).

    • ALTER statements on tbl2 fail.

8.12.4.3 Using Symbolic Links for Databases on Windows

On Windows, symbolic links can be used for database directories. This enables you to put a database directory at a different location (for example, on a different disk) by setting up a symbolic link to it. Use of database symlinks on Windows is similar to their use on Unix, although the procedure for setting up the link differs.

Suppose that you want to place the database directory for a database named mydb at D:\data\mydb. To do this, create a symbolic link in the MySQL data directory that points to D:\data\mydb. However, before creating the symbolic link, make sure that the D:\data\mydb directory exists by creating it if necessary. If you already have a database directory named mydb in the data directory, move it to D:\data. Otherwise, the symbolic link will be ineffective. To avoid problems, make sure that the server is not running when you move the database directory.

Windows Vista, Windows Server 2008, or newer have native symbolic link support, so you can create a symlink using the mklink command. This command requires administrative privileges.

  1. Change location into the data directory:

    C:\> cd \path\to\datadir
    
  2. In the data directory, create a symlink named mydb that points to the location of the database directory:

    C:\> mklink /d mydb D:\data\mydb
    

After this, all tables created in the database mydb are created in D:\data\mydb.

8.12.5 Optimizing Memory Use

8.12.5.1 How MySQL Uses Memory

MySQL allocates buffers and caches to improve performance of database operations. The default configuration is designed to allow a MySQL Server to start on a virtual machine with approximately 512MB of RAM. You can improve MySQL performance by increasing the values of certain cache and buffer-related system variables. You can also modify the default configuration to run MySQL on systems with limited memory.

The following list describes some of the ways that MySQL uses memory. Where applicable, relevant system variables are referenced. Some items are storage engine or feature specific.

  • The InnoDB buffer pool is a memory area that holds cached InnoDB data for tables, indexes, and other auxiliary buffers. For efficiency of high-volume read operations, the buffer pool is divided into pages that can potentially hold multiple rows. For efficiency of cache management, the buffer pool is implemented as a linked list of pages; data that is rarely used is aged out of the cache, using a variation of the LRU algorithm. For more information, see Section 8.10.1, “The InnoDB Buffer Pool”.

    The size of the buffer pool is important for system performance.

    • Typically, it is recommended that innodb_buffer_pool_size is configured to 50 to 75 percent of system memory.

    • InnoDB allocates memory for the entire buffer pool at server startup. Memory allocation is performed by malloc() operations. Buffer pool size is defined by the innodb_buffer_pool_size configuration option. As of MySQL 5.7, innodb_buffer_pool_size can be configured dynamically, while the server is running. For more information, see Section 14.4.3.7, “Configuring InnoDB Buffer Pool Size”.

    • On systems with a large amount of memory, you can improve concurrency by dividing the buffer pool into multiple buffer pool instances. The number of buffer pool instances is defined by innodb_buffer_pool_instances.

    • A buffer pool that is too small may cause excessive churning as pages are flushed from the buffer pool only to be required again a short time later.

    • A buffer pool that is too large may cause swapping due to competition for memory.

  • The MySQL Performance Schema is a feature for monitoring MySQL server execution at a low level. As of MySQL 5.7, the Performance Schema dynamically allocates memory incrementally, scaling its memory use to actual server load, instead of allocating required memory during server startup. Once memory is allocated, it is not freed until the server is restarted. For more information, see Section 22.14, “The Performance Schema Memory-Allocation Model”.

  • All threads share the MyISAM key buffer; its size is determined by the key_buffer_size variable. Other buffers used by the server are allocated as needed. See Section 8.12.2, “Tuning Server Parameters”.

    For each MyISAM table that is opened, the index file is opened once; the data file is opened once for each concurrently running thread. For each concurrent thread, a table structure, column structures for each column, and a buffer of size 3 * N are allocated (where N is the maximum row length, not counting BLOB columns). A BLOB column requires five to eight bytes plus the length of the BLOB data. The MyISAM storage engine maintains one extra row buffer for internal use.

  • Each thread that is used to manage client connections uses some thread-specific space. The following list indicates these and which variables control their size:

    The connection buffer and result buffer each begin with a size equal to net_buffer_length bytes, but are dynamically enlarged up to max_allowed_packet bytes as needed. The result buffer shrinks to net_buffer_length bytes after each SQL statement. While a statement is running, a copy of the current statement string is also allocated.

    Each connection thread uses memory for computing statement digests (see Section 22.7, “Performance Schema Statement Digests”): Before MySQL 5.7.4, 1024 bytes per session if the Performance Schema is compiled in with statement instrumentation. In 5.7.4 and 5.7.5, 1024 bytes per session. In 5.7.6 and higher, max_digest_length bytes per session.

  • All threads share the same base memory.

  • When a thread is no longer needed, the memory allocated to it is released and returned to the system unless the thread goes back into the thread cache. In that case, the memory remains allocated.

  • The myisam_use_mmap system variable can be set to 1 to enable memory-mapping for all MyISAM tables.

  • Each request that performs a sequential scan of a table allocates a read buffer (variable read_buffer_size).

  • When reading rows in an arbitrary sequence (for example, following a sort), a random-read buffer (variable read_rnd_buffer_size) may be allocated to avoid disk seeks.

  • All joins are executed in a single pass, and most joins can be done without even using a temporary table. Most temporary tables are memory-based hash tables. Temporary tables with a large row length (calculated as the sum of all column lengths) or that contain BLOB columns are stored on disk.

    If an internal in-memory temporary table becomes too large, MySQL handles this automatically by changing the table from in-memory to on-disk format, handled by the storage engine defined by internal_tmp_disk_storage_engine. You can increase the permissible temporary table size as described in Section 8.4.4, “Internal Temporary Table Use in MySQL”.

  • Most requests that perform a sort allocate a sort buffer and zero to two temporary files depending on the result set size. See Section B.5.3.5, “Where MySQL Stores Temporary Files”.

  • Almost all parsing and calculating is done in thread-local and reusable memory pools. No memory overhead is needed for small items, so the normal slow memory allocation and freeing is avoided. Memory is allocated only for unexpectedly large strings.

  • For each table having BLOB columns, a buffer is enlarged dynamically to read in larger BLOB values. If you scan a table, a buffer as large as the largest BLOB value is allocated.

  • MySQL requires memory and descriptors for the table cache. Handler structures for all in-use tables are saved in the table cache and managed as First In, First Out (FIFO). The initial table cache size is defined by the table_open_cache system variable; see Section 8.4.3.1, “How MySQL Opens and Closes Tables”.

    MySQL also requires memory for the table definition cache. The table_definition_cache system variable defines the number of table definitions (from .frm files) that can be stored in the table definition cache. If you use a large number of tables, you can create a large table definition cache to speed up the opening of tables. The table definition cache takes less space and does not use file descriptors, unlike the table cache.

  • A FLUSH TABLES statement or mysqladmin flush-tables command closes all tables that are not in use at once and marks all in-use tables to be closed when the currently executing thread finishes. This effectively frees most in-use memory. FLUSH TABLES does not return until all tables have been closed.

  • The server caches information in memory as a result of GRANT, CREATE USER, CREATE SERVER, and INSTALL PLUGIN statements. This memory is not released by the corresponding REVOKE, DROP USER, DROP SERVER, and UNINSTALL PLUGIN statements, so for a server that executes many instances of the statements that cause caching, there will be an increase in memory use. This cached memory can be freed with FLUSH PRIVILEGES.

ps and other system status programs may report that mysqld uses a lot of memory. This may be caused by thread stacks on different memory addresses. For example, the Solaris version of ps counts the unused memory between stacks as used memory. To verify this, check available swap with swap -s. We test mysqld with several memory-leakage detectors (both commercial and Open Source), so there should be no memory leaks.

8.12.5.2 Enabling Large Page Support

Some hardware/operating system architectures support memory pages greater than the default (usually 4KB). The actual implementation of this support depends on the underlying hardware and operating system. Applications that perform a lot of memory accesses may obtain performance improvements by using large pages due to reduced Translation Lookaside Buffer (TLB) misses.

In MySQL, large pages can be used by InnoDB, to allocate memory for its buffer pool and additional memory pool.

Standard use of large pages in MySQL attempts to use the largest size supported, up to 4MB. Under Solaris, a super large pages feature enables uses of pages up to 256MB. This feature is available for recent SPARC platforms. It can be enabled or disabled by using the --super-large-pages or --skip-super-large-pages option.

MySQL also supports the Linux implementation of large page support (which is called HugeTLB in Linux).

Before large pages can be used on Linux, the kernel must be enabled to support them and it is necessary to configure the HugeTLB memory pool. For reference, the HugeTBL API is documented in the Documentation/vm/hugetlbpage.txt file of your Linux sources.

The kernel for some recent systems such as Red Hat Enterprise Linux appear to have the large pages feature enabled by default. To check whether this is true for your kernel, use the following command and look for output lines containing huge:

shell> cat /proc/meminfo | grep -i huge
HugePages_Total:       0
HugePages_Free:        0
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:       4096 kB

The nonempty command output indicates that large page support is present, but the zero values indicate that no pages are configured for use.

If your kernel needs to be reconfigured to support large pages, consult the hugetlbpage.txt file for instructions.

Assuming that your Linux kernel has large page support enabled, configure it for use by MySQL using the following commands. Normally, you put these in an rc file or equivalent startup file that is executed during the system boot sequence, so that the commands execute each time the system starts. The commands should execute early in the boot sequence, before the MySQL server starts. Be sure to change the allocation numbers and the group number as appropriate for your system.

# Set the number of pages to be used.
# Each page is normally 2MB, so a value of 20 = 40MB.
# This command actually allocates memory, so this much
# memory must be available.
echo 20 > /proc/sys/vm/nr_hugepages

# Set the group number that is permitted to access this
# memory (102 in this case). The mysql user must be a
# member of this group.
echo 102 > /proc/sys/vm/hugetlb_shm_group

# Increase the amount of shmem permitted per segment
# (12G in this case).
echo 1560281088 > /proc/sys/kernel/shmmax

# Increase total amount of shared memory.  The value
# is the number of pages. At 4KB/page, 4194304 = 16GB.
echo 4194304 > /proc/sys/kernel/shmall

For MySQL usage, you normally want the value of shmmax to be close to the value of shmall.

To verify the large page configuration, check /proc/meminfo again as described previously. Now you should see some nonzero values:

shell> cat /proc/meminfo | grep -i huge
HugePages_Total:      20
HugePages_Free:       20
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:       4096 kB

The final step to make use of the hugetlb_shm_group is to give the mysql user an unlimited value for the memlock limit. This can by done either by editing /etc/security/limits.conf or by adding the following command to your mysqld_safe script:

ulimit -l unlimited

Adding the ulimit command to mysqld_safe causes the root user to set the memlock limit to unlimited before switching to the mysql user. (This assumes that mysqld_safe is started by root.)

Large page support in MySQL is disabled by default. To enable it, start the server with the --large-pages option. For example, you can use the following lines in your server's my.cnf file:

[mysqld]
large-pages

With this option, InnoDB uses large pages automatically for its buffer pool and additional memory pool. If InnoDB cannot do this, it falls back to use of traditional memory and writes a warning to the error log: Warning: Using conventional memory pool

To verify that large pages are being used, check /proc/meminfo again:

shell> cat /proc/meminfo | grep -i huge
HugePages_Total:      20
HugePages_Free:       20
HugePages_Rsvd:        2
HugePages_Surp:        0
Hugepagesize:       4096 kB

8.12.6 Optimizing Network Use

8.12.6.1 How MySQL Uses Threads for Client Connections

Connection manager threads handle client connection requests on the network interfaces that the server listens to. On all platforms, one manager thread handles TCP/IP connection requests. On Unix, this manager thread also handles Unix socket file connection requests. On Windows, a manager thread handles shared-memory connection requests, and another handles named-pipe connection requests. The server does not create threads to handle interfaces that it does not listen to. For example, a Windows server that does not have support for named-pipe connections enabled does not create a thread to handle them.

Connection manager threads associate each client connection with a thread dedicated to it that handles authentication and request processing for that connection. Manager threads create a new thread when necessary but try to avoid doing so by consulting the thread cache first to see whether it contains a thread that can be used for the connection. When a connection ends, its thread is returned to the thread cache if the cache is not full.

In this connection thread model, there are as many threads as there are clients currently connected, which has some disadvantages when server workload must scale to handle large numbers of connections. For example, thread creation and disposal becomes expensive. Also, each thread requires server and kernel resources, such as stack space. To accommodate a large number of simultaneous connections, the stack size per thread must be kept small, leading to a situation where it is either too small or the server consumes large amounts of memory. Exhaustion of other resources can occur as well, and scheduling overhead can become significant.

To control and monitor how the server manages threads that handle client connections, several system and status variables are relevant. (See Section 5.1.4, “Server System Variables”, and Section 5.1.6, “Server Status Variables”.)

The thread cache has a size determined by the thread_cache_size system variable. The default value is 0 (no caching), which causes a thread to be set up for each new connection and disposed of when the connection terminates. Set thread_cache_size to N to enable N inactive connection threads to be cached. thread_cache_size can be set at server startup or changed while the server runs. A connection thread becomes inactive when the client connection with which it was associated terminates.

To monitor the number of threads in the cache and how many threads have been created because a thread could not be taken from the cache, monitor the Threads_cached and Threads_created status variables.

You can set max_connections at server startup or at runtime to control the maximum number of clients that can connect simultaneously.

When the thread stack is too small, this limits the complexity of the SQL statements which the server can handle, the recursion depth of stored procedures, and other memory-consuming actions. To set a stack size of N bytes for each thread, start the server with --thread_stack=N.

8.12.6.2 DNS Lookup Optimization and the Host Cache

The MySQL server maintains a host cache in memory that contains information about clients: IP address, host name, and error information. The server uses this cache for nonlocal TCP connections. It does not use the cache for TCP connections established using a loopback interface address (127.0.0.1 or ::1), or for connections established using a Unix socket file, named pipe, or shared memory.

For each new client connection, the server uses the client IP address to check whether the client host name is in the host cache. If not, the server attempts to resolve the host name. First, it resolves the IP address to a host name and resolves that host name back to an IP address. Then it compares the result to the original IP address to ensure that they are the same. The server stores information about the result of this operation in the host cache. If the cache is full, the least recently used entry is discarded.

The host_cache Performance Schema table exposes the contents of the host cache so that it can be examined using SELECT statements. This may help you diagnose the causes of connection problems. See Section 22.9.15.1, “The host_cache Table”.

The server handles entries in the host cache like this:

  1. When the first TCP client connection reaches the server from a given IP address, a new entry is created to record the client IP, host name, and client lookup validation flag. Initially, the host name is set to NULL and the flag is false. This entry is also used for subsequent client connections from the same originating IP.

  2. If the validation flag for the client IP entry is false, the server attempts an IP-to-host name DNS resolution. If that is successful, the host name is updated with the resolved host name and the validation flag is set to true. If resolution is unsuccessful, the action taken depends on whether the error is permanent or transient. For permanent failures, the host name remains NULL and the validation flag is set to true. For transient failures, the host name and validation flag remain unchanged. (Another DNS resolution attempt occurs the next time a client connects from this IP.)

  3. If an error occurs while processing an incoming client connection from a given IP address, the server updates the corresponding error counters in the entry for that IP. For a description of the errors recorded, see Section 22.9.15.1, “The host_cache Table”.

The server performs host name resolution using the thread-safe gethostbyaddr_r() and gethostbyname_r() calls if the operating system supports them. Otherwise, the thread performing the lookup locks a mutex and calls gethostbyaddr() and gethostbyname() instead. In this case, no other thread can resolve host names that are not in the host cache until the thread holding the mutex lock releases it.

The server uses the host cache for several purposes:

  • By caching the results of IP-to-host name lookups, the server avoids doing a DNS lookup for each client connection. Instead, for a given host, it needs to perform a lookup only for the first connection from that host.

  • The cache contains information about errors that occur during the connection process. Some errors are considered blocking. If too many of these occur successively from a given host without a successful connection, the server blocks further connections from that host. The max_connect_errors system variable determines the number of permitted errors before blocking occurs. See Section B.5.2.6, “Host 'host_name' is blocked”.

To unblock blocked hosts, flush the host cache by issuing a FLUSH HOSTS statement or executing a mysqladmin flush-hosts command.

It is possible for a blocked host to become unblocked even without FLUSH HOSTS if activity from other hosts has occurred since the last connection attempt from the blocked host. This can occur because the server discards the least recently used cache entry to make room for a new entry if the cache is full when a connection arrives from a client IP not in the cache. If the discarded entry is for a blocked host, that host becomes unblocked.

The host cache is enabled by default. To disable it, set the host_cache_size system variable to 0, either at server startup or at runtime.

To disable DNS host name lookups, start the server with the --skip-name-resolve option. In this case, the server uses only IP addresses and not host names to match connecting hosts to rows in the MySQL grant tables. Only accounts specified in those tables using IP addresses can be used. (Be sure that an account exists that specifies an IP address or you may not be able to connect.)

If you have a very slow DNS and many hosts, you might be able to improve performance either by disabling DNS lookups with --skip-name-resolve or by increasing the value of host_cache_size to make the host cache larger.

To disallow TCP/IP connections entirely, start the server with the --skip-networking option.

Some connection errors are not associated with TCP connections, occur very early in the connection process (even before an IP address is known), or are not specific to any particular IP address (such as out-of-memory conditions). For information about these errors, check the Connection_errors_xxx status variables (see Section 5.1.6, “Server Status Variables”).

8.12.7 The Thread Pool Plugin

Note

MySQL Thread Pool is an extension included in MySQL Enterprise Edition, a commercial product. To learn more about commercial products, http://www.mysql.com/products/.

As of MySQL 5.7.9, MySQL Enterprise Edition includes MySQL Thread Pool, implemented using a server plugin. The default thread-handling model in MySQL Server executes statements using one thread per client connection. As more clients connect to the server and execute statements, overall performance degrades. The thread pool plugin provides an alternative thread-handling model designed to reduce overhead and improve performance. The plugin implements a thread pool that increases server performance by efficiently managing statement execution threads for large numbers of client connections.

The thread pool addresses several problems of the one thread per connection model:

  • Too many thread stacks make CPU caches almost useless in highly parallel execution workloads. The thread pool promotes thread stack reuse to minimize the CPU cache footprint.

  • With too many threads executing in parallel, context switching overhead is high. This also presents a challenging task to the operating system scheduler. The thread pool controls the number of active threads to keep the parallelism within the MySQL server at a level that it can handle and that is appropriate for the server host on which MySQL is executing.

  • Too many transactions executing in parallel increases resource contention. In InnoDB, this increases the time spent holding central mutexes. The thread pool controls when transactions start to ensure that not too many execute in parallel.

The thread pool plugin is included only in MySQL Enterprise Edition. It is not included in MySQL community distributions.

On Windows, the thread pool plugin requires Windows Vista or newer. On Linux, the plugin requires kernel 2.6.9 or newer.

Additional Resources

Section A.14, “MySQL 5.7 FAQ: MySQL Enterprise Thread Pool”

8.12.7.1 Thread Pool Components and Installation

The thread pool feature comprises these components:

To be usable by the server, the thread pool library object file must be located in the MySQL plugin directory (the directory named by the plugin_dir system variable). To enable thread pool capability, load the plugins to be used by starting the server with the --plugin-load option. For example, if you name just the plugin object file, the server loads all plugins that it contains (that is, the thread pool plugin and all the INFORMATION_SCHEMA tables). To do this, put these lines in your my.cnf file:

[mysqld]
plugin-load=thread_pool.so

That is equivalent to loading all thread pool plugins by naming them individually:

[mysqld]
plugin-load=thread_pool.so
plugin-load=thread_pool=thread_pool.so;tp_thread_state=thread_pool.so;tp_thread_group_state=thread_pool.so;tp_thread_group_stats=thread_pool.so

If object files have a suffix different from .so on your system, substitute the correct suffix (for example, .dll on Windows).

If necessary, set the value of the plugin_dir system variable to tell the server the location of the plugin directory.

If desired, you can load individual plugins from the library file. To load the thread pool plugin but not the INFORMATION_SCHEMA tables, use an option like this:

[mysqld]
plugin-load=thread_pool=thread_pool.so

To load the thread pool plugin and only the TP_THREAD_STATE INFORMATION_SCHEMA table, use an option like this:

[mysqld]
plugin-load=thread_pool=thread_pool.so;TP_THREAD_STATE=thread_pool.so

However, if you do not load all the INFORMATION_SCHEMA tables, some or all MySQL Enterprise Monitor thread pool graphs will be empty.

To verify plugin installation, examine the INFORMATION_SCHEMA.PLUGINS table or use the SHOW PLUGINS statement. See Section 5.1.8.2, “Obtaining Server Plugin Information”.

If the server loads the thread plugin successfully, it sets the thread_handling system variable to dynamically-loaded. If the plugin fails to load, the server writes a message to the error log.

8.12.7.2 Thread Pool Operation

The thread pool consists of a number of thread groups, each of which manages a set of client connections. As connections are established, the thread pool assigns them to thread groups in round-robin fashion.

The number of thread groups is configurable using the thread_pool_size system variable. The default number of groups is 16. For guidelines on setting this variable, see Section 8.12.7.3, “Thread Pool Tuning”.

The maximum number of threads per group is 4096 (or 4095 on some systems where one thread is used internally).

The thread pool separates connections and threads, so there is no fixed relationship between connections and the threads that execute statements received from those connections. This differs from the default thread-handling model that associates one thread with one connection such that the thread executes all statements from the connection.

The thread pool tries to ensure a maximum of one thread executing in each group at any time, but sometimes permits more threads to execute temporarily for best performance. The algorithm works in the following manner:

  • Each thread group has a listener thread that listens for incoming statements from the connections assigned to the group. When a statement arrives, the thread group either begins executing it immediately or queues it for later execution:

    • Immediate execution occurs if the statement is the only one received and no statements are queued or currently executing.

    • Queuing occurs if the statement cannot begin executing immediately.

  • If immediate execution occurs, execution is performed by the listener thread. (This means that temporarily no thread in the group is listening.) If the statement finishes quickly, the executing thread returns to listening for statements. Otherwise, the thread pool considers the statement stalled and starts another thread as a listener thread (creating it if necessary). To ensure that no thread group becomes blocked by stalled statements, the thread pool has a background thread that regularly monitors thread group states.

    By using the listening thread to execute a statement that can begin immediately, there is no need to create an additional thread if the statement finishes quickly. This ensures the most efficient execution possible in the case of a low number of concurrent threads.

    When the thread pool plugin starts, it creates one thread per group (the listener thread), plus the background thread. Additional threads are created as necessary to execute statements.

  • The value of the thread_pool_stall_limit system variable determines the meaning of finishes quickly in the previous item. The default time before threads are considered stalled is 60ms but can be set to a maximum of 6s. This parameter is configurable to enable you to strike a balance appropriate for the server work load. Short wait values permit threads to start more quickly. Short values are also better for avoiding deadlock situations. Long wait values are useful for workloads that include long-running statements, to avoid starting too many new statements while the current ones execute.

  • The thread pool focuses on limiting the number of concurrent short-running statements. Before an executing statement reaches the stall time, it prevents other statements from beginning to execute. If the statement executes past the stall time, it is permitted to continue but no longer prevents other statements from starting. In this way, the thread pool tries to ensure that in each thread group there is never more than one short-running statement, although there might be multiple long-running statements. It is undesirable to let long-running statements prevent other statements from executing because there is no limit on the amount of waiting that might be necessary. For example, on a replication master, a thread that is sending binary log events to a slave effectively runs forever.

  • A statement becomes blocked if it encounters a disk I/O operation or a user level lock (row lock or table lock). The block would cause the thread group to become unused, so there are callbacks to the thread pool to ensure that the thread pool can immediately start a new thread in this group to execute another statement. When a blocked thread returns, the thread pool permits it to restart immediately.

  • There are two queues, a high-priority queue and a low-priority queue. The first statement in a transaction goes to the low-priority queue. Any following statements for the transaction go to the high-priority queue if the transaction is ongoing (statements for it have begun executing), or to the low-priority queue otherwise. Queue assignment can be affected by enabling the thread_pool_high_priority_connection system variable, which causes all queued statements for a session to go into the high-priority queue.

    Statements for a nontransactional storage engine, or a transactional engine if autocommit is enabled, are treated as low-priority statements because in this case each statement is a transaction. Thus, given a mix of statements for InnoDB and MyISAM tables, the thread pool prioritizes those for InnoDB over those for MyISAM unless autocommit is enabled. With autocommit enabled, all statements will be low priority.

  • When the thread group selects a queued statement for execution, it first looks in the high-priority queue, then in the low-priority queue. If a statement is found, it is removed from its queue and begins to execute.

  • If a statement stays in the low-priority queue too long, the thread pool moves to the high-priority queue. The value of the thread_pool_prio_kickup_timer system variable controls the time before movement. For each thread group, a maximum of one statement per 10ms or 100 per second will be moved from the low-priority queue to the high-priority queue.

  • The thread pool reuses the most active threads to obtain a much better use of CPU caches. This is a small adjustment that has a great impact on performance.

  • While a thread executes a statement from a user connection, Performance Schema instrumentation accounts thread activity to the user connection. Otherwise, Performance Schema accounts activity to the thread pool.

Here are examples of conditions under which a thread group might have multiple threads started to execute statements:

  • One thread begins executing a statement, but runs long enough to be considered stalled. The thread group permits another thread to begin executing another statement even through the first thread is still executing.

  • One thread begins executing a statement, then becomes blocked and reports this back to the thread pool. The thread group permits another thread to begin executing another statement.

  • One thread begins executing a statement, becomes blocked, but does not report back that it is blocked because the block does not occur in code that has been instrumented with thread pool callbacks. In this case, the thread appears to the thread group to be still running. If the block lasts long enough for the statement to be considered stalled, the group permits another thread to begin executing another statement.

The thread pool is designed to be scalable across an increasing number of connections. It is also designed to avoid deadlocks that can arise from limiting the number of actively executing statements. It is important that threads that do not report back to the thread pool do not prevent other statements from executing and thus cause the thread pool to become deadlocked. Examples of such statements follow:

  • Long-running statements. These would lead to all resources used by only a few statements and they could prevent all others from accessing the server.

  • Binary log dump threads that read the binary log and send it to slaves. This is a kind of long-running statement that runs for a very long time, and that should not prevent other statements from executing.

  • Statements blocked on a row lock, table lock, sleep, or any other blocking activity that has not been reported back to the thread pool by MySQL Server or a storage engine.

In each case, to prevent deadlock, the statement is moved to the stalled category when it does not complete quickly, so that the thread group can permit another statement to begin executing. With this design, when a thread executes or becomes blocked for an extended time, the thread pool moves the thread to the stalled category and for the rest of the statement's execution, it does not prevent other statements from executing.

The maximum number of threads that can occur is the sum of max_connections and thread_pool_size. This can happen in a situation where all connections are in execution mode and an extra thread is created per group to listen for more statements. This is not necessarily a state that happens often, but it is theoretically possible.

8.12.7.3 Thread Pool Tuning

This section provides guidelines on setting thread pool system variables for best performance, measured using a metric such as transactions per second.

thread_pool_size is the most important parameter controlling thread pool performance. It can be set only at server startup. Our experience in testing the thread pool indicates the following:

  • If the primary storage engine is InnoDB, the optimal thread_pool_size setting is likely to be between 16 and 36, with the most common optimal values tending to be from 24 to 36. We have not seen any situation where the setting has been optimal beyond 36. There may be special cases where a value smaller than 16 is optimal.

    For workloads such as DBT2 and Sysbench, the optimum for InnoDB seems to be usually around 36. For very write-intensive workloads, the optimal setting can sometimes be lower.

  • If the primary storage engine is MyISAM, the thread_pool_size setting should be fairly low. We tend to get optimal performance for values from 4 to 8. Higher values tend to have a slightly negative but not dramatic impact on performance.

Another system variable, thread_pool_stall_limit, is important for handling of blocked and long-running statements. If all calls that block the MySQL Server are reported to the thread pool, it would always know when execution threads are blocked. However, this may not always be true. For example, blocks could occur in code that has not been instrumented with thread pool callbacks. For such cases, the thread pool must be able to identify threads that appear to be blocked. This is done by means of a timeout, the length of which can be tuned using the thread_pool_stall_limit system variable. This parameter ensures that the server does not become completely blocked. The value of thread_pool_stall_limit has an upper limit of 6 seconds to prevent the risk of a deadlocked server.

thread_pool_stall_limit also enables the thread pool to handle long-running statements. If a long-running statement was permitted to block a thread group, all other connections assigned to the group would be blocked and unable to start execution until the long-running statement completed. In the worst case, this could take hours or even days.

The value of thread_pool_stall_limit should be chosen such that statements that execute longer than its value are considered stalled. Stalled statements generate a lot of extra overhead since they involve extra context switches and in some cases even extra thread creations. On the other hand, setting the thread_pool_stall_limit parameter too high means that long-running statements will block a number of short-running statements for longer than necessary. Short wait values permit threads to start more quickly. Short values are also better for avoiding deadlock situations. Long wait values are useful for workloads that include long-running statements, to avoid starting too many new statements while the current ones execute.

Suppose a server executes a workload where 99.9% of the statements complete within 100ms even when the server is loaded, and the remaining statements take between 100ms and 2 hours fairly evenly spread. In this case, it would make sense to set thread_pool_stall_limit to 10 (meaning 100ms). The default value of 60ms is okay for servers that primarily execute very simple statements.

The thread_pool_stall_limit parameter can be changed at runtime to enable you to strike a balance appropriate for the server work load. Assuming that the TP_THREAD_GROUP_STATS table is enabled, you can use the following query to determine the fraction of executed statements that stalled:

SELECT SUM(STALLED_QUERIES_EXECUTED) / SUM(QUERIES_EXECUTED)
FROM information_schema.TP_THREAD_GROUP_STATS;

This number should be as low as possible. To decrease the likelihood of statements stalling, increase the value of thread_pool_stall_limit.

When a statement arrives, what is the maximum time it can be delayed before it actually starts executing? Suppose that the following conditions apply:

In the worst case, the 10 high-priority statements represent 10 transactions that continue executing for a long time. Thus, in the worst case, no statements will be moved to the high-priority queue because it will always already contain statements awaiting execution. After 10 seconds, the new statement is eligible to be moved to the high-priority queue. However, before it can be moved, all the statements before it must be moved as well. This could take another 2 seconds because a maximum of 100 statements per second are moved to the high-priority queue. Now when the statement reaches the high-priority queue, there could potentially be many long-running statements ahead of it. In the worst case, every one of those will become stalled and it will take 1 second for each statement before the next statement is retrieved from the high-priority queue. Thus, in this scenario, it will take 222 seconds before the new statement starts executing.

This example shows a worst case for an application. How to handle it depends on the application. If the application has high requirements for the response time, it should most likely throttle users at a higher level itself. Otherwise, it can use the thread pool configuration parameters to set some kind of a maximum waiting time.

8.13 Measuring Performance (Benchmarking)

To measure performance, consider the following factors:

  • Whether you are measuring the speed of a single operation on a quiet system, or how a set of operations (a workload) works over a period of time. With simple tests, you usually test how changing one aspect (a configuration setting, the set of indexes on a table, the SQL clauses in a query) affects performance. Benchmarks are typically long-running and elaborate performance tests, where the results could dictate high-level choices such as hardware and storage configuration, or how soon to upgrade to a new MySQL version.

  • For benchmarking, sometimes you must simulate a heavy database workload to get an accurate picture.

  • Performance can vary depending on so many different factors that a difference of a few percentage points might not be a decisive victory. The results might shift the opposite way when you test in a different environment.

  • Certain MySQL features help or do not help performance depending on the workload. For completeness, always test performance with those features turned on and turned off. The two most important features to try with each workload are the MySQL query cache, and the adaptive hash index for InnoDB tables.

This section progresses from simple and direct measurement techniques that a single developer can do, to more complicated ones that require additional expertise to perform and interpret the results.

8.13.1 Measuring the Speed of Expressions and Functions

To measure the speed of a specific MySQL expression or function, invoke the BENCHMARK() function using the mysql client program. Its syntax is BENCHMARK(loop_count,expression). The return value is always zero, but mysql prints a line displaying approximately how long the statement took to execute. For example:

mysql> SELECT BENCHMARK(1000000,1+1);
+------------------------+
| BENCHMARK(1000000,1+1) |
+------------------------+
|                      0 |
+------------------------+
1 row in set (0.32 sec)

This result was obtained on a Pentium II 400MHz system. It shows that MySQL can execute 1,000,000 simple addition expressions in 0.32 seconds on that system.

The built-in MySQL functions are typically highly optimized, but there may be some exceptions. BENCHMARK() is an excellent tool for finding out if some function is a problem for your queries.

8.13.2 Using Your Own Benchmarks

Benchmark your application and database to find out where the bottlenecks are. After fixing one bottleneck (or by replacing it with a dummy module), you can proceed to identify the next bottleneck. Even if the overall performance for your application currently is acceptable, you should at least make a plan for each bottleneck and decide how to solve it if someday you really need the extra performance.

A free benchmark suite is the Open Source Database Benchmark, available at http://osdb.sourceforge.net/.

It is very common for a problem to occur only when the system is very heavily loaded. We have had many customers who contact us when they have a (tested) system in production and have encountered load problems. In most cases, performance problems turn out to be due to issues of basic database design (for example, table scans are not good under high load) or problems with the operating system or libraries. Most of the time, these problems would be much easier to fix if the systems were not already in production.

To avoid problems like this, benchmark your whole application under the worst possible load:

These programs or packages can bring a system to its knees, so be sure to use them only on your development systems.

8.13.3 Measuring Performance with performance_schema

You can query the tables in the performance_schema database to see real-time information about the performance characteristics of your server and the applications it is running. See Chapter 22, MySQL Performance Schema for details.

8.14 Examining Thread Information

When you are attempting to ascertain what your MySQL server is doing, it can be helpful to examine the process list, which is the set of threads currently executing within the server. Process list information is available from these sources:

Access to threads does not require a mutex and has minimal impact on server performance. INFORMATION_SCHEMA.PROCESSLIST and SHOW PROCESSLIST have negative performance consequences because they require a mutex. threads also shows information about background threads, which INFORMATION_SCHEMA.PROCESSLIST and SHOW PROCESSLIST do not. This means that threads can be used to monitor activity the other thread information sources cannot.

You can always view information about your own threads. To view information about threads being executed for other accounts, you must have the PROCESS privilege.

Each process list entry contains several pieces of information:

  • Id is the connection identifier for the client associated with the thread.

  • User and Host indicate the account associated with the thread.

  • db is the default database for the thread, or NULL if none is selected.

  • Command