Chapter 7. Optimization

Table of Contents

7.1. Optimization Overview
7.1.1. MySQL Design Limitations and Tradeoffs
7.1.2. Designing Applications for Portability
7.1.3. What We Have Used MySQL For
7.1.4. The MySQL Benchmark Suite
7.1.5. Using Your Own Benchmarks
7.2. Optimizing SELECT and Other Statements
7.2.1. Optimizing Queries with EXPLAIN
7.2.2. Estimating Query Performance
7.2.3. Speed of SELECT Queries
7.2.4. WHERE Clause Optimization
7.2.5. Range Optimization
7.2.6. Index Merge Optimization
7.2.7. Condition Pushdown Optimization
7.2.8. IS NULL Optimization
7.2.9. LEFT JOIN and RIGHT JOIN Optimization
7.2.10. Nested-Loop Join Algorithms
7.2.11. Nested Join Optimization
7.2.12. Outer Join Simplification
7.2.13. ORDER BY Optimization
7.2.14. GROUP BY Optimization
7.2.15. DISTINCT Optimization
7.2.16. Optimizing IN/=ANY Subqueries
7.2.17. LIMIT Optimization
7.2.18. How to Avoid Table Scans
7.2.19. Speed of INSERT Statements
7.2.20. Speed of UPDATE Statements
7.2.21. Speed of DELETE Statements
7.2.22. Other Optimization Tips
7.3. Locking Issues
7.3.1. Internal Locking Methods
7.3.2. Table Locking Issues
7.3.3. Concurrent Inserts
7.3.4. External Locking
7.4. Optimizing Database Structure
7.4.1. Make Your Data as Small as Possible
7.4.2. Column Indexes
7.4.3. Multiple-Column Indexes
7.4.4. How MySQL Uses Indexes
7.4.5. The MyISAM Key Cache
7.4.6. MyISAM Index Statistics Collection
7.4.7. How MySQL Opens and Closes Tables
7.4.8. Disadvantages of Creating Many Tables in the Same Database
7.5. Optimizing the MySQL Server
7.5.1. How Compiling and Linking Affects the Speed of MySQL
7.5.2. System Factors and Startup Parameter Tuning
7.5.3. Tuning Server Parameters
7.5.4. Controlling Query Optimizer Performance
7.5.5. The MySQL Query Cache
7.5.6. Examining Thread Information
7.5.7. How MySQL Uses Threads for Client Connections
7.5.8. How MySQL Uses Memory
7.5.9. Enabling Large Page Support
7.5.10. How MySQL Uses Internal Temporary Tables
7.5.11. How MySQL Uses DNS
7.6. Disk Issues
7.6.1. Using Symbolic Links

Optimization is a complex task because ultimately it requires understanding of the entire system to be optimized. Although it may be possible to perform some local optimizations with little knowledge of your system or application, the more optimal you want your system to become, the more you must know about it.

This chapter tries to explain and give some examples of different ways to optimize MySQL. Remember, however, that there are always additional ways to make the system even faster, although they may require increasing effort to achieve.

7.1. Optimization Overview

The most important factor in making a system fast is its basic design. You must also know what kinds of processing your system is doing, and what its bottlenecks are. In most cases, system bottlenecks 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 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 we have the data in main memory, we need to process it to get our result. Having small 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.

MySQL Enterprise For instant notification of system bottlenecks subscribe to the MySQL Enterprise Monitor. For more information, see http://www.mysql.com/products/enterprise/advisors.html.

7.1.1. MySQL Design Limitations and Tradeoffs

When using the MyISAM storage engine, MySQL uses extremely fast table locking that allows multiple readers or a single writer. The biggest problem with this storage engine occurs when you have a steady stream of mixed updates and slow selects on a single table. If this is a problem for certain tables, you can use another storage engine for them. See Chapter 13, Storage Engines.

MySQL can work with both transactional and nontransactional tables. To make it easier to work smoothly with nontransactional tables (which cannot roll back if something goes wrong), MySQL has the following rules. Note that these rules apply only when not running in strict SQL mode or if you use the IGNORE specifier for INSERT or UPDATE.

  • All columns have default values.

  • If you insert an inappropriate or out-of-range value into a column, MySQL sets the column to the “best possible value” instead of reporting an error. For numerical values, this is 0, the smallest possible value or the largest possible value. For strings, this is either the empty string or as much of the string as can be stored in the column.

  • All calculated expressions return a value that can be used instead of signaling an error condition. For example, 1/0 returns NULL.

To change the preceding behaviors, you can enable stricter data handling by setting the server SQL mode appropriately. For more information about data handling, see Section 1.7.6, “How MySQL Deals with Constraints”, Section 5.1.7, “Server SQL Modes”, and Section 12.2.5, “INSERT Syntax”.

7.1.2. Designing Applications for Portability

Because all SQL servers implement different parts of standard SQL, it takes work to write portable database applications. It is very easy to achieve portability for very simple selects and inserts, but becomes more difficult the more capabilities you require. If you want an application that is fast with many database systems, it becomes even more difficult.

All database systems have some weak points. That is, they have different design compromises that lead to different behavior.

To make a complex application portable, you need to determine which SQL servers it must work with, and then determine what features those servers support. You can use the MySQL crash-me program to find functions, types, and limits that you can use with a selection of database servers. crash-me does not check for every possible feature, but it is still reasonably comprehensive, performing about 450 tests. An example of the type of information crash-me can provide is that you should not use column names that are longer than 18 characters if you want to be able to use Informix or DB2.

The crash-me program and the MySQL benchmarks are all very database independent. By taking a look at how they are written, you can get a feeling for what you must do to make your own applications database independent. The programs can be found in the sql-bench directory of MySQL source distributions. They are written in Perl and use the DBI database interface. Use of DBI in itself solves part of the portability problem because it provides database-independent access methods. See Section 7.1.4, “The MySQL Benchmark Suite”.

If you strive for database independence, you need to get a good feeling for each SQL server's bottlenecks. For example, MySQL is very fast in retrieving and updating rows for MyISAM tables, but has a problem in mixing slow readers and writers on the same table. Transactional database systems in general are not very good at generating summary tables from log tables, because in this case row locking is almost useless.

MySQL Enterprise For expert advice on choosing the database engine suitable to your circumstances subscribe to the MySQL Enterprise Monitor. For more information, see http://www.mysql.com/products/enterprise/advisors.html.

To make your application really database independent, you should define an easily extendable interface through which you manipulate your data. For example, C++ is available on most systems, so it makes sense to use a C++ class-based interface to the databases.

If you use some feature that is specific to a given database system (such as the REPLACE statement, which is specific to MySQL), you should implement the same feature for other SQL servers by coding an alternative method. Although the alternative might be slower, it enables the other servers to perform the same tasks.

With MySQL, you can use the /*! */ syntax to add MySQL-specific keywords to a statement. The code inside /* */ is treated as a comment (and ignored) by most other SQL servers. For information about writing comments, see Section 8.5, “Comment Syntax”.

If high performance is more important than exactness, as for some Web applications, it is possible to create an application layer that caches all results to give you even higher performance. By letting old results expire after a while, you can keep the cache reasonably fresh. This provides a method to handle high load spikes, in which case you can dynamically increase the cache size and set the expiration timeout higher until things get back to normal.

In this case, the table creation information should contain information about the initial cache size and how often the table should normally be refreshed.

An attractive alternative to implementing an application cache is to use the MySQL query cache. By enabling the query cache, the server handles the details of determining whether a query result can be reused. This simplifies your application. See Section 7.5.5, “The MySQL Query Cache”.

7.1.3. What We Have Used MySQL For

This section describes an early application for MySQL.

During MySQL initial development, the features of MySQL were made to fit our largest customer, which handled data warehousing for a couple of the largest retailers in Sweden.

From all stores, we got weekly summaries of all bonus card transactions, and were expected to provide useful information for the store owners to help them find how their advertising campaigns were affecting their own customers.

The volume of data was quite huge (about seven million summary transactions per month), and we had data for 4–10 years that we needed to present to the users. We got weekly requests from our customers, who wanted instant access to new reports from this data.

We solved this problem by storing all information per month in compressed “transaction tables.” We had a set of simple macros that generated summary tables grouped by different criteria (product group, customer id, store, and so on) from the tables in which the transactions were stored. The reports were Web pages that were dynamically generated by a small Perl script. This script parsed a Web page, executed the SQL statements in it, and inserted the results. We would have used PHP or mod_perl instead, but they were not available at the time.

For graphical data, we wrote a simple tool in C that could process SQL query results and produce GIF images based on those results. This tool also was dynamically executed from the Perl script that parses the Web pages.

In most cases, a new report could be created simply by copying an existing script and modifying the SQL query that it used. In some cases, we needed to add more columns to an existing summary table or generate a new one. This also was quite simple because we kept all transaction-storage tables on disk. (This amounted to about 50GB of transaction tables and 200GB of other customer data.)

We also let our customers access the summary tables directly with ODBC so that the advanced users could experiment with the data themselves.

This system worked well and we had no problems handling the data with quite modest Sun Ultra SPARCstation hardware (2×200MHz). Eventually the system was migrated to Linux.

7.1.4. The MySQL Benchmark Suite

This benchmark suite is meant to tell any user what operations a given SQL implementation performs well or poorly. You can get a good idea for how the benchmarks work by looking at the code and results in the sql-bench directory in any MySQL source distribution.

Note that this benchmark is single-threaded, so it measures the minimum time for the operations performed. We plan to add multi-threaded tests to the benchmark suite in the future.

To use the benchmark suite, the following requirements must be satisfied:

After you obtain a MySQL source distribution, you can find the benchmark suite located in its sql-bench directory. To run the benchmark tests, build MySQL, and then change location into the sql-bench directory and execute the run-all-tests script:

shell> cd sql-bench
shell> perl run-all-tests --server=server_name

server_name should be the name of one of the supported servers. To get a list of all options and supported servers, invoke this command:

shell> perl run-all-tests --help

The crash-me script also is located in the sql-bench directory. crash-me tries to determine what features a database system supports and what its capabilities and limitations are by actually running queries. For example, it determines:

  • What data types are supported

  • How many indexes are supported

  • What functions are supported

  • How big a query can be

  • How big a VARCHAR column can be

For more information about benchmark results, visit http://www.mysql.com/why-mysql/benchmarks/.

7.1.5. Using Your Own Benchmarks

You should definitely 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.

For examples of portable benchmark programs, look at those in the MySQL benchmark suite. See Section 7.1.4, “The MySQL Benchmark Suite”. You can take any program from this suite and modify it for your own needs. By doing this, you can try different solutions to your problem and test which really is fastest for you.

Another 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, you should put some effort into benchmarking your whole application under the worst possible load. You can use Super Smack, available at http://jeremy.zawodny.com/mysql/super-smack/. As suggested by its name, it can bring a system to its knees, so make sure to use it only on your development systems.

7.2. Optimizing SELECT and Other Statements

First, one factor affects all statements: The more complex your permissions setup, the more overhead you have. Using simpler permissions when you issue 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, it may be worth the time to use a simplified grant structure to reduce permission-checking overhead.

If your problem is with a specific MySQL expression or function, you can perform a timing test by invoking 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.

All MySQL functions should be 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.

7.2.1. Optimizing Queries with EXPLAIN

The EXPLAIN statement can be used either as a synonym for DESCRIBE or as a way to obtain information about how MySQL executes a SELECT statement:

  • EXPLAIN tbl_name is synonymous with DESCRIBE tbl_name or SHOW COLUMNS FROM tbl_name:

    EXPLAIN tbl_name
    

  • When you precede a SELECT statement with the keyword EXPLAIN, MySQL displays information from the optimizer about the query execution plan. That is, MySQL explains how it would process the SELECT, including information about how tables are joined and in which order:

    EXPLAIN [EXTENDED] SELECT select_options
    

This section describes the second use of EXPLAIN for obtaining query execution plan information. See also Section 12.3.2, “EXPLAIN Syntax”. For a description of the DESCRIBE and SHOW COLUMNS statements, see Section 12.3.1, “DESCRIBE Syntax”, and Section 12.5.5.5, “SHOW COLUMNS Syntax”.

With the help of EXPLAIN, you can see where you should add indexes to tables to get a faster SELECT that uses 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 the SELECT statement, begin the statement with SELECT STRAIGHT_JOIN rather than just SELECT. (See Section 12.2.8, “SELECT Syntax”.)

If you have a problem with indexes not being used when you believe that they should be, you should run ANALYZE TABLE to update table statistics such as cardinality of keys, that can affect the choices the optimizer makes. See Section 12.5.2.1, “ANALYZE TABLE Syntax”.

EXPLAIN returns a row of information for each table used in the SELECT statement. The tables are listed in the output in the order that MySQL would read them while processing the query. MySQL resolves all joins using a single-sweep multi-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.

When the EXTENDED keyword is used, EXPLAIN produces extra information that can be viewed by issuing a SHOW WARNINGS statement following the EXPLAIN statement. This information 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.

Each output row from EXPLAIN provides information about one table, and each row contains the following columns:

  • id

    The SELECT identifier. This is the sequential number of the SELECT within the query.

  • select_type

    The type of SELECT, which can be any of those shown in the following table.

    SIMPLESimple SELECT (not using UNION or subqueries)
    PRIMARYOutermost SELECT
    UNIONSecond or later SELECT statement in a UNION
    DEPENDENT UNIONSecond or later SELECT statement in a UNION, dependent on outer query
    UNION RESULTResult of a UNION.
    SUBQUERYFirst SELECT in subquery
    DEPENDENT SUBQUERYFirst SELECT in subquery, dependent on outer query
    DERIVEDDerived table SELECT (subquery in FROM clause)
    UNCACHEABLE SUBQUERYA subquery for which the result cannot be cached and must be re-evaluated for each row of the outer query

    DEPENDENT typically signifies the use of a correlated subquery. See Section 12.2.9.7, “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 is subject to the restrictions detailed in Section 7.5.5.1, “How the Query Cache Operates”. For example, referring to user variables makes a subquery uncacheable.

  • table

    The table to which the row of output refers.

  • type

    The join type. The different join types are listed here, 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 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 7.2.8, “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 7.2.6, “Index Merge Optimization”.

    • unique_subquery

      This type replaces 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

      This join type is the same as ALL, except that only the index tree is scanned. This usually is faster than ALL because the index file usually is smaller than the data file.

      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 allow row retrieval from the table based on constant values or column values from earlier tables.

  • 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, 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 12.1.4, “ALTER TABLE Syntax”.

    To see what indexes a table has, use SHOW INDEX FROM tbl_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 12.2.8.2, “Index Hint Syntax”.

    For MyISAM and BDB tables, running ANALYZE TABLE helps the optimizer choose better indexes. For MyISAM tables, myisamchk --analyze does the same. See Section 12.5.2.1, “ANALYZE TABLE Syntax”, and Section 6.4, “Table Maintenance and Crash Recovery”.

  • key_len

    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

    The ref column shows which columns or constants are compared to the index named in the key column to select rows from the table.

  • 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.

  • Extra

    This column contains additional information about how MySQL resolves the query. The following list explains the values that can appear in this column. If you want to make your queries as fast as possible, you should look out for Extra values of Using filesort and Using temporary.

    • 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.

    • Full scan on NULL key

      This occurs for subquery optimization as a fallback strategy when the optimizer cannot use an index-lookup access method.

    • Impossible WHERE noticed after reading const tables

      MySQL has read all const (and system) tables and notice that the WHERE clause is always false.

    • No tables

      The query has no FROM clause, or has a FROM DUAL clause.

    • Not exists

      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.

    • Range checked for each record (index map: N)

      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 7.2.5, “Range Optimization”, and Section 7.2.6, “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.

    • Select tables optimized away

      The query contained only aggregate functions (MIN(), MAX()) that were all resolved using an index, or COUNT(*) for MyISAM, and no GROUP BY clause. The optimizer determined that only one row should be returned.

    • 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 7.2.13, “ORDER BY Optimization”.

    • 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.

    • 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 7.2.14, “GROUP BY Optimization”.

    • Using sort_union(...), Using union(...), Using intersect(...)

      These indicate how index scans are merged for the index_merge join type. See Section 7.2.6, “Index Merge Optimization”.

    • Using temporary

      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

      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 with pushed condition

      This item applies to NDBCLUSTER 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 7.2.7, “Condition Pushdown Optimization”.

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 7.5.3, “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

Note that the rows column in the output from EXPLAIN is an educated guess from the MySQL join optimizer. You should 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.

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 12.2.9.8, “Subqueries in the FROM clause”.

MySQL Enterprise Subscribers to the MySQL Enterprise Monitor regularly receive expert advice on optimization. For more information, see http://www.mysql.com/products/enterprise/advisors.html.

7.2.2. 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.

Note that 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 7.5.3, “Tuning Server Parameters”.

MySQL Enterprise The MySQL Enterprise Monitor provides a number of advisors specifically designed to improve query performance. For more information, see http://www.mysql.com/products/enterprise/advisors.html.

7.2.3. Speed of SELECT Queries

In general, when you want to make a slow SELECT ... WHERE query faster, the first thing to check is whether you can add an index. All references between different tables should usually be done with indexes. You can use the EXPLAIN statement to determine which indexes are used for a SELECT. See Section 7.2.1, “Optimizing Queries with EXPLAIN, and Section 7.4.4, “How MySQL Uses Indexes”.

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.

7.2.4. WHERE Clause Optimization

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.

Work on the MySQL optimizer is ongoing, so this section is incomplete. MySQL performs a great many optimizations, not all of which are documented here.

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, ... ;

7.2.5. 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 a detailed description of how intervals are extracted from the WHERE clause.

7.2.5.1. 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, so we speak of 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.

  • 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 types of indexes, 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';

Note that some nonconstant values may be converted to constants during the 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.

Currently, 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.

7.2.5.2. 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 7.2.5.1, “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). For example, for this condition:

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

    The single interval is:

    ('foo',10,10) < (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 7.2.5.1, “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.

7.2.6. 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.

Note

If you have upgraded from a previous version of MySQL, you should be aware that this type of join optimization is first introduced in MySQL 5.0, and represents a significant change in behavior with regard to indexes. (Formerly, MySQL was able to use at most only one index for each referenced table.)

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 a range scan is possible on some key, the optimizer will not consider using Index Merge Union or Index Merge Sort-Union algorithms. For example, consider this query:

    SELECT * FROM t1 WHERE (goodkey1 < 10 OR goodkey2 < 20) AND badkey < 30;
    

    For this query, two plans are possible:

    • An Index Merge scan using the (goodkey1 < 10 OR goodkey2 < 20) condition.

    • A range scan using the badkey < 30 condition.

    However, the optimizer considers only the second plan.

  • If your query has a complex WHERE clause with deep AND/OR nesting and MySQL doesn't 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 fulltext 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.

7.2.6.1. 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 or BDB 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 don't 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 or BDB table, it is not used for row retrieval, but is used to filter out rows retrieved using other conditions.

7.2.6.2. 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 or BDB 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;

7.2.6.3. 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.

7.2.7. Condition Pushdown Optimization

This optimization improves the efficiency of a direct comparison between a nonindexed column and a constant. In such cases, the condition is “pushed down” to the storage engine for evaluation. In MySQL 5.0, this optimization can be used only by the NDBCLUSTER storage engine, but we intend to implement it for additional storage engines in future versions of MySQL.

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=NDBCLUSTER;

Condition pushdown can be used with a query against this table such as the query shown here:

SELECT a,b FROM t1 WHERE b = 10;

This can be seen in the output of EXPLAIN SELECT:

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;

With regard to the first of these two queries, condition pushdown is not applicable 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.) In the case of the second query, condition pushdown cannot be employed 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 comparisons which are supported 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 11.4.1, “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.

Condition pushdown capability is not used by default. To enable it, you can start mysqld with the --engine-condition-pushdown option, or you can execute either of the following statements at runtime:

SET engine_condition_pushdown=ON;

SET engine_condition_pushdown=1;

Limitations.  Condition pushdown is subject to the following limitations:

  • In MySQL 5.0, condition pushdown is supported by the NDBCLUSTER storage engine only.

  • 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.

7.2.8. 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.

Note that 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);

7.2.9. 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 12.2.8, “SELECT Syntax”.)

7.2.10. 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 would be 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, tables processed in the inner loops typically are read 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.

  • 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 would be 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

One implication is that 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.

7.2.11. Nested Join Optimization

As of MySQL 5.0.1, the syntax for expressing joins allows nested joins. The following discussion refers to the join syntax described in Section 12.2.8.1, “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 versions of MySQL prior to 5.0.1, parentheses in table_references were just omitted and all join operations were grouped to the left. 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))

Note that 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.

Note that 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 5.0.)

7.2.12. 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

The algorithm that converts outer join operations into inner joins was implemented in full measure, as it has been described here, in MySQL 5.0.1. MySQL 4.1 performs only some simple conversions.

7.2.13. ORDER BY Optimization

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

The index can also be used even if the ORDER BY does not match the index exactly, as long as all of the unused portions of the index and all the 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;

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:

  • You use ORDER BY on different keys:

    SELECT * FROM t1 ORDER BY key1, key2;
    
  • You use ORDER BY on nonconsecutive parts of a key:

    SELECT * FROM t1 WHERE key2=constant ORDER BY key_part2;
    
  • You mix ASC and DESC:

    SELECT * FROM t1 ORDER BY key_part1 DESC, key_part2 ASC;
    
  • The key used to fetch the rows is not the same as the one used in the ORDER BY:

    SELECT * FROM t1 WHERE key2=constant ORDER BY key1;
    
  • You use ORDER BY with an expression that includes terms other than the key column name:

    SELECT * FROM t1 ORDER BY ABS(key);
    SELECT * FROM t1 ORDER BY -key;
    
  • You are joining 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.)

  • You have different ORDER BY and GROUP BY expressions.

  • You index 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 you have a CHAR(20) column, but index only the first 10 bytes, the index cannot distinguish values past the 10th byte and a filesort will be needed.

  • The type of table index used 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 ORDER BY clause explicitly that contains the same column list, MySQL optimizes it away without any speed penalty, although the sorting still occurs. 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;

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 7.2.1, “Optimizing Queries with EXPLAIN.

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 used in 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.

The original filesort algorithm works as follows:

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

  2. For each row, store a pair of values in a buffer (the sort key and the row pointer). The size of the buffer is the value of the sort_buffer_size system variable.

  3. When the buffer gets full, run a qsort (quicksort) on it and store the result in a temporary file. Save a pointer to the sorted block. (If all pairs fit into the sort buffer, no temporary file is created.)

  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 pointer to the row (the last part of the sort key) is written to a result file.

  8. Read the rows in sorted order by using the row pointers in the result file. To optimize this, we read in a big block of row pointers, sort them, and use them to read the rows in sorted order into a row buffer. The size of the buffer is the value of the read_rnd_buffer_size system variable. 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 when evaluating the WHERE clause, and again after sorting the pair values. 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 such that it records not only the sort key value and row position, but also the columns required for the query. This avoids reading the rows twice. The modified filesort algorithm works like this:

  1. Read the rows that match the WHERE clause.

  2. For each row, record a tuple of values consisting of the sort key value and row position, and also the columns required for the query.

  3. Sort the tuples by sort key value

  4. Retrieve the rows in sorted order, but read the required columns directly from the sorted tuples rather than by accessing the table a second time.

Using the modified filesort algorithm, the tuples are longer than the pairs used in the original method, and fewer of them fit in the sort buffer (the size of which is given by sort_buffer_size). As a result, it is possible for the extra I/O to make the modified approach slower, not faster. To avoid a slowdown, the optimization is used 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 that you should see high disk activity and low CPU activity.)

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

If you want 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 size of the sort_buffer_size variable.

  • Increase the size of the read_rnd_buffer_size variable.

  • 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 tmpdir to point to a dedicated file system with large amounts of free space. Also, this option accepts several paths that are used in round-robin fashion, so 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, NetWare, and OS/2. The paths should be for directories in file systems that are located on different physical disks, not different partitions on the same disk.

7.2.14. 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.

In MySQL, GROUP BY is used for sorting, so the server may also apply ORDER BY optimizations to grouping. See Section 7.2.13, “ORDER BY Optimization”.

There are two ways to execute a GROUP BY query via 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.

7.2.14.1. Loose index scan

The most efficient way to process GROUP BY is when the index is used to directly retrieve the group 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 7.2.1, “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 includes the first consecutive parts of the index. (If, instead of GROUP BY, the query has a DISTINCT clause, all distinct attributes refer to the beginning of the index.)

  • The only aggregate functions used (if any) are MIN() and MAX(), and all of them refer to the same column.

  • 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.

The EXPLAIN output for such queries shows Using index for group-by in the Extra column.

The following queries fall into this category, assuming that there is an index idx(c1,c2,c3) on table t1(c1,c2,c3,c4):

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(), for example:

    SELECT c1, SUM(c2) FROM t1 GROUP BY c1;
    
  • The columns in the GROUP BY clause do not refer to the beginning of the index, as shown here:

    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, an example being:

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

7.2.14.2. 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 is still 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. Notice that 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.

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 (assuming that there is an index idx(c1,c2,c3) on table t1(c1,c2,c3,c4)).

  • 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;
    

7.2.15. DISTINCT Optimization

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

Because DISTINCT may use GROUP BY, you should be aware of how MySQL works with columns in ORDER BY or HAVING clauses that are not part of the selected columns. See Section 11.11.3, “GROUP BY and HAVING with Hidden Columns”.

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 7.2.14, “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;

7.2.16. Optimizing IN/=ANY Subqueries

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. Suggestions on what you can do to help the optimizer are given at the end of the discussion.

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)

The following discussion assumes a single pair of outer and inner expression values for simplicity.

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 don't care.)

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:

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 subqery-specific access methods also have or-null variants. However, 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. In MySQL 5.0 before 5.0.36, the optimizer chose speed over distinguishing a NULL from FALSE result, so for some queries, you might get a FALSE result rather than NULL.

As of MySQL 5.0.36, 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 that 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 doesn't 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:

  • A column must be declared as NOT NULL if it really is. (This also helps other aspects of the optimizer.)

  • If you don't 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 ...) will never be evaluated because MySQL stops evaluating AND parts as soon as the expression result is clear.

7.2.17. LIMIT Optimization

In some cases, MySQL handles a query differently when you are using LIMIT row_count and not using HAVING:

  • If you are selecting 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 use 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 must be selected, and most or all of them must be sorted, before it can be ascertained that the first row_count rows have been found. In either case, after the initial rows have been found, there is no need to sort any remainder of the result set, and MySQL does not do so.

  • When combining 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 key in order (or doing a sort on the key) and then calculating summaries until the key 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.

  • LIMIT 0 quickly returns an empty set. This can be useful for checking the validity of a query. When using one of the MySQL APIs, it can also be employed for obtaining the types of the result columns. (This trick does not work in the MySQL Monitor (the mysql program), which merely displays Empty set in such cases; you should instead use SHOW COLUMNS or DESCRIBE for this purpose.)

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

7.2.18. How to Avoid Table Scans

The output from EXPLAIN shows ALL in the type column when MySQL uses a 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 7.2.4, “WHERE Clause Optimization”.

  • 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.

MySQL Enterprise For expert advice on avoiding excessive table scans subscribe to the MySQL Enterprise Monitor. For more information, see http://www.mysql.com/products/enterprise/advisors.html.

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:

7.2.19. Speed of INSERT Statements

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:

  • If you are inserting many rows from the same client at the same time, use INSERT statements with multiple VALUES lists to insert several rows at a time. This is considerably faster (many times faster in some cases) than using separate single-row INSERT statements. If you are adding data to a nonempty table, you can tune the bulk_insert_buffer_size variable to make data insertion even faster. See Section 5.1.3, “Server System Variables”.

  • If multiple clients are inserting a lot of rows, you can get higher speed by using the INSERT DELAYED statement. See Section 12.2.5.2, “INSERT DELAYED Syntax”.

  • 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 7.3.3, “Concurrent Inserts”.

  • When loading a table from a text file, use LOAD DATA INFILE. This is usually 20 times faster than using INSERT statements. See Section 12.2.6, “LOAD DATA INFILE Syntax”.

  • 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. Optionally create the table with CREATE TABLE.

    2. Execute a FLUSH TABLES statement or a mysqladmin flush-tables command.

    3. Use myisamchk --keys-used=0 -rq /path/to/db/tbl_name. This removes all use of indexes for the table.

    4. Insert data into the table with LOAD DATA INFILE. This does not update any indexes and therefore is very fast.

    5. If you intend only to read from the table in the future, use myisampack to compress it. See Section 13.1.3.3, “Compressed Table Characteristics”.

    6. 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 that updating the index during LOAD DATA INFILE because it avoids lots of disk seeks. The resulting index tree is also perfectly balanced.

    7. 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.

    To obtain faster insertions for transactional tables, you should use START TRANSACTION and COMMIT instead of LOCK TABLES.

    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 allow other threads access to the 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 7.5.3, “Tuning Server Parameters”.

MySQL Enterprise For more advice on optimizing the performance of your server, subscribe to the MySQL Enterprise Monitor. Numerous advisors are dedicated to monitoring performance. For more information, see http://www.mysql.com/products/enterprise/advisors.html.

7.2.20. 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 12.5.2.5, “OPTIMIZE TABLE Syntax”.

7.2.21. Speed of DELETE Statements

The time required to delete individual rows 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 7.5.3, “Tuning Server Parameters”.

To delete all rows from a table, TRUNCATE TABLE tbl_name is faster than 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 12.2.10, “TRUNCATE Syntax”.

7.2.22. Other Optimization Tips

This section lists a number of miscellaneous tips for improving query processing speed:

  • Use persistent connections to the database to avoid connection overhead. If you cannot use persistent connections and you are initiating many new connections to the database, you may want to change the value of the thread_cache_size variable. See Section 7.5.3, “Tuning Server Parameters”.

  • Always check whether all your queries really use the indexes that you have created in the tables. In MySQL, you can do this with the EXPLAIN statement. See Section 7.2.1, “Optimizing Queries with EXPLAIN.

  • 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, you should 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 allow concurrent inserts), even in tables that have deleted rows. See Section 7.3.3, “Concurrent Inserts”.

    MySQL Enterprise For optimization tips geared to your specific circumstances subscribe to the MySQL Enterprise Monitor. For more information, see http://www.mysql.com/products/enterprise/advisors.html.

  • To fix any compression issues that may have occurred with ARCHIVE tables, you can use OPTIMIZE TABLE. See Section 13.8, “The ARCHIVE Storage Engine”.

  • 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.

  • In some cases, it may make sense to introduce a column that is “hashed” based on information from other columns. If this column is short, reasonably unique, and indexed, it may be much 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(col1,col2))
      AND col1='constant' AND col2='constant';
    
  • For MyISAM tables that change frequently, you should 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 13, 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 13, Storage Engines.

  • 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.

  • If you need to collect statistics from large log tables, use summary tables instead of scanning the entire log table. Maintaining the summaries should be much faster than trying to calculate statistics “live.” Regenerating new summary tables from the logs when things change (depending on business decisions) is faster than changing the running application.

  • If possible, you should 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.

  • Take advantage of the fact that columns have default values. Insert values explicitly only when the value to be inserted differs from the default. This reduces the parsing that MySQL must do and improves the insert speed.

  • In some cases, it is convenient to 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 may save a lot of accesses at some stage. This is practical when you have data that does not conform well to a rows-and-columns table structure.

  • Normally, you should try to keep all data nonredundant (observing what is referred to in database theory as third normal form). However, there may be situations in which it can be advantageous to duplicate information or create summary tables to gain more speed.

  • Stored routines or UDFs (user-defined functions) may be a good way to gain performance for some tasks. See Section 18.2, “Using Stored Routines (Procedures and Functions)”, and Section 21.2, “Adding New Functions to MySQL”, for more information.

  • You can increase performance by caching queries or answers in your application and then executing many inserts or updates together. If your database system supports table locks, this should help to ensure 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 7.5.5, “The MySQL Query Cache”.

  • Use INSERT DELAYED when you do not need to know when your data is written. This reduces the overall insertion impact because many rows can be written with a single disk write.

  • Use INSERT LOW_PRIORITY when you want to give SELECT statements higher priority than your inserts.

    Use SELECT HIGH_PRIORITY to get retrievals that jump the queue. That is, the SELECT is executed even if there is another client waiting to do a write.

    LOW_PRIORITY and HIGH_PRIORITY have an effect only for storage engines that use only table-level locking (MyISAM, MEMORY, MERGE).

  • Use multiple-row INSERT statements to store many rows with one SQL statement. Many SQL servers support this, including MySQL.

  • Use LOAD DATA INFILE to load large amounts of data. This is faster than using INSERT statements.

  • Use AUTO_INCREMENT columns so that each row in a table can be identified by a single unique value. unique values.

  • Use OPTIMIZE TABLE once in a while to avoid fragmentation with dynamic-format MyISAM tables. See Section 13.1.3, “MyISAM Table Storage Formats”.

  • Use MEMORY (HEAP) tables when possible to get more speed. See Section 13.4, “The MEMORY (HEAP) Storage Engine”. MEMORY tables are useful for noncritical data that is accessed often, such as information about the last displayed banner for users who don't have cookies enabled in their Web browser. User sessions are another alternative available in many Web application environments for handling volatile state data.

  • With Web servers, images and other binary assets should normally be stored as files. That is, store only a reference to the file rather than the file itself in the database. Most Web servers are better at caching files than database contents, so using files is generally faster.

  • Columns with identical information in different tables should be declared to have identical data types so that joins based on the corresponding columns will be faster.

  • Try to keep column names simple. 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, you should keep them shorter than 18 characters.

  • If you need really high speed, you should take a look at the low-level interfaces for data storage that the different SQL servers support. For example, by accessing the MySQL MyISAM storage engine directly, you could get a speed increase of two to five times compared to using the SQL interface. To be able to do this, the data must be on the same server as the application, and usually it should only be accessed by one process (because external file locking is really slow). One could eliminate these problems by introducing low-level MyISAM commands in the MySQL server (this could be one easy way to get more performance if needed). By carefully designing the database interface, it should be quite easy to support this type of optimization.

  • If you are using numerical 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 need not parse your text files to find line and column boundaries.

  • 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 16, Replication.

  • 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 should ensure that the table is okay by running the server with the --myisam-recover 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.)

7.3. Locking Issues

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 7.3.1, “Internal Locking Methods”.

  • External locking occurs when the server and other programs lock table files to coordinate among themselves which program can access the tables at which time. See Section 7.3.4, “External Locking”.

7.3.1. 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. External locking occurs when the server and other programs lock table files to coordinate among themselves which program can access the tables at which time. See Section 7.3.4, “External Locking”.

MySQL uses table-level locking for MyISAM, MEMORY and MERGE tables, page-level locking for BDB tables, and row-level locking for InnoDB tables.

In many cases, you can make an educated guess about which locking type is best for an application, but generally it is difficult to say that a given lock type is better than another. Everything depends on the application and different parts of an application may require different lock types.

To decide whether you want to use a storage engine with row-level locking, you should look at what your application does and what mix of select and update statements it uses. For example, most Web applications perform many selects, relatively few deletes, updates based mainly on key values, and inserts into a few specific tables. The base MySQL MyISAM setup is very well tuned for this.

MySQL Enterprise The MySQL Enterprise Monitor provides expert advice on when to use table-level locking and when to use row-level locking. To subscribe see http://www.mysql.com/products/enterprise/advisors.html.

Table locking in MySQL is deadlock-free for storage engines that use table-level locking. Deadlock avoidance is managed by always requesting all needed locks at once at the beginning of a query and always locking the tables in the same order.

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 if there is heavy SELECT activity for the table. However, if you have 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 7.3.2, “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 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. This behavior is altered by the concurrent_insert system variable. See Section 7.3.3, “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 real_table when concurrent inserts are not possible, you can insert rows into a temporary table temp_table and update the real table with the rows from the temporary table periodically. This can be done with the following code:

mysql> LOCK TABLES real_table WRITE, temp_table WRITE;
mysql> INSERT INTO real_table SELECT * FROM temp_table;
mysql> DELETE FROM temp_table;
mysql> UNLOCK TABLES;

InnoDB uses row locks and BDB uses page locks. Deadlocks are possible for these storage engines because they automatically acquire locks during the processing of SQL statements, not at the start of the transaction.

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

Disadvantages of row-level locking:

  • Requires more memory than page-level or table-level locks

  • Slower than page-level or table-level locks when used on a large part of the table because you must acquire many more locks

  • Slower than other locks if you often do GROUP BY operations on a large part of the data or if you must scan the entire table frequently

Generally, table locks are superior to page-level or 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 or page-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 page-level or 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 11.10.4, “Miscellaneous Functions”.

7.3.2. Table Locking Issues

To achieve a very high lock speed, MySQL uses table locking (instead of page, row, or column locking) for all storage engines except InnoDB, BDB, and NDBCLUSTER.

For InnoDB and BDB tables, MySQL uses table locking only if you explicitly lock the table with LOCK TABLES. For these storage engines, avoid using LOCK TABLES at all, because InnoDB uses automatic row-level locking and BDB uses page-level locking to ensure transaction isolation.

For large tables, table locking is often better than row locking, but there are some disadvantages:

  • 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. During the update, all other sessions that want to access this particular table must wait until the update is done.

  • Table locking causes problems in cases such as 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.

Table locking is also disadvantageous under the following scenario:

  • A session issues a SELECT that takes a long time to run.

  • Another session then issues an UPDATE on the same table. This session waits until the SELECT is finished.

  • Another session issues another SELECT statement on the same table. Because UPDATE has higher priority than SELECT, this SELECT waits for the UPDATE to finish, after waiting for the first SELECT to finish.

The following items describe some ways to avoid or reduce contention caused by table locking:

Here are some tips concerning table locks in MySQL:

7.3.3. 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), inserts can be performed to add rows to the end of the table at the same time that SELECT statements are reading rows from the table.

The concurrent_insert system variable can be set to modify the concurrent-insert processing. By default, the variable is set to 1 and concurrent inserts are handled as just described. If concurrent_insert is set to 0, concurrent inserts are disabled. If the variable is set to 2, concurrent inserts at the end of the table are allowed even for tables that have deleted rows. See also the description of the concurrent_insert system variable.

Under circumstances where concurrent inserts can be used, there is seldom any need to use the DELAYED modifier for INSERT statements. See Section 12.2.5.2, “INSERT DELAYED Syntax”.

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 5.2.3, “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 allows 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.

7.3.4. External Locking

External locking is the use of file system locking to manage contention for 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 don't stop mysqld, you should 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 7.5.2, “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. From MySQL 4.0 on, external locking is disabled by default. Before MySQL 4.0, external locking is enabled by default on Linux or when MySQL is configured to use MIT-pthreads.

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:

  • You should not use the query cache for queries that use tables that are updated by another process.

  • You should 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.)

7.4. Optimizing Database Structure

7.4.1. Make Your Data as Small as Possible

One of the most basic optimizations is to design your tables to take as little space on the disk as possible. This can result in huge improvements because disk reads are faster, and smaller tables normally require less main memory while their contents are being actively processed during query execution. Indexing also is a lesser resource burden if done on smaller columns.

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 may give you a big performance gain. See Chapter 13, Storage Engines.

You can get better performance for a table and minimize storage space by using the techniques listed here:

  • 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 everything faster and you save one bit per column. If you really need NULL in your application, you should definitely use it. Just avoid having it on all columns by default.

  • 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 unfortunately may waste some space. See Section 13.1.3, “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.

  • Starting with MySQL 5.0.3, InnoDB tables use a more compact storage format. In earlier versions of MySQL, InnoDB rows contain some redundant information, such as the number of columns and the length of each column, even for fixed-size columns. By default, tables are created in the compact format (ROW_FORMAT=COMPACT). If you wish to downgrade to older versions of MySQL, you can request the old format with ROW_FORMAT=REDUNDANT.

    The presence of 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 UTF-8 data are stored. With ROW_FORMAT=REDUNDANT, a UTF-8 CHAR(N) occupies 3 × N bytes, given that the maximum length of a UTF-8 encoded character is three bytes. Many languages can be written primarily using single-byte UTF-8 characters, so a fixed storage length often wastes space. With ROW_FORMAT=COMPACT format, InnoDB allocates a variable amount of storage in the range from N to 3 × N bytes 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.

  • The primary index of a table should be as short as possible. This makes identification of each row easy and efficient.

  • Create only the indexes that you really need. Indexes are good for retrieval but bad when you need to store data quickly. If you access a table mostly by searching on a combination of columns, create an index on them. 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 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 12.1.8, “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 7.5.3, “Tuning Server Parameters”.

  • 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.

7.4.2. Column Indexes

All MySQL data types can be indexed. Use of indexes on the relevant columns is the best way to improve the performance of SELECT operations.

The maximum number of indexes per table and the maximum index length is defined per storage engine. See Chapter 13, 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.

With col_name(N) syntax in an index specification, you can create an index that uses only the first N characters of a string 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). Note that prefix limits are measured in bytes, whereas the prefix length in CREATE TABLE statements is interpreted as number of characters. Be sure to take this into account when specifying a prefix length for a column that uses a multi-byte character set.

You can also create FULLTEXT indexes. These are used for full-text searches. Only the MyISAM storage engine supports 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 11.8, “Full-Text Search Functions”.

You can also create indexes on spatial data types. Currently, only MyISAM supports R-tree indexes on spatial types. As of MySQL 5.0.16, other storage engines use B-trees for indexing spatial types (except for ARCHIVE and NDBCLUSTER, which do not support spatial type indexing).

The MEMORY storage engine uses HASH indexes by default, but also supports BTREE indexes.

7.4.3. Multiple-Column Indexes

MySQL can create composite indexes (that is, indexes on multiple columns). An index may consist of up to 15 columns. For certain data types, you can index a prefix of the column (see Section 7.4.2, “Column Indexes”).

A multiple-column index can be considered a sorted array containing values that are created by concatenating the values of the indexed columns.

MySQL uses multiple-column indexes in such a way that queries are fast when you specify a known quantity for the first column of the index in a WHERE clause, even if you do not specify values for the other columns.

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 queries that specify values in a known range for last_name, or for both last_name and first_name. Therefore, the name index is used 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 in the following queries:

SELECT * FROM test WHERE first_name='Michael';

SELECT * FROM test
  WHERE last_name='Widenius' OR first_name='Michael';

The manner in which MySQL uses indexes to improve query performance is discussed further in Section 7.4.4, “How MySQL Uses Indexes”.

7.4.4. 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. If a table has 1,000 rows, this is at least 100 times faster than reading sequentially. If you need to access most of the rows, it is faster to read sequentially, because this minimizes disk seeks.

Most MySQL indexes (PRIMARY KEY, UNIQUE, INDEX, and FULLTEXT) are stored in B-trees. Exceptions are that indexes on spatial data types use R-trees, and that MEMORY tables also support hash indexes.

Strings are automatically prefix- and end-space compressed. See Section 12.1.8, “CREATE INDEX Syntax”.

In general, indexes are used as described in the following discussion. Characteristics specific to hash indexes (as used in MEMORY tables) are described at the end of this section.

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.

  • 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.

    Comparison of dissimilar columns may prevent use of indexes if values cannot be compared directly without conversion. Suppose that a numeric column is compared to a string column. 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 key (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 7.2.13, “ORDER BY Optimization”, and Section 7.2.14, “GROUP BY Optimization”.

  • In some cases, a query can be optimized to retrieve values without consulting the data rows. If a query uses only columns from a table that are numeric and that form a leftmost prefix for some key, the selected values may be retrieved from the index tree for greater speed:

    SELECT key_part3 FROM tbl_name
      WHERE key_part1=1
    

Suppose that you issue the following SELECT statement:

mysql> 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 will attempt to use the Index Merge optimization (see Section 7.2.6, “Index Merge Optimization”), or attempt to find the most restrictive index by deciding which index finds fewer 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 find 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 an index 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).

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 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.

  • 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 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.)

MySQL Enterprise Often, it is not possible to predict exactly what indexes will be required or will be most efficient — actual table usage is the best indicator. The MySQL Enterprise Monitor provides expert advice on this topic. For more information, see http://www.mysql.com/products/enterprise/advisors.html.

7.4.5. 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:

  • Access to the key cache no longer is serialized among threads. 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).

MySQL Enterprise For expert advice on identifying the optimum size for key_buffer_size, subscribe to the MySQL Enterprise Monitor. See http://www.mysql.com/products/enterprise/advisors.html.

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 a special queue (LRU chain) of all used blocks. When a block is accessed, it is placed at the end of the queue. When blocks need to be replaced, blocks at the beginning of the queue are the least recently used and become the first candidates for eviction.

7.4.5.1. 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.

7.4.5.2. 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 12.5.6.1, “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;

Note that you cannot destroy the default key cache. Any attempt to do this will be 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.

CACHE INDEX 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

MySQL Enterprise For advice on how best to configure your my.cnf/my.ini option file subscribe to MySQL Enterprise Monitor. Recommendations are based on actual table usage. For more information, see http://www.mysql.com/products/enterprise/advisors.html.

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

7.4.5.3. Midpoint Insertion Strategy

By default, the key cache management system uses the 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 sub-chain and a warm sub-chain. 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 sub-chain. After a certain number of hits (accesses of the block), it is promoted to the hot sub-chain. 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 sub-chain is placed at the end of the chain. The block then circulates within this sub-chain. If the block stays at the beginning of the sub-chain for a long enough time, it is demoted to the warm chain. 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 sub-chain not accessed within the last N × key_cache_age_threshold / 100 hits is to be moved to the beginning of the warm sub-chain. It then becomes the first candidate for eviction, because blocks for replacement always are taken from the beginning of the warm sub-chain.

The midpoint insertion strategy allows 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 sub-chain during an index scan operation as well.

7.4.5.4. 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 allows 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.

7.4.5.5. 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.

7.4.5.6. 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.

7.4.6. 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 allow that average size to be collected or used.

For MyISAM tables, you have some control over collection of table statistics by means of the myisam_stats_method system variable. This variable has three possible values, which differ as follows:

  • When myisam_stats_method is 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 myisam_stats_method is 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 myisam_stats_method is 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 myisam_stats_method system variable has global and session values. Setting the global value affects MyISAM statistics collection for all MyISAM tables. 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 table statistics, you can use any of the following methods:

Some caveats regarding the use of 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 myisam_stats_method has at the time. Thus, if you collect statistics using one method, but myisam_stats_method 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 MyISAM table.

  • myisam_stats_method applies only to MyISAM tables. Other storage engines have only one method for collecting table statistics. Usually it is closer to the nulls_equal method.

7.4.7. 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_cache, max_connections, and max_tmp_tables system variables affect the maximum number of files the server keeps open. If you increase one or more 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 allow 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_cache is related to max_connections. For example, for 200 concurrent running connections, you should have 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_cache setting. If table_cache is set too high, MySQL may run out of file descriptors and refuse connections, fail to perform queries, and be very unreliable. You also have to take into account that the MyISAM storage engine needs two file descriptors for each unique open table. You can increase the number of file descriptors available to MySQL using the --open-files-limit startup option to mysqld. See Section B.1.2.18, “'File' Not Found and Similar Errors”.

The cache of open tables is kept at a level of table_cache entries. The default value is 64; this can be changed with the --table_cache option to mysqld. Note that MySQL may temporarily open more tables than this to execute queries.

MySQL Enterprise Performance may suffer if table_cache is set too low. For expert advice on the optimum value for this variable, subscribe to the MySQL Enterprise Monitor. For more information, see http://www.mysql.com/products/enterprise/advisors.html.

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_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 12.2.4, “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, you should increase the table cache size. See Section 5.1.3, “Server System Variables”, and Section 5.1.6, “Server Status Variables”.

7.4.8. 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 allowed in the table cache.

7.5. Optimizing the MySQL Server

7.5.1. How Compiling and Linking Affects the Speed of MySQL

Most of the following tests were performed on Linux with the MySQL benchmarks, but they should give some indication for other operating systems and workloads.

You obtain the fastest executables when you link with -static.

On Linux, it is best to compile the server with pgcc and -O3. You need about 200MB memory to compile sql_yacc.cc with these options, because gcc or pgcc needs a great deal of memory to make all functions inline. You should also set CXX=gcc when configuring MySQL to avoid inclusion of the libstdc++ library, which is not needed. Note that with some versions of pgcc, the resulting binary runs only on true Pentium processors, even if you use the compiler option indicating that you want the resulting code to work on all x586-type processors (such as AMD).

By using a better compiler and compilation options, you can obtain a 10–30% speed increase in applications. This is particularly important if you compile the MySQL server yourself.

When we tested both the Cygnus CodeFusion and Fujitsu compilers, neither was sufficiently bug-free to allow MySQL to be compiled with optimizations enabled.

The standard MySQL binary distributions are compiled with support for all character sets. When you compile MySQL yourself, you should include support only for the character sets that you are going to use. This is controlled by the --with-charset option to configure.

Here is a list of some measurements that we have made:

  • If you use pgcc and compile everything with -O6, the mysqld server is 1% faster than with gcc 2.95.2.

  • If you link dynamically (without -static), the result is 13% slower on Linux. Note that you still can use a dynamically linked MySQL library for your client applications. It is the server that is most critical for performance.

  • For a connection from a client to a server running on the same host, if you connect using TCP/IP rather than a Unix socket file, performance is 7.5% slower. (On Unix, if you connect to the host name localhost, MySQL uses a socket file by default.)

  • For TCP/IP connections from a client to a server, connecting to a remote server on another host is 8–11% slower than connecting to a server on the same host, even for connections faster than 100Mb/s Ethernet.

  • When running our benchmark tests using secure connections (all data encrypted with internal SSL support) performance was 55% slower than with unencrypted connections.

  • If you compile with --with-debug=full, most queries are 20% slower. Some queries may take substantially longer; for example, the MySQL benchmarks run 35% slower. If you use --with-debug (without =full), the speed decrease is only 15%. For a version of mysqld that has been compiled with --with-debug=full, you can disable memory checking at runtime by starting it with the --skip-safemalloc option. The execution speed should then be close to that obtained when configuring with --with-debug.

  • On a Sun UltraSPARC-IIe, a server compiled with Forte 5.0 is 4% faster than one compiled with gcc 3.2.

  • On a Sun UltraSPARC-IIe, a server compiled with Forte 5.0 is 4% faster in 32-bit mode than in 64-bit mode.

  • Compiling with gcc 2.95.2 for UltraSPARC with the -mcpu=v8 -Wa,-xarch=v8plusa options gives 4% more performance.

  • On Solaris 2.5.1, MIT-pthreads is 8–12% slower than Solaris native threads on a single processor. With greater loads or more CPUs, the difference should be larger.

  • Compiling on Linux-x86 using gcc without frame pointers (-fomit-frame-pointer or -fomit-frame-pointer -ffixed-ebp) makes mysqld 1–4% faster.

Binary MySQL distributions for Linux that are provided by us used to be compiled with pgcc. We had to go back to regular gcc due to a bug in pgcc that would generate binaries that do not run on AMD. We will continue using gcc until that bug is resolved. In the meantime, if you have a non-AMD machine, you can build a faster binary by compiling with pgcc. The standard MySQL Linux binary is linked statically to make it faster and more portable.

7.5.2. 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.

The operating system to use is very important. To get the best use of multiple-CPU machines, you should use Solaris (because its threads implementation works well) or Linux (because the 2.4 and later kernels have good SMP support). Note that older Linux kernels have a 2GB filesize limit by default. If you have such a kernel and a need for files larger than 2GB, you should get the Large File Support (LFS) patch for the ext2 file system. Other file systems such as ReiserFS and XFS do not have this 2GB limitation.

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. Since MySQL 4.0, the default has been for external locking to be disabled on all systems. The --external-locking and --skip-external-locking options explicitly enable and disable external locking.

    Note that 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.

7.5.3. 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:

help                              TRUE
abort-slave-event-count           0
allow-suspicious-udfs             FALSE
auto-increment-increment          1
auto-increment-offset             1
automatic-sp-privileges           TRUE
basedir                           /home/jon/bin/mysql-5.0/
bdb                               FALSE
bind-address                      (No default value)
character-set-client-handshake    TRUE
character-set-filesystem          binary
character-set-server              latin1
character-sets-dir                /home/jon/bin/mysql-5.0/share/mysql/charsets/
chroot                            (No default value)
collation-server                  latin1_swedish_ci
completion-type                   0
concurrent-insert                 1
console                           FALSE
datadir                           /home/jon/bin/mysql-5.0/var/
default-character-set             latin1
default-collation                 latin1_swedish_ci
default-time-zone                 (No default value)
disconnect-slave-event-count      0
enable-locking                    FALSE
enable-pstack                     FALSE
engine-condition-pushdown         FALSE
external-locking                  FALSE
federated                         TRUE
gdb                               FALSE
large-pages                       FALSE
init-connect                      (No default value)
init-file                         (No default value)
init-slave                        (No default value)
innodb                            TRUE
innodb_checksums                  TRUE
innodb_data_home_dir              (No default value)
innodb_adaptive_hash_index        TRUE
innodb_doublewrite                TRUE
innodb_fast_shutdown              1
innodb_file_per_table             FALSE
innodb_flush_log_at_trx_commit    1
innodb_flush_method               (No default value)
innodb_locks_unsafe_for_binlog    FALSE
innodb_log_arch_dir               (No default value)
innodb_log_group_home_dir         (No default value)
innodb_max_dirty_pages_pct        90
innodb_max_purge_lag              0
innodb_rollback_on_timeout        FALSE
innodb_status_file                FALSE
innodb_support_xa                 TRUE
innodb_table_locks                TRUE
isam                              FALSE
language                          /home/jon/bin/mysql-5.0/share/mysql/english/
lc-time-names                     en_US
local-infile                      TRUE
log                               (No default value)
log-bin                           (No default value)
log-bin-index                     (No default value)
log-bin-trust-function-creators   FALSE
log-bin-trust-routine-creators    FALSE
log-error
log-isam                          myisam.log
log-queries-not-using-indexes     FALSE
log-short-format                  FALSE
log-slave-updates                 FALSE
log-slow-admin-statements         FALSE
log-slow-queries                  (No default value)
log-tc                            tc.log
log-tc-size                       24576
log-update                        (No default value)
log-warnings                      1
low-priority-updates              FALSE
master-connect-retry              60
master-host                       (No default value)
master-info-file                  master.info
master-password                   (No default value)
master-port                       3306
master-retry-count                86400
master-ssl                        FALSE
master-ssl-ca                     (No default value)
master-ssl-capath                 (No default value)
master-ssl-cert                   (No default value)
master-ssl-cipher                 (No default value)
master-ssl-key                    (No default value)
master-user                       test
max-binlog-dump-events            0
memlock                           FALSE
merge                             TRUE
myisam-recover                    OFF
ndbcluster                        FALSE
new                               FALSE
old-passwords                     FALSE
old-style-user-limits             FALSE
pid-file                          /home/jon/bin/mysql-5.0/var/tonfisk.pid
port                              3306
port-open-timeout                 0
relay-log                         (No default value)
relay-log-index                   (No default value)
relay-log-info-file               relay-log.info
replicate-same-server-id          FALSE
report-host                       (No default value)
report-password                   (No default value)
report-port                       3306
report-user                       (No default value)
rpl-recovery-rank                 0
safe-user-create                  FALSE
secure-auth                       FALSE
secure-file-priv                  (No default value)
server-id                         0
show-slave-auth-info              FALSE
skip-grant-tables                 FALSE
skip-slave-start                  FALSE
slave-load-tmpdir                 /tmp/
socket                            /tmp/mysql.sock
sporadic-binlog-dump-fail         FALSE
sql-mode                          OFF
symbolic-links                    TRUE
sysdate-is-now                    FALSE
tc-heuristic-recover              (No default value)
temp-pool                         TRUE
timed_mutexes                     FALSE
tmpdir                            (No default value)
use-symbolic-links                TRUE
verbose                           TRUE
warnings                          1
back_log                          50
binlog_cache_size                 32768
bulk_insert_buffer_size           8388608
connect_timeout                   10
date_format                       (No default value)
datetime_format                   (No default value)
default_week_format               0
delayed_insert_limit              100
delayed_insert_timeout            300
delayed_queue_size                1000
div_precision_increment           4
expire_logs_days                  0
flush_time                        0
ft_max_word_len                   84
ft_min_word_len                   4
ft_query_expansion_limit          20
ft_stopword_file                  (No default value)
group_concat_max_len              1024
innodb_additional_mem_pool_size   1048576
innodb_autoextend_increment       8
innodb_buffer_pool_awe_mem_mb     0
innodb_buffer_pool_size           8388608
innodb_commit_concurrency         0
innodb_concurrency_tickets        500
innodb_file_io_threads            4
innodb_force_recovery             0
innodb_lock_wait_timeout          50
innodb_log_buffer_size            1048576
innodb_log_file_size              5242880
innodb_log_files_in_group         2
innodb_mirrored_log_groups        1
innodb_open_files                 300
innodb_sync_spin_loops            20
innodb_thread_concurrency         8
innodb_thread_sleep_delay         10000
interactive_timeout               28800
join_buffer_size                  131072
keep_files_on_create              FALSE
key_buffer_size                   8384512
key_cache_age_threshold           300
key_cache_block_size              1024
key_cache_division_limit          100
long_query_time                   10
lower_case_table_names            0
max_allowed_packet                1048576
max_binlog_cache_size             18446744073709547520
max_binlog_size                   1073741824
max_connect_errors                10
max_connections                   100
max_delayed_threads               20
max_error_count                   64
max_heap_table_size               16777216
max_join_size                     18446744073709551615
max_length_for_sort_data          1024
max_prepared_stmt_count           16382
max_relay_log_size                0
max_seeks_for_key                 18446744073709551615
max_sort_length                   1024
max_sp_recursion_depth            0
max_tmp_tables                    32
max_user_connections              0
max_write_lock_count              18446744073709551615
multi_range_count                 256
myisam_block_size                 1024
myisam_data_pointer_size          6
myisam_max_extra_sort_file_size   2147483648
myisam_max_sort_file_size         9223372036853727232
myisam_repair_threads             1
myisam_sort_buffer_size           8388608
myisam_stats_method               nulls_unequal
net_buffer_length                 16384
net_read_timeout                  30
net_retry_count                   10
net_write_timeout                 60
open_files_limit                  0
optimizer_prune_level             1
optimizer_search_depth            62
plugin_dir
preload_buffer_size               32768
query_alloc_block_size            8192
query_cache_limit                 1048576
query_cache_min_res_unit          4096
query_cache_size                  0
query_cache_type                  1
query_cache_wlock_invalidate      FALSE
query_prealloc_size               8192
range_alloc_block_size            4096
read_buffer_size                  131072
read_only                         FALSE
read_rnd_buffer_size              262144
record_buffer                     131072
relay_log_purge                   TRUE
relay_log_space_limit             0
slave_compressed_protocol         FALSE
slave_net_timeout                 3600
slave_transaction_retries         10
slow_launch_time                  2
sort_buffer_size                  2097144
sync-binlog                       0
sync-frm                          TRUE
table_cache                       64
table_lock_wait_timeout           50
thread_cache_size                 0
thread_concurrency                10
thread_stack                      262144
time_format                       (No default value)
tmp_table_size                    33554432
transaction_alloc_block_size      8192
transaction_prealloc_size         4096
updatable_views_with_limit        1
wait_timeout                      28800

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 5.1.3, “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 you get better performance by giving MySQL more memory.

When tuning a MySQL server, the two most important variables to configure are key_buffer_size and table_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 256MB of memory and many tables and want maximum performance with a moderate number of clients, you should use something like this:

    shell> mysqld_safe --key_buffer_size=64M --table_cache=256 \
               --sort_buffer_size=4M --read_buffer_size=1M &
    
  • If you have only 128MB 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=16M --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_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, you should 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 4.2.3.3.2, “Preconfigured Option Files”.

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=32M --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 tuning the InnoDB storage engine, see Section 13.2.13.1, “InnoDB Performance Tuning Tips”.

MySQL Enterprise For expert advice on tuning system parameters subscribe to the MySQL Enterprise Monitor. For more information, see http://www.mysql.com/products/enterprise/advisors.html.

7.5.4. Controlling Query Optimizer Performance

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 7–10) 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.

MySQL 5.0.1 introduces a more flexible method for query optimization that allows 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 via 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.

7.5.5. 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 is extremely 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.

Note

The query cache is not used for prepared statements. If you are using prepared statements, consider that these statements will not be satisfied by the query cache. See Section 20.9.4, “C API Prepared Statements”.

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. If you build MySQL from source, query cache capabilities can be excluded from the server entirely by invoking configure with the --without-query-cache option.

The query cache offers the potential for substantial performance improvement, but you should 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 7.5.5.2, “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.

7.5.5.1. How the Query Cache Operates

This section describes how the query cache works when it is operational. Section 7.5.5.3, “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.

Because comparison of a query against those in the cache occurs before parsing, the cache is not used for queries of the following types:

  • Prepared statements

  • Queries that are a subquery of an outer query

  • Queries executed within the body of a stored function or trigger

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 7.5.5.4, “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, ALTER TABLE, DROP TABLE, or DROP DATABASE.

The query cache also works within transactions when using InnoDB tables.

In MySQL 5.0, the result from a SELECT query on a view is cached.

Before MySQL 5.0, a query that began with a leading comment could be cached, but could not be fetched from the cache. This problem is fixed in MySQL 5.0.

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.

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 or INFORMATION_SCHEMA system database.

  • 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 MyODBC section of Chapter 20, Connectors and APIs.

  • It was issued as a prepared statement, even if no placeholders were employed. For example, the query used here is not cached:

    char *my_sql_stmt = "SELECT a, b FROM table_c";
    /* ... */
    mysql_stmt_prepare(stmt, my_sql_stmt, strlen(my_sql_stmt));
    

    See Section 20.9.4, “C API Prepared Statements”.

  • 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.

7.5.5.2. 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 query result is not cached.

Examples:

SELECT SQL_CACHE id, name FROM customer;
SELECT SQL_NO_CACHE id, name FROM customer;

7.5.5.3. 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 5.1.3, “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. The default size is 0, so the query cache is disabled by default.

MySQL Enterprise For expert advice on configuring the query cache subscribe to the MySQL Enterprise Monitor. For more information, see http://www.mysql.com/products/enterprise/advisors.html.

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. You should 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 allows 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.

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 allowed.

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.

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, you should 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).

MySQL Enterprise If the query cache is under-utilized, performance will suffer. Advice on avoiding this problem is provided to subscribers to the MySQL Enterprise Monitor. For more information, see http://www.mysql.com/products/enterprise/advisors.html.

7.5.5.4. Query Cache Status and Maintenance

You can check whether the query cache is present in your MySQL server using 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 7.5.5.3, “Query Cache Configuration”.

7.5.6. 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:

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 and State indicate what the thread is doing.

    Most states correspond to very quick operations. If a thread stays in a given state for many seconds, there might be a problem that needs to be investigated.

  • Time indicates how long the thread has been in its current state. The thread's notion of the current time may be altered in some cases: The thread can change the time with SET TIMESTAMP = value. For a thread running on a slave that is processing events from the master, the thread time is set to the time found in the events and thus reflects current time on the master and not the slave.

  • Info contains the text of the statement being executed by the thread, or NULL if it is not executing one. By default, this value contains only the first 100 characters of the statement. To see the complete statements, use SHOW FULL PROCESSLIST.

The following sections list the possible Command values, and State values grouped by category. The meaning for some of these values is self-evident. For others, additional description is provided.

7.5.6.1. Thread Command Values

A thread can have any of the following Command values:

  • Binlog Dump

    This is a thread on a master server for sending binary log contents to a slave server.

  • Change user

    The thread is executing a change-user operation.

  • Close stmt

    The thread is closing a prepared statement.

  • Connect

    A replication slave is connected to its master.

  • Connect Out

    A replication slave is connecting to its master.

  • Create DB

    The thread is executing a create-database operation.

  • Daemon

    This thread is internal to the server, not a thread that services a client connection.

  • Debug

    The thread is generating debugging information.

  • Delayed insert

    The thread is a delayed-insert handler.

  • Drop DB

    The thread is executing a drop-database operation.

  • Error

  • Execute

    The thread is executing a prepared statement.

  • Fetch

    The thread is fetching the results from executing a prepared statement.

  • Field List

    The thread is retrieving information for table columns.

  • Init DB

    The thread is selecting a default database.

  • Kill

    The thread is killing another thread.

  • Long Data

    The thread is retrieving long data in the result of executing a prepared statement.

  • Ping

    The thread is handling a server-ping request.

  • Prepare

    The thread is preparing a prepared statement.

  • Processlist

    The thread is producing information about server threads.

  • Query

    The thread is executing a statement.

  • Quit

    The thread is terminating.

  • Refresh

    The thread is flushing table, logs, or caches, or resetting status variable or replication server information.

  • Register Slave

    The thread is registering a slave server.

  • Reset stmt

    The thread is resetting a prepared statement.

  • Set option

    The thread is setting or resetting a client statement-execution option.

  • Shutdown

    The thread is shutting down the server.

  • Sleep

    The thread is waiting for the client to send a new statement to it.

  • Statistics

    The thread is producing server-status information.

  • Table Dump

    The thread is sending table contents to a slave server.

  • Time

    Unused.

7.5.6.2. General Thread States

The following list describes thread State values that are associated with general query processing and not more specialized activities such as replication. Many of these are useful only for finding bugs in the server.

  • After create

    Occurs when the thread creates a table (including internal temporary tables), at the end of the function that creates the table. This state is used even if the table could not be created due to some error.

  • Analyzing

    The thread is calculating a MyISAM table key distributions (for example, for ANALYZE TABLE).

  • checking permissions

    The thread is checking whether the server has the required privileges to execute the statement.

  • Checking table

    The thread is performing a table check operation.

  • cleaning up

    The thread has processed one command and is preparing to free memory and reset certain state variables.

  • closing tables

    Means that the thread is flushing the changed table data to disk and closing the used tables. This should be a fast operation. If not, you should verify that you do not have a full disk and that the disk is not in very heavy use.

  • converting HEAP to MyISAM

    The thread is converting an internal temporary table from a MEMORY table to an on-disk MyISAM table.

  • copy to tmp table

    The thread is processing an ALTER TABLE statement. This state occurs after the table with the new structure has been created but before rows are copied into it.

  • Copying to group table

    If a statement has different ORDER BY and GROUP BY criteria, the rows are sorted by group and copied to a temporary table.

  • Copying to tmp table

    The server is copying to a temporary table in memory.

  • Copying to tmp table on disk

    The server is copying to a temporary table on disk. The temporary result set was larger than tmp_table_size and the thread is changing the temporary table from in-memory to disk-based format to save memory.

  • Creating index

    The thread is processing ALTER TABLE ... ENABLE KEYS for a MyISAM table.

  • Creating sort index

    The thread is processing a SELECT that is resolved using an internal temporary table.

  • creating table

    The thread is creating a table. This includes creation of temporary tables.

  • Creating tmp table

    The thread is creating a temporary table in memory or on disk. If the table is created in memory but later is converted to an on-disk table, the state during that operation will be Copying to tmp table on disk.

  • deleting from main table

    The server is executing the first part of a multiple-table delete. It is deleting only from the first table, and saving columns and offsets to be used for deleting from the other (reference) tables.

  • deleting from reference tables

    The server is executing the second part of a multiple-table delete and deleting the matched rows from the other tables.

  • discard_or_import_tablespace

    The thread is processing an ALTER TABLE ... DISCARD TABLESPACE or ALTER TABLE ... IMPORT TABLESPACE statement.

  • end

    This occurs at the end but before the cleanup of ALTER TABLE, CREATE VIEW, DELETE, INSERT, SELECT, or UPDATE statements.

  • executing

    The thread has begun executing a statement.

  • Execution of init_command

    The thread is executing statements in the value of the init_command system variable.

  • freeing items

    The thread has executed a command. This state is usually followed by cleaning up.

  • Flushing tables

    The thread is executing FLUSH TABLES and is waiting for all threads to close their tables.

  • FULLTEXT initialization

    The server is preparing to perform a natural-language full-text search.

  • init

    This occurs before the initialization of ALTER TABLE, DELETE, INSERT, SELECT, or UPDATE statements.

  • Killed

    Someone has sent a KILL statement to the thread and it should abort next time it checks the kill flag. The flag is checked in each major loop in MySQL, but in some cases it might still take a short time for the thread to die. If the thread is locked by some other thread, the kill takes effect as soon as the other thread releases its lock.

  • Locked

    The query is locked by another query.

  • logging slow query

    The thread is writing a statement to the slow-query log.

  • NULL

    This state is used for the SHOW PROCESSLIST state.

  • login

    The initial state for a connection thread until the client has been authenticated successfully.

  • Opening tables, Opening table

    The thread is trying to open a table. This is should be very fast procedure, unless something prevents opening. For example, an ALTER TABLE or a LOCK TABLE statement can prevent opening a table until the statement is finished.

  • preparing

    This state occurs during query optimization.

  • Purging old relay logs

    The thread is removing unneeded relay log files.

  • query end

    This state occurs after processing a query but before the freeing items state.

  • Reading from net

    The server is reading a packet from the network.

  • Removing duplicates

    The query was using SELECT DISTINCT in such a way that MySQL could not optimize away the distinct operation at an early stage. Because of this, MySQL requires an extra stage to remove all duplicated rows before sending the result to the client.

  • removing tmp table

    The thread is removing an internal temporary table after processing a SELECT statement. This state is not used if no temporary table was created.

  • rename

    The thread is renaming a table.

  • rename result table

    The thread is processing an ALTER TABLE statement, has created the new table, and is renaming it to replace the original table.

  • Reopen tables

    The thread got a lock for the table, but noticed after getting the lock that the underlying table structure changed. It has freed the lock, closed the table, and is trying to reopen it.

  • Repair by sorting

    The repair code is using a sort to create indexes.

  • Repair done

    The thread has completed a multi-threaded repair for a MyISAM table.

  • Repair with keycache

    The repair code is using creating keys one by one through the key cache. This is much slower than Repair by sorting.

  • Rolling back

    The thread is rolling back a transaction.

  • Saving state

    For MyISAM table operations such as repair or analysis, the thread is saving the new table state to the .MYI file header. State includes information such as number of rows, the AUTO_INCREMENT counter, and key distributions.

  • Searching rows for update

    The thread is doing a first phase to find all matching rows before updating them. This has to be done if the UPDATE is changing the index that is used to find the involved rows.

  • Sending data

    The thread is processing rows for a SELECT statement and also is sending data to the client.

  • setup

    The thread is beginning an ALTER TABLE operation.

  • Sorting for group

    The thread is doing a sort to satisfy a GROUP BY.

  • Sorting for order

    The thread is doing a sort to satisfy a ORDER BY.

  • Sorting index

    The thread is sorting index pages for more efficient access during a MyISAM table optimization operation.

  • Sorting result

    For a SELECT statement, this is similar to Creating sort index, but for nontemporary tables.

  • statistics

    The server is calculating statistics to develop a query execution plan.

  • System lock

    The thread is going to request or is waiting for an internal or external system lock for the table. If this state is being caused by requests for external locks and you are not using multiple mysqld servers that are accessing the same tables, you can disable external system locks with the --skip-external-locking option. However, external locking is disabled by default, so it is likely that this option will have no effect. For SHOW PROFILE, this state means the thread is requesting the lock (not waiting for it).

  • Table lock

    The next thread state after System lock. The thread has acquired an external lock and is going to request an internal table lock.

  • Updating

    The thread is searching for rows to update and is updating them.

  • updating main table

    The server is executing the first part of a multiple-table update. It is updating only the first table, and saving columns and offsets to be used for updating the other (reference) tables.

  • updating reference tables

    The server is executing the second part of a multiple-table update and updating the matched rows from the other tables.

  • User lock

    The thread is going to request or is waiting for an advisory lock requested with a GET_LOCK() call. For SHOW PROFILE, this state means the thread is requesting the lock (not waiting for it).

  • Waiting for tables, Waiting for table

    The thread got a notification that the underlying structure for a table has changed and it needs to reopen the table to get the new structure. However, to reopen the table, it must wait until all other threads have closed the table in question.

    This notification takes place if another thread has used FLUSH TABLES or one of the following statements on the table in question: FLUSH TABLES tbl_name, ALTER TABLE, RENAME TABLE, REPAIR TABLE, ANALYZE TABLE, or OPTIMIZE TABLE.

  • Waiting on cond

    A generic state in which the thread is waiting for a condition to become true. No specific state information is available.

  • Writing to net

    The server is writing a packet to the network.

7.5.6.3. Delayed-Insert Thread States

These thread states are associated with processing for DELAYED inserts (see Section 12.2.5.2, “INSERT DELAYED Syntax”). Some states are associated with connection threads that process INSERT DELAYED statements from clients. Other states are associated with delayed-insert handler threads that insert the rows. There is a delayed-insert handler thread for each table for which INSERT DELAYED statements are issued.

States associated with a connection thread that processes an INSERT DELAYED statement from the client:

  • allocating local table

    The thread is preparing to feed rows to the delayed-insert handler thread.

  • Creating delayed handler

    The thread is creating a handler for DELAYED inserts.

  • got handler lock

    This occurs before the allocating local table state and after the waiting for handler lock state, when the connection thread gets access to the delayed-insert handler thread.

  • got old table

    This occurs after the waiting for handler open state. The delayed-insert handler thread has signaled that it has ended its initialization phase, which includes opening the table for delayed inserts.

  • storing row into queue

    The thread is adding a new row to the list of rows that the delayed-insert handler thread must insert.

  • update

  • waiting for delay_list

    This occurs during the initialization phase when the thread is trying to find the delayed-insert handler thread for the table, and before attempting to gain access to the list of delayed-insert threads.

  • waiting for handler insert

    An INSERT DELAYED handler has processed all pending inserts and is waiting for new ones.

  • waiting for handler lock

    This occurs before the allocating local table state when the connection thread waits for access to the delayed-insert handler thread.

  • waiting for handler open

    This occurs after the Creating delayed handler state and before the got old table state. The delayed-insert handler thread has just been started, and the connection thread is waiting for it to initialize.

States associated with a delayed-insert handler thread that inserts the rows:

  • insert

    The state that occurs just before inserting rows into the table.

  • reschedule

    After inserting a number of rows, the delayed-insert thread sleeps to let other threads do work.

  • upgrading lock

    A delayed-insert handler is trying to get a lock for the table to insert rows.

  • Waiting for INSERT

    A delayed-insert handler is waiting for a connection thread to add rows to the queue (see storing row into queue).

7.5.6.4. Replication Master Thread States

The following list shows the most common states you may see in the State column for the master's Binlog Dump thread. If you see no Binlog Dump threads on a master server, this means that replication is not running — that is, that no slaves are currently connected.

  • Sending binlog event to slave

    Binary logs consist of events, where an event is usually an update plus some other information. The thread has read an event from the binary log and is now sending it to the slave.

  • Finished reading one binlog; switching to next binlog

    The thread has finished reading a binary log file and is opening the next one to send to the slave.

  • Has sent all binlog to slave; waiting for binlog to be updated

    The thread has read all outstanding updates from the binary logs and sent them to the slave. The thread is now idle, waiting for new events to appear in the binary log resulting from new updates occurring on the master.

  • Waiting to finalize termination

    A very brief state that occurs as the thread is stopping.

7.5.6.5. Replication Slave I/O Thread States

The following list shows the most common states you see in the State column for a slave server I/O thread. This state also appears in the Slave_IO_State column displayed by SHOW SLAVE STATUS, so you can get a good view of what is happening by using that statement.

  • Waiting for master update

    The initial state before Connecting to master.

  • Connecting to master

    The thread is attempting to connect to the master.

  • Checking master version

    A state that occurs very briefly, after the connection to the master is established.

  • Registering slave on master

    A state that occurs very briefly after the connection to the master is established.

  • Requesting binlog dump

    A state that occurs very briefly, after the connection to the master is established. The thread sends to the master a request for the contents of its binary logs, starting from the requested binary log file name and position.

  • Waiting to reconnect after a failed binlog dump request

    If the binary log dump request failed (due to disconnection), the thread goes into this state while it sleeps, then tries to reconnect periodically. The interval between retries can be specified using the CHANGE MASTER TO statement or the --master-connect-retry option.

  • Reconnecting after a failed binlog dump request

    The thread is trying to reconnect to the master.

  • Waiting for master to send event

    The thread has connected to the master and is waiting for binary log events to arrive. This can last for a long time if the master is idle. If the wait lasts for slave_net_timeout seconds, a timeout occurs. At that point, the thread considers the connection to be broken and makes an attempt to reconnect.

  • Queueing master event to the relay log

    The thread has read an event and is copying it to the relay log so that the SQL thread can process it.

  • Waiting to reconnect after a failed master event read

    An error occurred while reading (due to disconnection). The thread is sleeping for the number of seconds set by the CHANGE MASTER TO statement or --master-connect-retry option (default 60) before attempting to reconnect.

  • Reconnecting after a failed master event read

    The thread is trying to reconnect to the master. When connection is established again, the state becomes Waiting for master to send event.

  • Waiting for the slave SQL thread to free enough relay log space

    You are using a nonzero relay_log_space_limit value, and the relay logs have grown large enough that their combined size exceeds this value. The I/O thread is waiting until the SQL thread frees enough space by processing relay log contents so that it can delete some relay log files.

  • Waiting for slave mutex on exit

    A state that occurs briefly as the thread is stopping.

7.5.6.6. Replication Slave SQL Thread States

The following list shows the most common states you may see in the State column for a slave server SQL thread:

  • Waiting for the next event in relay log

    The initial state before Reading event from the relay log.

  • Reading event from the relay log

    The thread has read an event from the relay log so that the event can be processed.

  • Has read all relay log; waiting for the slave I/O thread to update it

    The thread has processed all events in the relay log files, and is now waiting for the I/O thread to write new events to the relay log.

  • Making temp file

    The thread is executing a LOAD DATA INFILE statement and is creating a temporary file containing the data from which the slave will read rows.

  • Waiting for slave mutex on exit

    A very brief state that occurs as the thread is stopping.

The State column for the I/O thread may also show the text of a statement. This indicates that the thread has read an event from the relay log, extracted the statement from it, and is executing it.

7.5.6.7. Replication Slave Connection Thread States

These thread states occur on a replication slave but are associated with connection threads, not with the I/O or SQL threads.

  • Changing master

    The thread is processing a CHANGE MASTER TO statement.

  • Creating table from master dump

    The slave is creating a table using the CREATE TABLE statement contained in the dump from the master. Used for LOAD TABLE FROM MASTER and LOAD DATA FROM MASTER.

  • Killing slave

    The thread is processing a SLAVE STOP statement.

  • Opening master dump table

    This state occurs after Creating table from master dump.

  • Reading master dump table data

    This state occurs after Opening master dump table.

  • Rebuilding the index on master dump table

    This state occurs after Reading master dump table data.

  • starting slave

    The thread is starting the slave threads after processing a successful LOAD DATA FROM MASTER load operation.

7.5.6.8. MySQL Cluster Thread States

  • Committing events to binlog

  • Opening mysql.ndb_apply_status

  • Processing events

    The thread is processing events for binary logging.

  • Processing events from schema table

    The thread is doing the work of schema replication.

  • Shutting down

  • Syncing ndb table schema operation and binlog

    This is used to have a correct binary log of schema operations for NDB.

  • Waiting for event from ndbcluster

    The server is acting as an SQL node in a MySQL Cluster, and is connected to a cluster management node.

  • Waiting for first event from ndbcluster

  • Waiting for ndbcluster binlog update to reach current position

  • Waiting for ndbcluster to start

  • Waiting for schema epoch

    The thread is waiting for a schema epoch (that is, a global checkpoint).

7.5.7. 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 5.1.3, “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 allow 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.

7.5.8. How MySQL Uses Memory

The following list indicates some of the ways that the mysqld server uses memory. Where applicable, the name of the system variable relevant to the memory use is given:

  • The key buffer is shared by all threads; its size is determined by the key_buffer_size variable. Other buffers used by the server are allocated as needed. See Section 7.5.3, “Tuning Server Parameters”.

  • 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 both begin with a size given by net_buffer_length but are dynamically enlarged up to max_allowed_packet bytes as needed. The result buffer shrinks to net_buffer_length after each SQL statement. While a statement is running, a copy of the current statement string is also allocated.

  • 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.

  • Only compressed MyISAM tables are memory mapped. This is because the 32-bit memory space of 4GB is not large enough for most big tables. When systems with a 64-bit address space become more common, we may add general support for memory mapping.

  • 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 in order 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 heap table exceeds the size of tmp_table_size, MySQL handles this automatically by changing the in-memory heap table to a disk-based MyISAM table as necessary. You can also increase the temporary table size by setting the tmp_table_size option to mysqld, or by setting the SQL option sql_big_tables in the client program. See Section 5.1.4, “Session System Variables”.

    MySQL Enterprise Subscribers to the MySQL Enterprise Monitor are alerted when temporary tables exceed tmp_table_size. Advisors make recommendations for the optimum value of tmp_table_size based on actual table usage. For more information about the MySQL Enterprise Monitor please see http://www.mysql.com/products/enterprise/advisors.html.

  • 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.1.4.4, “Where MySQL Stores Temporary Files”.

  • Almost all parsing and calculating is done in a local memory store. 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. This is done with malloc() and free().

  • 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.

  • 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.

  • Handler structures for all in-use tables are saved in a cache and managed as a FIFO. By default, the cache has 64 entries. If a table has been used by two running threads at the same time, the cache contains two entries for the table. See Section 7.4.7, “How MySQL Opens and Closes Tables”.

  • 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 and CREATE USER statements. This memory is not released by the corresponding REVOKE and DROP USER 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. You can verify this by checking 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.

7.5.9. 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.

Currently, MySQL supports only 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 allowed 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 allowed 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

You can verify that large pages are being used by checking /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

7.5.10. How MySQL Uses Internal Temporary Tables

In some cases, the server creates internal temporary tables while processing queries. A temporary table can be held in memory and processed by the MEMORY storage engine, or stored on disk and processed by the MyISAM storage engine. Temporary tables can be created under conditions such as these:

  • 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.

  • DISTINCT combined with ORDER BY may require a temporary table.

You can tell whether a query requires a temporary table by using EXPLAIN and checking the Extra column to see whether it says Using temporary. See Section 7.2.1, “Optimizing Queries with EXPLAIN.

Some conditions prevent the use of a MEMORY temporary table, in which case the server uses a MyISAM table instead:

  • Presence of a TEXT or BLOB column in the table

  • Presence of any column in a GROUP BY or DISTINCT clause larger than 512 bytes

  • Presence of any column larger than 512 bytes in the SELECT list, if UNION or UNION ALL is used.

A temporary table that is created initially as a MEMORY table might be converted to a MyISAM table and stored on disk if it becomes too large. The max_heap_table_size system variable determines how large MEMORY tables are allowed to grow. It applies to all MEMORY tables, including those created with CREATE TABLE. However, for internal MEMORY tables, the actual maximum size is determined by max_heap_table_size in combination with tmp_table_size: Whichever value is smaller is the one that applies. If the size of an internal MEMORY table exceeds the limit, MySQL automatically converts it to an on-disk MyISAM table.

7.5.11. How MySQL Uses DNS

When a new client connects to mysqld, mysqld spawns a new thread to handle the request. This thread first checks whether the host name is in the host name cache. If not, the thread attempts to resolve the host name:

  • The thread takes the IP address and resolves it to a host name (using gethostbyaddr()). It then takes that host name and resolves it back to the IP address (using gethostbyname()) and compares to ensure it is the original IP address.

  • If the operating system supports the thread-safe gethostbyaddr_r() and gethostbyname_r() calls, the thread uses them to perform host name resolution.

  • If the operating system does not support the thread-safe calls, the thread 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 name cache until the first thread unlocks the mutex.

You can disable DNS host name lookups by starting mysqld with the --skip-name-resolve option. However, in this case, you can use only IP numbers in the MySQL grant tables.

If you have a very slow DNS and many hosts, you can get more performance by either disabling DNS lookups with --skip-name-resolve or by increasing the HOST_CACHE_SIZE define (default value: 128) and recompiling mysqld.

You can disable the host name cache by starting the server with the --skip-host-cache option. To clear the host name cache, issue a FLUSH HOSTS statement or execute the mysqladmin flush-hosts command.

To disallow TCP/IP connections entirely, start mysqld with the --skip-networking option.

7.6. Disk Issues

  • 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 7.6.1, “Using Symbolic Links”.

    • 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 7.1.5, “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 more 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
    

    Note that 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 more performance without sacrificing too much reliability. (This flag is on by default on Linux.)

7.6.1. Using Symbolic Links

You can move tables and databases 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 disk.

The recommended way to do this is simply to symlink databases to a different disk. Symlink tables only as a last resort.

7.6.1.1. 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 symlink 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.

However, if you really need to do this, it is possible by altering the source file mysys/my_symlink.c, in which you should look for the following statement:

if (!(MyFlags & MY_RESOLVE_LINK) ||
    (!lstat(filename,&stat_buff) && S_ISLNK(stat_buff.st_mode)))

Change the statement to this:

if (1)

7.6.1.2. Using Symbolic Links for Tables on Unix

You should not symlink tables on systems that do not have a fully operational realpath() call. (Linux and Solaris support realpath()). You can check whether your system supports symbolic links by issuing a SHOW VARIABLES LIKE 'have_symlink' statement.

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.

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.

  • You can instruct a running MySQL server to perform the symlinking by using the DATA DIRECTORY and INDEX DIRECTORY options to CREATE TABLE. See Section 12.1.10, “CREATE TABLE Syntax”. Alternatively, symlinking can be accomplished manually from the command line using ln -s if mysqld is not running.

    Note

    Beginning with MySQL 5.0.60, 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 why you should not run mysqld as the system root or allow 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, you should use the --skip-symbolic-links option to mysqld to ensure that no one can use mysqld to drop or rename a file outside of the data directory.

Table symlink operations that are not yet supported:

  • ALTER TABLE ignores the DATA DIRECTORY and INDEX DIRECTORY table options.

  • BACKUP TABLE and RESTORE TABLE do not respect symbolic links.

  • The .frm file must never be a symbolic link (as indicated previously, only the data and index files can be symbolic links). Attempting to do this (for example, to make synonyms) 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.

7.6.1.3. Using Symbolic Links for Databases on Windows

Symbolic links are enabled by default for all Windows servers. This enables you to put a database directory on a different disk by setting up a symbolic link to it. This is similar to the way that database symbolic links work on Unix, although the procedure for setting up the link is different. If you do not need symbolic links, you can disable them using the --skip-symbolic-links option.

On Windows, create a symbolic link to a MySQL database by creating a file in the data directory that contains the path to the destination directory. The file should be named db_name.sym, where db_name is the database name.

Suppose that the MySQL data directory is C:\mysql\data and you want to have database foo located at D:\data\foo. Set up a symlink using this procedure:

  1. Make sure that the D:\data\foo directory exists by creating it if necessary. If you already have a database directory named foo in the data directory, you should 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.

  2. Create a text file C:\mysql\data\foo.sym that contains the path name D:\data\foo\.

    Note

    The path name to the new database and tables should be absolute. If you specify a relative path, the location will be relative to the foo.sym file.

After this, all tables created in the database foo are created in D:\data\foo.

The following limitations apply to the use of .sym files for database symbolic linking on Windows:

  • The symbolic link is not used if a directory with the same name as the database exists in the MySQL data directory.

  • The --innodb_file_per_table option cannot be used.

  • If you run mysqld as a service, you cannot use a mapped drive to a remote server as the destination of the symbolic link. As a workaround, you can use the full path (\\servername\path\).