MySQL Performance: Difference between revisions
Created page with "== Connections == The MySQL protocol is half-duplex, which means that it can send or receive messages, but not both at the same time. == Query cache == Looking up an item i..." |
|||
(3 intermediate revisions by the same user not shown) | |||
Line 98: | Line 98: | ||
ALTER TABLE my_table ADD KEY (my_column) | ALTER TABLE my_table ADD KEY (my_column) | ||
== Database dumps == | |||
Fastest way to create a database dump is to pass <code>--quick</code> and <code>--single-transaction</code> to <code>mysqldump</code>. This will only work if all your tables are InnoDB (which they should be, unless you have very specific use cases). In theory you should also pass <code>--opt</code>, but in practice this is enabled by default. | |||
[[Category:MySQL]] |
Latest revision as of 10:56, 8 November 2021
Connections
The MySQL protocol is half-duplex, which means that it can send or receive messages, but not both at the same time.
Query cache
Looking up an item in the query cache is a hash operation.
Profiling queries
Enable the profiler for this session (note this method is deprecated in later versions of Percona):
SET PROFILING = 1
Run query:
SELECT * FROM table
Show the profiles (summary of query and time taken):
SHOW PROFILES
Show detailed information for a specific query:
SHOW PROFILE FOR QUERY 1
To clear out the query cache:
RESET QUERY CACHE
Using EXPLAIN
A simple test to see if indexes are being used:
EXPLAIN SELECT * FROM table WHERE column = ?
If the rows
value returned is significantly higher than the number of rows fetched by the query and the type
value is ALL
then it is likely that MySQL is performing a full table scan. Adding an index can significantly reduce the number of rows scanned (though not always down to 1:1 scanned:fetched).
If the extra
value is Using where
then this also suggests that MySQL is discarding rows after scanning them, which may have performance implications.
Locking
Locking has a trade-off effect on performance. On the one hand, a high-level lock (e.g. table lock) is cheap to obtain, but prevents reads and writes. A row-level lock is more expensive to obtain, but allows reads and writes to other rows.
In some circumstances, the locking strategy is fixed. For example, ALTER TABLE
commands will always require a table lock.
Transactions
Transactions help to ensure that a collection of statements are executed as one atomic commit. There is a cost to this, and it may not be required for all applications.
Storage engines
To find the storage engine for a given table:
SELECT engine FROM information_schema.tables WHERE table_schema = 'database' AND table_name = 'table'
Storage engines for all tables in a given database:
SELECT table_name, engine FROM information_schema.tables WHERE table_schema = 'database'
MyISAM
MyISAM supports compressed tables using myisampack
. Compressed tables use up less disk space and therefore require fewer seeks to load. They may be a good choice for read-only data, e.g. lookup tables.
The major downside to MyISAM is that it uses table locking, and therefore it is not suitable for data which is regularly updated (e.g. sessions, logs).
Converting between storage engines
The simplest and easiest way to convert a table to another storage engine is with the ALTER TABLE
command:
ALTER TABLE mytable ENGINE = 'InnoDB';
However, this performs a copy of each row in the table, and therefore is unlikely to be suitable for live data.
Data types
Some broad advice:
- Simpler data types are easier to process.
- Integers are cheaper to compare than strings.
- Use integers for IP addresses.
- Avoid NULL where possible as it is harder for MySQL to optimise queries which may involve NULL values.
Joins
MySQL has a limit of 61 tables per join.
Indexes
Adding a b-tree index to an InnoDB table:
ALTER TABLE my_table ADD KEY (my_column)
Database dumps
Fastest way to create a database dump is to pass --quick
and --single-transaction
to mysqldump
. This will only work if all your tables are InnoDB (which they should be, unless you have very specific use cases). In theory you should also pass --opt
, but in practice this is enabled by default.