# Performance Tuning
# Add the correct index
This is a huge topic, but it is also the most important "performance" issue.
The main lesson for a novice is to learn of "composite" indexes. Here's a quick example:
is excellent for these:
WHERE last_name = '...' WHERE first_name = '...' AND last_name = '...' -- (order in WHERE does not matter)
but not for
WHERE first_name = '...' -- order in INDEX _does_ matter WHERE last_name = '...' OR first_name = '...' -- "OR" is a killer
# Don't hide in function
A common mistake is to hide an indexed column inside a function call. For example, this can't be helped by an index:
WHERE DATE(dt) = '2000-01-01'
INDEX(dt) then these may use the index:
WHERE dt = '2000-01-01' -- if `dt` is datatype `DATE`
This works for
TIMESTAMP, and even
WHERE dt >= '2000-01-01' AND dt < '2000-01-01' + INTERVAL 1 DAY
OR kills optimization.
WHERE a = 12 OR b = 78
INDEX(a,b), and may or may not use
INDEX(a), INDEX(b) via "index merge". Index merge is better than nothing, but only barely.
WHERE x = 3 OR x = 5
is turned into
WHERE x IN (3, 5)
which may use an index with
x in it.
# Set the cache correctly
innodb_buffer_pool_size should be about 70% of available RAM.
Here are some things that are not likely to help performance. They stem from out-of-date information and/or naivety.
- InnoDB has improved to the point where MyISAM is unlikely to be better.
PARTITIONingrarely provides performance benefits; it can even hurt performance.
query_cache_sizebigger than 100M will usually hurt performance.
- Increasing lots of values in
my.cnfmay lead to 'swapping', which is a serious performance problem.
- "Prefix indexes" (such as
INDEX(foo(20))) are generally useless.
OPTIMIZE TABLEis almost always useless. (And it involves locking the table.)
# Have an INDEX
The most important thing for speeding up a query on any non-tiny table is to have a suitable index.
WHERE a = 12 --> INDEX(a) WHERE a > 12 --> INDEX(a) WHERE a = 12 AND b > 78 --> INDEX(a,b) is more useful than INDEX(b,a) WHERE a > 12 AND b > 78 --> INDEX(a) or INDEX(b); no way to handle both ranges ORDER BY x --> INDEX(x) ORDER BY x, y --> INDEX(x,y) in that order ORDER BY x DESC, y ASC --> No index helps - because of mixing ASC and DESC
Subqueries come in several flavors, and they have different optimization potential. First, note that subqueries can be either "correlated" or "uncorrelated". Correlated means that they depend on some value from outside the subquery. This generally implies that the subquery must be re-evaluated for each outer value.
This correlated subquery is often pretty good. Note: It must return at most 1 value. It is often useful as an alternative to, though not necessarily faster than, a
SELECT a, b, ( SELECT ... FROM t WHERE t.x = u.x ) AS c FROM u ... SELECT a, b, ( SELECT MAX(x) ... ) AS c FROM u ... SELECT a, b, ( SELECT x FROM t ORDER BY ... LIMIT 1 ) AS c FROM u ...
This is usually uncorrelated:
SELECT ... FROM ( SELECT ... ) AS a JOIN b ON ...
Notes on the
- If it returns 1 row, great.
- A good paradigm (again "1 row") is for the subquery to be
( SELECT @n := 0 ), thereby initializing an `@variable for use in the rest or the query.
- If it returns many rows and the
( SELECT ... )with many rows, then efficiency can be terrible. Pre-5.6, there was no index, so it became a
CROSS JOIN; 5.6+ involves deducing the best index on the temp tables and then generating it, only to throw it away when finished with the
# JOIN + GROUP BY
A common problem that leads to an inefficient query goes something like this:
SELECT ... FROM a JOIN b ON ... WHERE ... GROUP BY a.id
JOIN expands the number of rows; then the
GROUP BY whittles it back down the the number of rows in
There may not be any good choices to solve this explode-implode problem. One possible option is to turn the
JOIN into a correlated subquery in the
SELECT. This also eliminates the
# Avoid inefficient constructs
x IN ( SELECT ... )
turn into a
When possible, avoid
Do not 'hide' an indexed column in a function, such as
WHERE DATE(x) = ...; reformulate as
WHERE x = ...
You can generally avoid
WHERE LCASE(name1) = LCASE(name2) by having a suitable collation.
Do no use
OFFSET for "pagination", instead 'remember where you left off'.
SELECT * ... (unless debugging).
Note to Maria Deleva, Barranka, Batsu: This is a place holder; please make remove these items as you build full-scale examples. After you have done the ones you can, I will move in to elaborate on the rest and/or toss them.
See also discussions about ORDER BY, LIKE, REGEXP, etc. Note: this needs editing with links and more Topics.