MySQL下的RAND()优化案例分析
众所周知,在mysql中,如果直接 order by rand() 的话,效率非常差,因为会多次执行。事实上,如果等值查询也是用 rand() 的话也如此,我们先来看看下面这几个sql的不同执行计划和执行耗时。
首先,看下建表ddl,这是一个没有显式自增主键的innodb表:
[yejr@imysql]> show create table t_innodb_random\g *************************** 1. row *************************** table: t_innodb_random create table: create table `t_innodb_random` ( `id` int(10) unsigned not null, `user` varchar(64) not null default '', key `idx_id` (`id`) ) engine=innodb default charset=latin1
往这个表里灌入一些测试数据,至少10万以上, id 字段也是乱序的。
[yejr@imysql]> select count(*) from t_innodb_random\g *************************** 1. row *************************** count(*): 393216
1、常量等值检索:
[yejr@imysql]> explain select id from t_innodb_random where id = 13412\g *************************** 1. row *************************** id: 1 select_type: simple table: t_innodb_random type: ref possible_keys: idx_id key: idx_id key_len: 4 ref: const rows: 1 extra: using index
[yejr@imysql]> select id from t_innodb_random where id = 13412; 1 row in set (0.00 sec)
可以看到执行计划很不错,是常量等值查询,速度非常快。
2、使用rand()函数乘以常量,求得随机数后检索:
[yejr@imysql]> explain select id from t_innodb_random where id = round(rand()*13241324)\g *************************** 1. row *************************** id: 1 select_type: simple table: t_innodb_random type: index possible_keys: null key: idx_id key_len: 4 ref: null rows: 393345 extra: using where; using index
[yejr@imysql]> select id from t_innodb_random where id = round(rand()*13241324)\g empty set (0.26 sec)
可以看到执行计划很糟糕,虽然是只扫描索引,但是做了全索引扫描,效率非常差。因为where条件中包含了rand(),使得mysql把它当做变量来处理,无法用常量等值的方式查询,效率很低。
我们把常量改成取t_innodb_random表的最大id值,再乘以rand()求得随机数后检索看看什么情况:
[yejr@imysql]> explain select id from t_innodb_random where id = round(rand()*(select max(id) from t_innodb_random))\g *************************** 1. row *************************** id: 1 select_type: primary table: t_innodb_random type: index possible_keys: null key: idx_id key_len: 4 ref: null rows: 393345 extra: using where; using index *************************** 2. row *************************** id: 2 select_type: subquery table: null type: null possible_keys: null key: null key_len: null ref: null rows: null extra: select tables optimized away
[yejr@imysql]> select id from t_innodb_random where id = round(rand()*(select max(id) from t_innodb_random))\g empty set (0.27 sec)
可以看到,执行计划依然是全索引扫描,执行耗时也基本相当。
3、改造成普通子查询模式 ,这里有两次子查询
[yejr@imysql]> explain select id from t_innodb_random where id = (select round(rand()*(select max(id) from t_innodb_random)) as nid)\g *************************** 1. row *************************** id: 1 select_type: primary table: t_innodb_random type: index possible_keys: null key: idx_id key_len: 4 ref: null rows: 393345 extra: using where; using index *************************** 2. row *************************** id: 3 select_type: subquery table: null type: null possible_keys: null key: null key_len: null ref: null rows: null extra: select tables optimized away
[yejr@imysql]> select id from t_innodb_random where id = (select round(rand()*(select max(id) from t_innodb_random)) as nid)\g empty set (0.27 sec)
可以看到,执行计划也不好,执行耗时较慢。
4、改造成join关联查询,不过最大值还是用常量表示
[yejr@imysql]> explain select id from t_innodb_random t1 join (select round(rand()*13241324) as id2) as t2 where t1.id = t2.id2\g *************************** 1. row *************************** id: 1 select_type: primary table: <derived2> type: system possible_keys: null key: null key_len: null ref: null rows: 1 extra: *************************** 2. row *************************** id: 1 select_type: primary table: t1 type: ref possible_keys: idx_id key: idx_id key_len: 4 ref: const rows: 1 extra: using where; using index *************************** 3. row *************************** id: 2 select_type: derived table: null type: null possible_keys: null key: null key_len: null ref: null rows: null extra: no tables used
[yejr@imysql]> select id from t_innodb_random t1 join (select round(rand()*13241324) as id2) as t2 where t1.id = t2.id2\g empty set (0.00 sec)
这时候执行计划就非常完美了,和最开始的常量等值查询是一样的了,执行耗时也非常之快。
这种方法虽然很好,但是有可能查询不到记录,改造范围查找,但结果limit 1就可以了:
[yejr@imysql]> explain select id from t_innodb_random where id > (select round(rand()*(select max(id) from t_innodb_random)) as nid) limit 1\g *************************** 1. row *************************** id: 1 select_type: primary table: t_innodb_random type: index possible_keys: null key: idx_id key_len: 4 ref: null rows: 393345 extra: using where; using index *************************** 2. row *************************** id: 3 select_type: subquery table: null type: null possible_keys: null key: null key_len: null ref: null rows: null extra: select tables optimized away
[yejr@imysql]> select id from t_innodb_random where id > (select round(rand()*(select max(id) from t_innodb_random)) as nid) limit 1\g *************************** 1. row *************************** id: 1301 1 row in set (0.00 sec)
可以看到,虽然执行计划也是全索引扫描,但是因为有了limit 1,只需要找到一条记录,即可终止扫描,所以效率还是很快的。
小结:
从数据库中随机取一条记录时,可以把rand()生成随机数放在join子查询中以提高效率。
5、再来看看用ordrr by rand()方式一次取得多个随机值的方式:
[yejr@imysql]> explain select id from t_innodb_random order by rand() limit 1000\g *************************** 1. row *************************** id: 1 select_type: simple table: t_innodb_random type: index possible_keys: null key: idx_id key_len: 4 ref: null rows: 393345 extra: using index; using temporary; using filesort
[yejr@imysql]> select id from t_innodb_random order by rand() limit 1000; 1000 rows in set (0.41 sec)
全索引扫描,生成排序临时表,太差太慢了。
6、把随机数放在子查询里看看:
[yejr@imysql]> explain select id from t_innodb_random where id > (select rand() * (select max(id) from t_innodb_random) as nid) limit 1000\g *************************** 1. row *************************** id: 1 select_type: primary table: t_innodb_random type: index possible_keys: null key: idx_id key_len: 4 ref: null rows: 393345 extra: using where; using index *************************** 2. row *************************** id: 3 select_type: subquery table: null type: null possible_keys: null key: null key_len: null ref: null rows: null extra: select tables optimized away
[yejr@imysql]> select id from t_innodb_random where id > (select rand() * (select max(id) from t_innodb_random) as nid) limit 1000\g 1000 rows in set (0.04 sec)
嗯,提速了不少,这个看起来还不赖:)
7、仿照上面的方法,改成join和随机数子查询关联
[yejr@imysql]> explain select id from t_innodb_random t1 join (select rand() * (select max(id) from t_innodb_random) as nid) t2 on t1.id > t2.nid limit 1000\g *************************** 1. row *************************** id: 1 select_type: primary table: <derived2> type: system possible_keys: null key: null key_len: null ref: null rows: 1 extra: *************************** 2. row *************************** id: 1 select_type: primary table: t1 type: range possible_keys: idx_id key: idx_id key_len: 4 ref: null rows: 196672 extra: using where; using index *************************** 3. row *************************** id: 2 select_type: derived table: null type: null possible_keys: null key: null key_len: null ref: null rows: null extra: no tables used *************************** 4. row *************************** id: 3 select_type: subquery table: null type: null possible_keys: null key: null key_len: null ref: null rows: null extra: select tables optimized away
[yejr@imysql]> select id from t_innodb_random t1 join (select rand() * (select max(id) from t_innodb_random) as nid) t2 on t1.id > t2.nid limit 1000\g 1000 rows in set (0.00 sec)
可以看到,全索引检索,发现符合记录的条件后,直接取得1000行,这个方法是最快的。
综上,想从mysql数据库中随机取一条或者n条记录时,最好把rand()生成随机数放在join子查询中以提高效率。
上面说了那么多的废话,最后简单说下,就是把下面这个sql:
select id from table order by rand() limit n;
改造成下面这个:
select id from table t1 join (select rand() * (select max(id) from table) as nid) t2 on t1.id > t2.nid limit n;
如果想要达到完全随机,还可以改成下面这种写法:
select id from table t1 join (select round(rand() * (select max(id) from table)) as nid from table limit n) t2 on t1.id = t2.nid;
就可以享受在sql中直接取得随机数了,不用再在程序中构造一串随机数去检索了。
下一篇: 分析一个MySQL的异常查询的案例