MySQL之select in 子查询优化的实现

(编辑：jimmy 日期: 2024/11/25 浏览：3 次 )

下面的演示基于MySQL5.7.27版本

一、关于MySQL子查询的优化策略介绍:

子查询优化策略

对于不同类型的子查询，优化器会选择不同的策略。

1. 对于 IN、=ANY 子查询，优化器有如下策略选择：

semijoin
Materialization
exists

2. 对于 NOT IN、<>ALL 子查询，优化器有如下策略选择：

Materialization
exists

3. 对于 derived 派生表，优化器有如下策略选择：
derived_merge，将派生表合并到外部查询中（5.7 引入）；
将派生表物化为内部临时表，再用于外部查询。
注意：update 和 delete 语句中子查询不能使用 semijoin、materialization 优化策略

二、创建数据进行模拟演示

为了方便分析问题先建两张表并插入模拟数据：

CREATE TABLE `test02` (
 `id` int(11) NOT NULL,
 `a` int(11) DEFAULT NULL,
 `b` int(11) DEFAULT NULL,
 PRIMARY KEY (`id`),
 KEY `a` (`a`)
) ENGINE=InnoDB;

drop procedure idata;
delimiter ;;
create procedure idata()
begin
 declare i int;
 set i=1;
 while(i<=10000)do
  insert into test02 values(i, i, i);
  set i=i+1;
 end while;
end;;
delimiter ;
call idata();

create table test01 like test02;
insert into test01 (select * from test02 where id<=1000)

三、举例分析SQL实例

子查询示例：

SELECT * FROM test01 WHERE test01.a IN (SELECT test02.b FROM test02 WHERE id < 10)

大部分人可定会简单的认为这个 SQL 会这样执行：

SELECT test02.b FROM test02 WHERE id < 10

结果:1,2,3,4,5,6,7,8,9

SELECT * FROM test01 WHERE test01.a IN (1,2,3,4,5,6,7,8,9);

但实际上 MySQL 并不是这样做的。MySQL 会将相关的外层表压到子查询中，优化器认为这样效率更高。也就是说，优化器会将上面的 SQL 改写成这样：

select * from test01 where exists(select b from test02 where id < 10 and test01.a=test02.b);

提示：针对mysql5.5以及之前的版本

查看执行计划如下，发现这条SQL对表test01进行了全表扫描1000，效率低下：

root@localhost [dbtest01]>desc select * from test01 where exists(select b from test02 where id < 10 and test01.a=test02.b);
+----+--------------------+--------+------------+-------+---------------+---------+---------+------+--------+----------+-------------+
| id | select_type    | table | partitions | type | possible_keys | key   | key_len | ref | rows  | filtered | Extra    |
+----+--------------------+--------+------------+-------+---------------+---------+---------+------+--------+----------+-------------+
| 1 | PRIMARY      | test01 | NULL    | ALL  | NULL     | NULL  | NULL  | NULL | 1000  |  100.00 | Using where |
| 2 | DEPENDENT SUBQUERY | test02 | NULL    | range | PRIMARY    | PRIMARY | 4    | NULL |   9 |  10.00 | Using where |
+----+--------------------+--------+------------+-------+---------------+---------+---------+------+--------+----------+-------------+
2 rows in set, 2 warnings (0.00 sec)

但是此时实际执行下面的SQL，发现也不慢啊,这不是自相矛盾嘛，别急，咱们继续往下分析:

SELECT * FROM test01 WHERE test01.a IN (SELECT test02.b FROM test02 WHERE id < 10)

查看此条SQL的执行计划如下：

root@localhost [dbtest01]>desc SELECT * FROM test01 WHERE test01.a IN (SELECT test02.b FROM test02 WHERE id < 10);
+----+--------------+-------------+------------+-------+---------------+---------+---------+---------------+------+----------+-------------+
| id | select_type | table    | partitions | type | possible_keys | key   | key_len | ref      | rows | filtered | Extra    |
+----+--------------+-------------+------------+-------+---------------+---------+---------+---------------+------+----------+-------------+
| 1 | SIMPLE    | <subquery2> | NULL    | ALL  | NULL     | NULL  | NULL  | NULL     | NULL |  100.00 | Using where |
| 1 | SIMPLE    | test01   | NULL    | ref  | a       | a    | 5    | <subquery2>.b |  1 |  100.00 | NULL    |
| 2 | MATERIALIZED | test02   | NULL    | range | PRIMARY    | PRIMARY | 4    | NULL     |  9 |  100.00 | Using where |
+----+--------------+-------------+------------+-------+---------------+---------+---------+---------------+------+----------+-------------+
3 rows in set, 1 warning (0.00 sec)

发现优化器使用到了策略MATERIALIZED。于是对此策略进行了资料查询和学习。
https://dev.mysql.com/doc/refman/5.6/en/subquery-optimization.html

原因是从MySQL5.6版本之后包括MySQL5.6版本，优化器引入了新的优化策略:materialization=[off|on],semijoin=[off|on],(off代表关闭此策略，on代表开启此策略)
可以采用show variables like 'optimizer_switch'; 来查看MySQL采用的优化器策略。当然这些策略都是可以在线进行动态修改的
set global optimizer_switch='materialization=on,semijoin=on';代表开启优化策略materialization和semijoin

MySQL5.7.27默认的优化器策略：

root@localhost [dbtest01]>show variables like 'optimizer_switch';                                                               
+------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| Variable_name  | Value                                                                                                                                                                                                      |
+------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| optimizer_switch | index_merge=on,index_merge_union=on,index_merge_sort_union=on,index_merge_intersection=on,engine_condition_pushdown=on,index_condition_pushdown=on,mrr=on,mrr_cost_based=on,block_nested_loop=on,batched_key_access=off,materialization=on,semijoin=on,loosescan=on,firstmatch=on,duplicateweedout=on,subquery_materialization_cost_based=on,use_index_extensions=on,condition_fanout_filter=on,derived_merge=on |
+------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

所以在MySQL5.6及以上版本时

执行下面的SQL是不会慢的。因为MySQL的优化器策略materialization和semijoin 对此SQL进行了优化

SELECT * FROM test01 WHERE test01.a IN (SELECT test02.b FROM test02 WHERE id < 10)

然而咱们把mysql的优化器策略materialization和semijoin 关闭掉测试，发现SQL确实对test01进行了全表的扫描(1000):

set global optimizer_switch='materialization=off,semijoin=off';

执行计划如下test01表确实进行了全表扫描:

root@localhost [dbtest01]>desc SELECT * FROM test01 WHERE test01.a IN (SELECT test02.b FROM test02 WHERE id < 10);
+----+--------------------+--------+------------+-------+---------------+---------+---------+------+--------+----------+-------------+
| id | select_type    | table | partitions | type | possible_keys | key   | key_len | ref | rows  | filtered | Extra    |
+----+--------------------+--------+------------+-------+---------------+---------+---------+------+--------+----------+-------------+
| 1 | PRIMARY      | test01 | NULL    | ALL  | NULL     | NULL  | NULL  | NULL | 1000  |  100.00 | Using where |
| 2 | DEPENDENT SUBQUERY | test02 | NULL    | range | PRIMARY    | PRIMARY | 4    | NULL |   9 |  10.00 | Using where |
+----+--------------------+--------+------------+-------+---------------+---------+---------+------+--------+----------+-------------+
2 rows in set, 1 warning (0.00 sec)

下面咱们分析下这个执行计划:

！！！！再次提示:如果是mysql5.5以及之前的版本，或者是mysql5.6以及之后的版本关闭掉优化器策略materialization=off,semijoin=off，得到的SQL执行计划和下面的是相同的

root@localhost [dbtest01]>desc select * from test01 where exists(select b from test02 where id < 10 and test01.a=test02.b);
+----+--------------------+--------+------------+-------+---------------+---------+---------+------+------+----------+-------------+
| id | select_type    | table | partitions | type | possible_keys | key   | key_len | ref | rows | filtered | Extra    |
+----+--------------------+--------+------------+-------+---------------+---------+---------+------+------+----------+-------------+
| 1 | PRIMARY      | test01 | NULL    | ALL  | NULL     | NULL  | NULL  | NULL | 1000 |  100.00 | Using where |
| 2 | DEPENDENT SUBQUERY | test02 | NULL    | range | PRIMARY    | PRIMARY | 4    | NULL |  9 |  10.00 | Using where |
+----+--------------------+--------+------------+-------+---------------+---------+---------+------+------+----------+-------------+
2 rows in set, 2 warnings (0.00 sec)

不相关子查询变成了关联子查询（select_type:DEPENDENT SUBQUERY），子查询需要根据 b 来关联外表 test01，因为需要外表的 test01 字段，所以子查询是没法先执行的。执行流程为：

扫描 test01，从 test01 取出一行数据 R；
从数据行 R 中，取出字段 a 执行子查询，如果得到结果为 TRUE，则把这行数据 R 放到结果集；
重复 1、2 直到结束。

总的扫描行数为 1000+1000*9=10000（这是理论值，但是实际值比10000还少，怎么来的一直没想明白，看规律是子查询结果集每多一行，总扫描行数就会少几行）。

Semi-join优化器:

这样会有个问题，如果外层表是一个非常大的表，对于外层查询的每一行，子查询都得执行一次，这个查询的性能会非常差。我们很容易想到将其改写成 join 来提升效率：

select test01.* from test01 join test02 on test01.a=test02.b and test02.id<10;

# 查看此SQL的执行计划:

desc select test01.* from test01 join test02 on test01.a=test02.b and test02.id<10;

root@localhost [dbtest01]>EXPLAIN extended select test01.* from test01 join test02 on test01.a=test02.b and test02.id<10;
+----+-------------+--------+------------+-------+---------------+---------+---------+-------------------+------+----------+-------------+
| id | select_type | table | partitions | type | possible_keys | key   | key_len | ref        | rows | filtered | Extra    |
+----+-------------+--------+------------+-------+---------------+---------+---------+-------------------+------+----------+-------------+
| 1 | SIMPLE   | test02 | NULL    | range | PRIMARY    | PRIMARY | 4    | NULL       |  9 |  100.00 | Using where |
| 1 | SIMPLE   | test01 | NULL    | ref  | a       | a    | 5    | dbtest01.test02.b |  1 |  100.00 | NULL    |
+----+-------------+--------+------------+-------+---------------+---------+---------+-------------------+------+----------+-------------+
2 rows in set, 2 warnings (0.00 sec)

这样优化可以让 t2 表做驱动表，t1 表关联字段有索引，查找效率非常高。

但这里会有个问题，join 是有可能得到重复结果的，而 in(select ...) 子查询语义则不会得到重复值。
而 semijoin 正是解决重复值问题的一种特殊联接。
在子查询中，优化器可以识别出 in 子句中每组只需要返回一个值，在这种情况下，可以使用 semijoin 来优化子查询，提升查询效率。
这是 MySQL 5.6 加入的新特性，MySQL 5.6 以前优化器只有 exists 一种策略来“优化”子查询。

经过 semijoin 优化后的 SQL 和执行计划分为：

root@localhost [dbtest01]>desc SELECT * FROM test01 WHERE test01.a IN (SELECT test02.b FROM test02 WHERE id < 10);
+----+--------------+-------------+------------+-------+---------------+---------+---------+---------------+------+----------+-------------+
| id | select_type | table    | partitions | type | possible_keys | key   | key_len | ref      | rows | filtered | Extra    |
+----+--------------+-------------+------------+-------+---------------+---------+---------+---------------+------+----------+-------------+
| 1 | SIMPLE    | <subquery2> | NULL    | ALL  | NULL     | NULL  | NULL  | NULL     | NULL |  100.00 | Using where |
| 1 | SIMPLE    | test01   | NULL    | ref  | a       | a    | 5    | <subquery2>.b |  1 |  100.00 | NULL    |
| 2 | MATERIALIZED | test02   | NULL    | range | PRIMARY    | PRIMARY | 4    | NULL     |  9 |  100.00 | Using where |
+----+--------------+-------------+------------+-------+---------------+---------+---------+---------------+------+----------+-------------+
3 rows in set, 1 warning (0.00 sec)

select 
  `test01`.`id`,`test01`.`a`,`test01`.`b` 
from `test01` semi join `test02` 
where
  ((`test01`.`a` = `<subquery2>`.`b`) 
  and (`test02`.`id` < 10));

##注意这是优化器改写的SQL，客户端上是不能用 semi join 语法的

semijoin 优化实现比较复杂，其中又分 FirstMatch、Materialize 等策略,上面的执行计划中 select_type=MATERIALIZED 就是代表使用了 Materialize 策略来实现的 semijoin
这里 semijoin 优化后的执行流程为：

先执行子查询，把结果保存到一个临时表中，这个临时表有个主键用来去重；
从临时表中取出一行数据 R；
从数据行 R 中，取出字段 b 到被驱动表 t1 中去查找，满足条件则放到结果集；
重复执行 2、3，直到结束。
这样一来，子查询结果有 9 行，即临时表也有 9 行（这里没有重复值），总的扫描行数为 9+9+9*1=27 行，比原来的 10000 行少了很多。

MySQL 5.6 版本中加入的另一种优化特性 materialization，就是把子查询结果物化成临时表，然后代入到外查询中进行查找，来加快查询的执行速度。内存临时表包含主键（hash 索引），消除重复行，使表更小。
如果子查询结果太大，超过 tmp_table_size 大小，会退化成磁盘临时表。这样子查询只需要执行一次，而不是对于外层查询的每一行都得执行一遍。
不过要注意的是，这样外查询依旧无法通过索引快速查找到符合条件的数据，只能通过全表扫描或者全索引扫描，

semijoin 和 materialization 的开启是通过 optimizer_switch 参数中的 semijoin={on|off}、materialization={on|off} 标志来控制的。
上文中不同的执行计划就是对 semijoin 和 materialization 进行开/关产生的
总的来说对于子查询，先检查是否满足各种优化策略的条件（比如子查询中有 union 则无法使用 semijoin 优化）
然后优化器会按成本进行选择，实在没得选就会用 exists 策略来“优化”子查询，exists 策略是没有参数来开启或者关闭的。

下面举一个delete相关的子查询例子：

把上面的2张测试表分别填充350万数据和50万数据来测试delete语句

root@localhost [dbtest01]>select count(*) from test02;
+----------+
| count(*) |
+----------+
| 3532986 |
+----------+
1 row in set (0.64 sec)
root@localhost [dbtest01]>create table test01 like test02;
Query OK, 0 rows affected (0.01 sec)

root@localhost [dbtest01]>insert into test01 (select * from test02 where id<=500000)

root@localhost [dbtest01]>select count(*) from test01;
+----------+
| count(*) |
+----------+
|  500000 |

执行delete删除语句执行了4s

root@localhost [dbtest01]>delete FROM test01 WHERE test01.a IN (SELECT test02.b FROM test02 WHERE id < 10);
Query OK, 9 rows affected (4.86 sec)

查看执行计划，对test01表进行了几乎全表扫描:

root@localhost [dbtest01]>desc delete FROM test01 WHERE test01.a IN (SELECT test02.b FROM test02 WHERE id < 10);
+----+--------------------+--------+------------+-------+---------------+---------+---------+------+--------+----------+-------------+
| id | select_type    | table | partitions | type | possible_keys | key   | key_len | ref | rows  | filtered | Extra    |
+----+--------------------+--------+------------+-------+---------------+---------+---------+------+--------+----------+-------------+
| 1 | DELETE       | test01 | NULL    | ALL  | NULL     | NULL  | NULL  | NULL | 499343 |  100.00 | Using where |
| 2 | DEPENDENT SUBQUERY | test02 | NULL    | range | PRIMARY    | PRIMARY | 4    | NULL |   9 |  10.00 | Using where |
+----+--------------------+--------+------------+-------+---------------+---------+---------+------+--------+----------+-------------+
2 rows in set (0.00 sec)

于是修改上面的delete SQL语句伪join语句

root@localhost [dbtest01]>desc delete test01.* from test01 join test02 on test01.a=test02.b and test02.id<10;
+----+-------------+--------+------------+-------+---------------+---------+---------+-------------------+------+----------+-------------+
| id | select_type | table | partitions | type | possible_keys | key   | key_len | ref        | rows | filtered | Extra    |
+----+-------------+--------+------------+-------+---------------+---------+---------+-------------------+------+----------+-------------+
| 1 | SIMPLE   | test02 | NULL    | range | PRIMARY    | PRIMARY | 4    | NULL       |  9 |  100.00 | Using where |
| 1 | DELETE   | test01 | NULL    | ref  | a       | a    | 5    | dbtest01.test02.b |  1 |  100.00 | NULL    |
+----+-------------+--------+------------+-------+---------------+---------+---------+-------------------+------+----------+-------------+
2 rows in set (0.01 sec)

执行非常的快
root@localhost [dbtest01]>delete test01.* from test01 join test02 on test01.a=test02.b and test02.id<10;
Query OK, 9 rows affected (0.01 sec)

root@localhost [dbtest01]>select test01.* from test01 join test02 on test01.a=test02.b and test02.id<10;
Empty set (0.00 sec)

下面的这个表执行要全表扫描，非常慢，基本对表test01进行了全表扫描:

root@lcalhost [dbtest01]>desc delete FROM test01 WHERE id IN (SELECT id FROM test02 WHERE id='350000');
+----+--------------------+--------+------------+-------+---------------+---------+---------+-------+--------+----------+-------------+
| id | select_type    | table | partitions | type | possible_keys | key   | key_len | ref  | rows  | filtered | Extra    |
+----+--------------------+--------+------------+-------+---------------+---------+---------+-------+--------+----------+-------------+
| 1 | DELETE       | test01 | NULL    | ALL  | NULL     | NULL  | NULL  | NULL | 499343 |  100.00 | Using where |
| 2 | DEPENDENT SUBQUERY | test02 | NULL    | const | PRIMARY    | PRIMARY | 4    | const |   1 |  100.00 | Using index |
+----+--------------------+--------+------------+-------+---------------+---------+---------+-------+--------+----------+-------------+
2 rows in set (0.00 sec)

然而采用join的话，效率非常的高:

root@localhost [dbtest01]>desc delete test01.* FROM test01 inner join test02 WHERE test01.id=test02.id and test02.id=350000 ;
+----+-------------+--------+------------+-------+---------------+---------+---------+-------+------+----------+-------------+
| id | select_type | table | partitions | type | possible_keys | key   | key_len | ref  | rows | filtered | Extra    |
+----+-------------+--------+------------+-------+---------------+---------+---------+-------+------+----------+-------------+
| 1 | DELETE   | test01 | NULL    | const | PRIMARY    | PRIMARY | 4    | const |  1 |  100.00 | NULL    |
| 1 | SIMPLE   | test02 | NULL    | const | PRIMARY    | PRIMARY | 4    | const |  1 |  100.00 | Using index |
+----+-------------+--------+------------+-------+---------------+---------+---------+-------+------+----------+-------------+
2 rows in set (0.01 sec)

 
root@localhost [dbtest01]> desc delete test01.* from test01 join test02 on test01.a=test02.b and test02.id=350000;
+----+-------------+--------+------------+-------+---------------+---------+---------+-------+------+----------+-------+
| id | select_type | table | partitions | type | possible_keys | key   | key_len | ref  | rows | filtered | Extra |
+----+-------------+--------+------------+-------+---------------+---------+---------+-------+------+----------+-------+
| 1 | SIMPLE   | test02 | NULL    | const | PRIMARY    | PRIMARY | 4    | const |  1 |  100.00 | NULL |
| 1 | DELETE   | test01 | NULL    | ref  | a       | a    | 5    | const |  1 |  100.00 | NULL |
+----+-------------+--------+------------+-------+---------------+---------+---------+-------+------+----------+-------+
2 rows in set (0.00 sec)

参考文档:

https://www.cnblogs.com/zhengyun_ustc/p/slowquery1.html
https://www.jianshu.com/p/3989222f7084
https://dev.mysql.com/doc/refman/5.6/en/subquery-optimization.html

上一篇：Mysql NULL导致的神坑
下一篇：MySQL循环插入千万级数据