问题语句
SELECT * FROM a
WHERE `type` = 'appointment'
AND `event` = 14
AND EXISTS (
SELECT * FROM b
WHERE a.`sheet_id` = b.`id`
AND `company_id` = 8
AND b.`deleted_at` IS NULL )
ORDER BY a.id DESC
LIMIT 6;
解读执行计划
在exists类型的子查询的执行计划中，select_type一栏分别是PRIMARY和DEPENDENT SUBQUERY
DEPENDENT SUBQUERY的意思是：子查询，依赖于外层的查询;DEPEND SUBQUERY是依赖于SQL的主体部分，它的执行次数最大可能和SQL主体部分结果的行数一样多（这里因为有limit6，所以看起来主表扫描行数是6，如果去掉这个limit6的话，这个值是1500W）
上面这句话解释得通俗一点就是外连接先执行查询，然后把查询的结果集放入子查询内进行匹配；外查询每执行一次查询，就要来子查询匹配一次
join的执行计划中，select_type一栏都是simple
join的第一行的就是外表
从上面的对比可以看出，无论是 exists类型的子查询 还是join，都基本可以看作遵循了第一行就是驱动表的规则（注意不是所有子查询都遵循这个规则，本篇只针对exists类型的dependent subquery）
分析
图一PRIMARY对应的表就是图二中的a表，DEPENDENT SUBQUERY表就是图二中的b表；a表有1500W行数据，b表有2W行数据
所以图一的SQL执行效率如此低下的原因就是大表驱动小表
优化
exists改写为join
SELECT a.*
FROM a join b on a.`sheet_id` = b.`id`
WHERE a.`type` = 'appointment'
AND a.`event` = 14
AND b.`company_id` = 8
AND b.`deleted_at` IS NULL
ORDER BY a.`id` DESC
LIMIT 6;
由于a表作为内表，因此在a.`sheet_id`，a.`type`，a.`event`上创建联合索引；语句中出现了b表的本地谓词，所以b表的b.`company_id`，b.`deleted_at`上也要创建联合索引
优化结果，执行时间：117s→0.36s，性能提升了2000倍
这个语句有一个更极端的取值，在b.`company_id` = 2的时候，小表不会搜出任何满足条件的结果，在这种情况下，原语句执行时间在350s以上，而新语句仅需要0.03s，性能提升万倍
优化案例
今天优化的这批语句中，大多数是exists子查询的问题，可以看出这个研发小哥非常的喜欢用exists这种写法；前面的那个exists语句是泛用型，后面的exists语句加了些新花样
eg.
SELECT SUM(`xxxx`) AS ag
FROM a
WHERE EXISTS (
SELECT * FROM b
WHERE a.`delivery_sheet_id` = b.`id`
AND (`status` = 4
OR `is_rejected` = '1')
AND `company_id` = 8
AND b.`deleted_at` IS NULL
)
AND `status` IN (0, 4)
AND `collection_type` IN (2, 3)
AND a.`deleted_at` IS NULL;
or的优化通常改写union，但这里是求sum不能这么改，需要改写成2个语句然后求和；对应的列要建好索引
select c.ag+d.ag as ag from
(SELECT SUM(a.`xxxx`) AS ag
FROM a join b
on a.`delivery_sheet_id` = b.`id`
where
b.`status` = 4
AND b.`company_id` = 8
AND b.`deleted_at` IS NULL
AND a.`status` IN (0, 4)
AND a.`collection_type` IN (2, 3)
AND a.`deleted_at` IS NULL) c,
(
SELECT SUM(a.`xxxx`) AS ag
FROM a join b
on a.`delivery_sheet_id` = b.`id`
where
b.`is_rejected` = '1'
AND b.`company_id` = 8
AND b.`deleted_at` IS NULL
AND a.`status` IN (0, 4)
AND a.`collection_type` IN (2, 3)
AND a.`deleted_at` IS NULL) d;
优化结果，执行时间：18s→0.2s
in改写join的思路和exists差不多
这里没有现成的例子，粘贴一篇郑松华老师公众号的分析过来
原语句
SELECT
COUNT( * ) AS totalNum,
sum( CASE WHEN F.ALARM_LEVEL = 1 THEN 1 ELSE 0 END ) AS LEVELS1,
sum( CASE WHEN F.ALARM_LEVEL = 2 THEN 1 ELSE 0 END ) AS LEVELS2,
sum( CASE WHEN F.ALARM_LEVEL = 3 THEN 1 ELSE 0 END ) AS LEVELS3,
sum( CASE WHEN F.DEAL_STATE = 0 THEN 1 ELSE 0 END ) AS DESTS
FROM
F
LEFT JOIN DC ON DC.ID = F.CONST_ID
LEFT JOIN V ON V.ID = F.VEHICLE_ID
LEFT JOIN AREA ON AREA.ID = V.SYS_DIVISION_ID
WHERE
DC.ID IS NOT NULL
AND V.ID IS NOT NULL
AND F.DEAL_STATE = 0
AND ALARM_LEVEL IN ( 1, 2, 3 )
AND F.VEHICLE_ID IN (
SELECT
VEHICLE_ID
FROM
GVLK
WHERE
GROUP_ID IN ( SELECT GROUP_ID FROM GULK WHERE USER_ID = 'ff8080816091b09c0161f9b825750a9a' )
UNION
SELECT
VEHICLE_ID
FROM
UVLK
WHERE
USER_ID = 'ff8080816091b09c0161f9b825750a9a'
)
AND date( F.ALARM_TIME ) BETWEEN '2000-01-01'
AND '2018-08-14'
AND AREA.PATH LIKE CONCAT( ( SELECT ARE.PATH FROM ARE WHERE ARE.ID = '0' ), '%' )
执行计划如下
改写如下（in改join）
explain extended
SELECT
COUNT( * ) AS totalNum,
sum( CASE WHEN F.ALARM_LEVEL = 1 THEN 1 ELSE 0 END ) AS LEVELS1,
sum( CASE WHEN F.ALARM_LEVEL = 2 THEN 1 ELSE 0 END ) AS LEVELS2,
sum( CASE WHEN F.ALARM_LEVEL = 3 THEN 1 ELSE 0 END ) AS LEVELS3,
sum( CASE WHEN F.DEAL_STATE = 0 THEN 1 ELSE 0 END ) AS DESTS
FROM
F
straight_join (
SELECT
VEHICLE_ID
FROM
GVLK
WHERE
GROUP_ID IN ( SELECT GROUP_ID FROM GULK WHERE USER_ID = 'ff8080816091b09c0161f9b825750a9a' )
UNION
SELECT
VEHICLE_ID
FROM
UVLK
WHERE
USER_ID = 'ff8080816091b09c0161f9b825750a9a'
) s on F.VEHICLE_ID = s.VEHICLE_ID
straight_join DC ON DC.ID = F.CONST_ID
straight_join V ON V.ID = F.VEHICLE_ID
straight_join AREA ON AREA.ID = V.SYS_DIVISION_ID
WHERE
DC.ID IS NOT NULL
AND V.ID IS NOT NULL
AND F.DEAL_STATE = 0
AND ALARM_LEVEL IN ( 1, 2, 3 )
AND date( F.ALARM_TIME ) BETWEEN '2000-01-01'
AND '2018-08-14'
AND AREA.PATH LIKE CONCAT( ( SELECT ARE.PATH FROM ARE WHERE ARE.ID = '0' ), '%' )