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
ANDb.`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'
ANDb.`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' ), '%' )