前言
mysql在2016年仍然保持强劲的数据库流行度增长趋势。越来越多的客户将自己的应用建立在mysql数据库之上，甚至是从oracle迁移到mysql上来。但也存在部分客户在使用mysql数据库的过程中遇到一些比如响应时间慢，cpu打满等情况。
阿里云rds专家服务团队帮助云上客户解决过很多紧急问题。现将《apsaradb专家诊断报告》中出现的部分常见sql问题总结如下，供大家参考。
1、limit 语句
分页查询是最常用的场景之一，但也通常也是最容易出问题的地方。
比如对于下面简单的语句，一般 dba 想到的办法是在 type, name, create_time 字段上加组合索引。这样条件排序都能有效的利用到索引，性能迅速提升。
select *
from operation
where type = 'sqlstats'
and name = 'slowlog'
order by create_time
limit 1000, 10;复制代码
好吧，可能90%以上的 dba 解决该问题就到此为止。
但当 limit 子句变成 “limit 1000000,10” 时，程序员仍然会抱怨：我只取10条记录为什么还是慢？
要知道数据库也并不知道第1000000条记录从什么地方开始，即使有索引也需要从头计算一次。出现这种性能问题，多数情形下是程序员偷懒了。
在前端数据浏览翻页，或者大数据分批导出等场景下，是可以将上一页的最大值当成参数作为查询条件的。sql 重新设计如下：
select * from operation where type = 'sqlstats' and name = 'slowlog' and create_time > '2017-03-16 14:00:00' order by create_time limit 10;复制代码
在新设计下查询时间基本固定，不会随着数据量的增长而发生变化。
2、隐式转换
sql语句中查询变量和字段定义类型不匹配是另一个常见的错误。比如下面的语句：
mysql> explain extended select *
> from my_balance b
> where b.bpn = 14000000123
> and b.isverified is null ;
mysql> show warnings;
| warning | 1739 | cannot use ref access on index 'bpn' due to type or collation conversion on field 'bpn'复制代码
其中字段 bpn 的定义为 varchar(20)，mysql 的策略是将字符串转换为数字之后再比较。函数作用于表字段，索引失效。
上述情况可能是应用程序框架自动填入的参数，而不是程序员的原意。现在应用框架很多很繁杂，使用方便的同时也小心它可能给自己挖坑。
3、关联更新、删除
虽然 mysql5.6 引入了物化特性，但需要特别注意它目前仅仅针对查询语句的优化。对于更新或删除需要手工重写成 join。
比如下面 update 语句，mysql 实际执行的是循环/嵌套子查询（dependent subquery)，其执行时间可想而知。
update operation o
set status = 'applying'
where o.id in (select id
from (select o.id,
o.status
from operation o
where o.group = 123
and o.status not in ( 'done' )
order by o.parent,
o.id
limit 1) t);复制代码
执行计划：
+----+--------------------+-------+-------+---------------+---------+---------+-------+------+-----------------------------------------------------+ | id | select_type | table | type | possible_keys | key | key_len | ref | rows | extra | +----+--------------------+-------+-------+---------------+---------+---------+-------+------+-----------------------------------------------------+ | 1 | primary | o | index | | primary | 8 | | 24 | using where; using temporary | | 2 | dependent subquery | | | | | | | | impossible where noticed after reading const tables | | 3 | derived | o | ref | idx_2,idx_5 | idx_5 | 8 | const | 1 | using where; using filesort | +----+--------------------+-------+-------+---------------+---------+---------+-------+------+-----------------------------------------------------+复制代码
重写为 join 之后，子查询的选择模式从 dependent subquery 变成 derived，执行速度大大加快，从7秒降低到2毫秒。
update operation o
join (select o.id,
o.status
from operation o
where o.group = 123
and o.status not in ( 'done' )
order by o.parent,
o.id
limit 1) t
on o.id = t.id
set status = 'applying' 复制代码
执行计划简化为：
+----+-------------+-------+------+---------------+-------+---------+-------+------+-----------------------------------------------------+ | id | select_type | table | type | possible_keys | key | key_len | ref | rows | extra | +----+-------------+-------+------+---------------+-------+---------+-------+------+-----------------------------------------------------+ | 1 | primary | | | | | | | | impossible where noticed after reading const tables | | 2 | derived | o | ref | idx_2,idx_5 | idx_5 | 8 | const | 1 | using where; using filesort | +----+-------------+-------+------+---------------+-------+---------+-------+------+-----------------------------------------------------+复制代码
4、混合排序
mysql 不能利用索引进行混合排序。但在某些场景，还是有机会使用特殊方法提升性能的。
select *
from my_order o
inner join my_appraise a on a.orderid = o.id
order by a.is_reply asc,
a.appraise_time desc
limit 0, 20 复制代码
执行计划显示为全表扫描：
+----+-------------+-------+--------+-------------+---------+---------+---------------+---------+-+ | id | select_type | table | type | possible_keys | key | key_len | ref | rows | extra +----+-------------+-------+--------+-------------+---------+---------+---------------+---------+-+ | 1 | simple | a | all | idx_orderid | null | null | null | 1967647 | using filesort | | 1 | simple | o | eq_ref | primary | primary | 122 | a.orderid | 1 | null | +----+-------------+-------+--------+---------+---------+---------+-----------------+---------+-+复制代码
由于 is_reply 只有0和1两种状态，我们按照下面的方法重写后，执行时间从1.58秒降低到2毫秒。
select *
from ((select *
from my_order o
inner join my_appraise a
on a.orderid = o.id
and is_reply = 0
order by appraise_time desc
limit 0, 20)
union all
(select *
from my_order o
inner join my_appraise a
on a.orderid = o.id
and is_reply = 1
order by appraise_time desc
limit 0, 20)) t
order by is_reply asc,
appraisetime desc
limit 20;复制代码
5、exists语句
mysql 对待 exists 子句时，仍然采用嵌套子查询的执行方式。如下面的 sql 语句：
select *
from my_neighbor n
left join my_neighbor_apply sra
on n.id = sra.neighbor_id
and sra.user_id = 'xxx'
where n.topic_status < 4
and exists(select 1
from message_info m
where n.id = m.neighbor_id
and m.inuser = 'xxx')
and n.topic_type <> 5 复制代码
执行计划为：
+----+--------------------+-------+------+-----+------------------------------------------+---------+-------+---------+ -----+ | id | select_type | table | type | possible_keys | key | key_len | ref | rows | extra | +----+--------------------+-------+------+ -----+------------------------------------------+---------+-------+---------+ -----+ | 1 | primary | n | all | | null | null | null | 1086041 | using where | | 1 | primary | sra | ref | | idx_user_id | 123 | const | 1 | using where | | 2 | dependent subquery | m | ref | | idx_message_info | 122 | const | 1 | using index condition; using where | +----+--------------------+-------+------+ -----+------------------------------------------+---------+-------+---------+ -----+复制代码
去掉 exists 更改为 join，能够避免嵌套子查询，将执行时间从1.93秒降低为1毫秒。
select *
from my_neighbor n
inner join message_info m
on n.id = m.neighbor_id
and m.inuser = 'xxx'
left join my_neighbor_apply sra
on n.id = sra.neighbor_id
and sra.user_id = 'xxx'
where n.topic_status < 4
and n.topic_type <> 5 复制代码
新的执行计划：
+----+-------------+-------+--------+ -----+------------------------------------------+---------+ -----+------+ -----+ | id | select_type | table | type | possible_keys | key | key_len | ref | rows | extra | +----+-------------+-------+--------+ -----+------------------------------------------+---------+ -----+------+ -----+ | 1 | simple | m | ref | | idx_message_info | 122 | const | 1 | using index condition | | 1 | simple | n | eq_ref | | primary | 122 | ighbor_id | 1 | using where | | 1 | simple | sra | ref | | idx_user_id | 123 | const | 1 | using where | +----+-------------+-------+--------+ -----+------------------------------------------+---------+ -----+------+ -----+复制代码
6、条件下推
外部查询条件不能够下推到复杂的视图或子查询的情况有：
-
聚合子查询；
-
含有 limit 的子查询；
-
union 或 union all 子查询；
-
输出字段中的子查询；
如下面的语句，从执行计划可以看出其条件作用于聚合子查询之后：
select *
from (select target,
count(*)
from operation
group by target) t
where target = 'rm-xxxx'复制代码
+----+-------------+------------+-------+---------------+-------------+---------+-------+------+-------------+ | id | select_type | table | type | possible_keys | key | key_len | ref | rows | extra | +----+-------------+------------+-------+---------------+-------------+---------+-------+------+-------------+ | 1 | primary | <derived2> | ref | <auto_key 0 > | <auto_key0> | 514 | const | 2 | using where | | 2 | derived | operation | index | idx_4 | idx_4 | 519 | null | 20 | using index | +----+-------------+------------+-------+---------------+-------------+---------+-------+------+-------------+复制代码
确定从语义上查询条件可以直接下推后，重写如下：
select target,
count(*)
from operation
where target = 'rm-xxxx'
group by target复制代码
执行计划变为：
+----+-------------+-----------+------+---------------+-------+---------+-------+------+--------------------+ | id | select_type | table | type | possible_keys | key | key_len | ref | rows | extra | +----+-------------+-----------+------+---------------+-------+---------+-------+------+--------------------+ | 1 | simple | operation | ref | idx_4 | idx_4 | 514 | const | 1 | using where; using index | +----+-------------+-----------+------+---------------+-------+---------+-------+------+--------------------+复制代码
关于 mysql 外部条件不能下推的详细解释说明请参考文章：
http://mysql.taobao.org/monthly/2016/07/08
7、提前缩小范围
先上初始 sql 语句：
select *
from my_order o
left join my_userinfo u
on o.uid = u.uid
left join my_productinfo p
on o.pid = p.pid
where ( o.display = 0 )
and ( o.ostaus = 1 )
order by o.selltime desc
limit 0, 15 复制代码
该sql语句原意是：先做一系列的左连接，然后排序取前15条记录。从执行计划也可以看出，最后一步估算排序记录数为90万，时间消耗为12秒。
+----+-------------+-------+--------+---------------+---------+---------+-----------------+--------+----------------------------------------------------+ | id | select_type | table | type | possible_keys | key | key_len | ref | rows | extra | +----+-------------+-------+--------+---------------+---------+---------+-----------------+--------+----------------------------------------------------+ | 1 | simple | o | all | null | null | null | null | 909119 | using where; using temporary; using filesort | | 1 | simple | u | eq_ref | primary | primary | 4 | o.uid | 1 | null | | 1 | simple | p | all | primary | null | null | null | 6 | using where; using join buffer (block nested loop) | +----+-------------+-------+--------+---------------+---------+---------+-----------------+--------+----------------------------------------------------+复制代码
由于最后 where 条件以及排序均针对最左主表，因此可以先对 my_order 排序提前缩小数据量再做左连接。sql 重写后如下，执行时间缩小为1毫秒左右。
select *
from (
select *
from my_order o
where ( o.display = 0 )
and ( o.ostaus = 1 )
order by o.selltime desc
limit 0, 15
) o
left join my_userinfo u
on o.uid = u.uid
left join my_productinfo p
on o.pid = p.pid
order by o.selltime desc
limit 0, 15复制代码
再检查执行计划：子查询物化后（select_type=derived)参与 join。虽然估算行扫描仍然为90万，但是利用了索引以及 limit 子句后，实际执行时间变得很小。
+----+-------------+------------+--------+---------------+---------+---------+-------+--------+----------------------------------------------------+ | id | select_type | table | type | possible_keys | key | key_len | ref | rows | extra | +----+-------------+------------+--------+---------------+---------+---------+-------+--------+----------------------------------------------------+ | 1 | primary | <derived2> | all | null | null | null | null | 15 | using temporary; using filesort | | 1 | primary | u | eq_ref | primary | primary | 4 | o.uid | 1 | null | | 1 | primary | p | all | primary | null | null | null | 6 | using where; using join buffer (block nested loop) | | 2 | derived | o | index | null | idx_1 | 5 | null | 909112 | using where | +----+-------------+------------+--------+---------------+---------+---------+-------+--------+----------------------------------------------------+复制代码
8、中间结果集下推
再来看下面这个已经初步优化过的例子(左连接中的主表优先作用查询条件)：
select a.*,
c.allocated
from (
select resourceid
from my_distribute d
where isdelete = 0
and cusmanagercode = '1234567'
order by salecode limit 20) a
left join
(
select resourcesid， sum(ifnull(allocation, 0) * 12345) allocated
from my_resources
group by resourcesid) c
on a.resourceid = c.resourcesid复制代码
那么该语句还存在其它问题吗？不难看出子查询 c 是全表聚合查询，在表数量特别大的情况下会导致整个语句的性能下降。
其实对于子查询 c，左连接最后结果集只关心能和主表 resourceid 能匹配的数据。因此我们可以重写语句如下，执行时间从原来的2秒下降到2毫秒。
select a.*,
c.allocated
from (
select resourceid
from my_distribute d
where isdelete = 0
and cusmanagercode = '1234567'
order by salecode limit 20) a
left join
(
select resourcesid， sum(ifnull(allocation, 0) * 12345) allocated
from my_resources r,
(
select resourceid
from my_distribute d
where isdelete = 0
and cusmanagercode = '1234567'
order by salecode limit 20) a
where r.resourcesid = a.resourcesid
group by resourcesid) c
on a.resourceid = c.resourcesid复制代码
但是子查询 a 在我们的sql语句中出现了多次。这种写法不仅存在额外的开销，还使得整个语句显的繁杂。使用 with 语句再次重写：
with a as
(
select resourceid
from my_distribute d
where isdelete = 0
and cusmanagercode = '1234567'
order by salecode limit 20)
select a.*,
c.allocated
from a
left join
(
select resourcesid， sum(ifnull(allocation, 0) * 12345) allocated
from my_resources r,
a
where r.resourcesid = a.resourcesid
group by resourcesid) c
on a.resourceid = c.resourcesid复制代码
总结
数据库编译器产生执行计划，决定着sql的实际执行方式。但是编译器只是尽力服务，所有数据库的编译器都不是尽善尽美的。
上述提到的多数场景，在其它数据库中也存在性能问题。了解数据库编译器的特性，才能避规其短处，写出高性能的sql语句。
程序员在设计数据模型以及编写sql语句时，要把算法的思想或意识带进来。
编写复杂sql语句要养成使用 with 语句的习惯。简洁且思路清晰的sql语句也能减小数据库的负担 。