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。

常见SQL错误用法

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。

  1. SELECT *
  2. FROM operation
  3. WHERE type = 'SQLStats'
  4. AND name = 'SlowLog'
  5. ORDER BY create_time
  6. LIMIT 1000, 10;

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:

  1. SELECT *
  2. FROM operation
  3. WHERE type = 'SQLStats'
  4. AND name = 'SlowLog'
  5. AND create_time > '2017-03-16 14:00:00'
  6. ORDER BY create_time limit 10;

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:

  1. mysql> explain extended SELECT *
  2. > FROM my_balance b
  3. > WHERE b.bpn = 14000000123
  4. > AND b.isverified IS NULL ;
  5. mysql> show warnings;
  6. | Warning | 1739 | Cannot use ref access on index 'bpn' due to type or collation conversion on field 'bpn'

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。

  1. UPDATE operation o
  2. SET status = 'applying'
  3. WHERE o.id IN (SELECT id
  4. FROM (SELECT o.id,
  5. o.status
  6. FROM operation o
  7. WHERE o.group = 123
  8. AND o.status NOT IN ( 'done' )
  9. ORDER BY o.parent,
  10. o.id
  11. LIMIT 1) t);

执行计划:

  1. +----+--------------------+-------+-------+---------------+---------+---------+-------+------+-----------------------------------------------------+
  2. | id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
  3. +----+--------------------+-------+-------+---------------+---------+---------+-------+------+-----------------------------------------------------+
  4. | 1 | PRIMARY | o | index | | PRIMARY | 8 | | 24 | Using where; Using temporary |
  5. | 2 | DEPENDENT SUBQUERY | | | | | | | | Impossible WHERE noticed after reading const tables |
  6. | 3 | DERIVED | o | ref | idx_2,idx_5 | idx_5 | 8 | const | 1 | Using where; Using filesort |
  7. +----+--------------------+-------+-------+---------------+---------+---------+-------+------+-----------------------------------------------------+

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。

  1. UPDATE operation o
  2. JOIN (SELECT o.id,
  3. o.status
  4. FROM operation o
  5. WHERE o.group = 123
  6. AND o.status NOT IN ( 'done' )
  7. ORDER BY o.parent,
  8. o.id
  9. LIMIT 1) t
  10. ON o.id = t.id
  11. SET status = 'applying'

执行计划简化为:

  1. +----+-------------+-------+------+---------------+-------+---------+-------+------+-----------------------------------------------------+
  2. | id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
  3. +----+-------------+-------+------+---------------+-------+---------+-------+------+-----------------------------------------------------+
  4. | 1 | PRIMARY | | | | | | | | Impossible WHERE noticed after reading const tables |
  5. | 2 | DERIVED | o | ref | idx_2,idx_5 | idx_5 | 8 | const | 1 | Using where; Using filesort |
  6. +----+-------------+-------+------+---------------+-------+---------+-------+------+-----------------------------------------------------+

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。

  1. SELECT *
  2. FROM my_order o
  3. INNER JOIN my_appraise a ON a.orderid = o.id
  4. ORDER BY a.is_reply ASC,
  5. a.appraise_time DESC
  6. LIMIT 0, 20

执行计划显示为全表扫描:

  1. +----+-------------+-------+--------+-------------+---------+---------+---------------+---------+-+
  2. | id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra
  3. +----+-------------+-------+--------+-------------+---------+---------+---------------+---------+-+
  4. | 1 | SIMPLE | a | ALL | idx_orderid | NULL | NULL | NULL | 1967647 | Using filesort |
  5. | 1 | SIMPLE | o | eq_ref | PRIMARY | PRIMARY | 122 | a.orderid | 1 | NULL |
  6. +----+-------------+-------+--------+---------+---------+---------+-----------------+---------+-+

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。

  1. SELECT *
  2. FROM ((SELECT *
  3. FROM my_order o
  4. INNER JOIN my_appraise a
  5. ON a.orderid = o.id
  6. AND is_reply = 0
  7. ORDER BY appraise_time DESC
  8. LIMIT 0, 20)
  9. UNION ALL
  10. (SELECT *
  11. FROM my_order o
  12. INNER JOIN my_appraise a
  13. ON a.orderid = o.id
  14. AND is_reply = 1
  15. ORDER BY appraise_time DESC
  16. LIMIT 0, 20)) t
  17. ORDER BY is_reply ASC,
  18. appraisetime DESC
  19. LIMIT 20;

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:

  1. SELECT *
  2. FROM my_neighbor n
  3. LEFT JOIN my_neighbor_apply sra
  4. ON n.id = sra.neighbor_id
  5. AND sra.user_id = 'xxx'
  6. WHERE n.topic_status < 4
  7. AND EXISTS(SELECT 1
  8. FROM message_info m
  9. WHERE n.id = m.neighbor_id
  10. AND m.inuser = 'xxx')
  11. AND n.topic_type <> 5

执行计划为:

  1. +----+--------------------+-------+------+-----+------------------------------------------+---------+-------+---------+ -----+
  2. | id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
  3. +----+--------------------+-------+------+ -----+------------------------------------------+---------+-------+---------+ -----+
  4. | 1 | PRIMARY | n | ALL | | NULL | NULL | NULL | 1086041 | Using where |
  5. | 1 | PRIMARY | sra | ref | | idx_user_id | 123 | const | 1 | Using where |
  6. | 2 | DEPENDENT SUBQUERY | m | ref | | idx_message_info | 122 | const | 1 | Using index condition; Using where |
  7. +----+--------------------+-------+------+ -----+------------------------------------------+---------+-------+---------+ -----+

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。

  1. SELECT *
  2. FROM my_neighbor n
  3. INNER JOIN message_info m
  4. ON n.id = m.neighbor_id
  5. AND m.inuser = 'xxx'
  6. LEFT JOIN my_neighbor_apply sra
  7. ON n.id = sra.neighbor_id
  8. AND sra.user_id = 'xxx'
  9. WHERE n.topic_status < 4
  10. AND n.topic_type <> 5

新的执行计划:

  1. +----+-------------+-------+--------+ -----+------------------------------------------+---------+ -----+------+ -----+
  2. | id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
  3. +----+-------------+-------+--------+ -----+------------------------------------------+---------+ -----+------+ -----+
  4. | 1 | SIMPLE | m | ref | | idx_message_info | 122 | const | 1 | Using index condition |
  5. | 1 | SIMPLE | n | eq_ref | | PRIMARY | 122 | ighbor_id | 1 | Using where |
  6. | 1 | SIMPLE | sra | ref | | idx_user_id | 123 | const | 1 | Using where |
  7. +----+-------------+-------+--------+ -----+------------------------------------------+---------+ -----+------+ -----+

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:

  1. 聚合子查询;
  2. 含有LIMIT的子查询;
  3. UNION 或UNION ALL子查询;
  4. 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:

  1. SELECT *
  2. FROM (SELECT target,
  3. Count(*)
  4. FROM operation
  5. GROUP BY target) t
  6. WHERE target = 'rm-xxxx'
  1. +----+-------------+------------+-------+---------------+-------------+---------+-------+------+-------------+
  2. | id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
  3. +----+-------------+------------+-------+---------------+-------------+---------+-------+------+-------------+
  4. | 1 | PRIMARY | <derived2> | ref | <auto_key0> | <auto_key0> | 514 | const | 2 | Using where |
  5. | 2 | DERIVED | operation | index | idx_4 | idx_4 | 519 | NULL | 20 | Using index |
  6. +----+-------------+------------+-------+---------------+-------------+---------+-------+------+-------------+

确定从语义上查询条件可以直接下推后,重写如下:

  1. SELECT target,
  2. Count(*)
  3. FROM operation
  4. WHERE target = 'rm-xxxx'
  5. GROUP BY target

执行计划变为:

  1. +----+-------------+-----------+------+---------------+-------+---------+-------+------+--------------------+
  2. | id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
  3. +----+-------------+-----------+------+---------------+-------+---------+-------+------+--------------------+
  4. | 1 | SIMPLE | operation | ref | idx_4 | idx_4 | 514 | const | 1 | Using where; Using index |
  5. +----+-------------+-----------+------+---------------+-------+---------+-------+------+--------------------+

关于MySQL外部条件不能下推的详细解释说明请参考以前文章:MySQL · 性能优化 · 条件下推到物化表

7. 提前缩小范围

先上初始SQL语句:

  1. SELECT *
  2. FROM my_order o
  3. LEFT JOIN my_userinfo u
  4. ON o.uid = u.uid
  5. LEFT JOIN my_productinfo p
  6. ON o.pid = p.pid
  7. WHERE ( o.display = 0 )
  8. AND ( o.ostaus = 1 )
  9. ORDER BY o.selltime DESC
  10. LIMIT 0, 15

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。

  1. +----+-------------+-------+--------+---------------+---------+---------+-----------------+--------+----------------------------------------------------+
  2. | id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
  3. +----+-------------+-------+--------+---------------+---------+---------+-----------------+--------+----------------------------------------------------+
  4. | 1 | SIMPLE | o | ALL | NULL | NULL | NULL | NULL | 909119 | Using where; Using temporary; Using filesort |
  5. | 1 | SIMPLE | u | eq_ref | PRIMARY | PRIMARY | 4 | o.uid | 1 | NULL |
  6. | 1 | SIMPLE | p | ALL | PRIMARY | NULL | NULL | NULL | 6 | Using where; Using join buffer (Block Nested Loop) |
  7. +----+-------------+-------+--------+---------------+---------+---------+-----------------+--------+----------------------------------------------------+

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。

  1. SELECT *
  2. FROM (
  3. SELECT *
  4. FROM my_order o
  5. WHERE ( o.display = 0 )
  6. AND ( o.ostaus = 1 )
  7. ORDER BY o.selltime DESC
  8. LIMIT 0, 15
  9. ) o
  10. LEFT JOIN my_userinfo u
  11. ON o.uid = u.uid
  12. LEFT JOIN my_productinfo p
  13. ON o.pid = p.pid
  14. ORDER BY o.selltime DESC
  15. limit 0, 15

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。

  1. +----+-------------+------------+--------+---------------+---------+---------+-------+--------+----------------------------------------------------+
  2. | id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
  3. +----+-------------+------------+--------+---------------+---------+---------+-------+--------+----------------------------------------------------+
  4. | 1 | PRIMARY | <derived2> | ALL | NULL | NULL | NULL | NULL | 15 | Using temporary; Using filesort |
  5. | 1 | PRIMARY | u | eq_ref | PRIMARY | PRIMARY | 4 | o.uid | 1 | NULL |
  6. | 1 | PRIMARY | p | ALL | PRIMARY | NULL | NULL | NULL | 6 | Using where; Using join buffer (Block Nested Loop) |
  7. | 2 | DERIVED | o | index | NULL | idx_1 | 5 | NULL | 909112 | Using where |
  8. +----+-------------+------------+--------+---------------+---------+---------+-------+--------+----------------------------------------------------+

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):

  1. SELECT a.*,
  2. c.allocated
  3. FROM (
  4. SELECT resourceid
  5. FROM my_distribute d
  6. WHERE isdelete = 0
  7. AND cusmanagercode = '1234567'
  8. ORDER BY salecode limit 20) a
  9. LEFT JOIN
  10. (
  11. SELECT resourcesid sum(ifnull(allocation, 0) * 12345) allocated
  12. FROM my_resources
  13. GROUP BY resourcesid) c
  14. ON a.resourceid = c.resourcesid

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。

  1. SELECT a.*,
  2. c.allocated
  3. FROM (
  4. SELECT resourceid
  5. FROM my_distribute d
  6. WHERE isdelete = 0
  7. AND cusmanagercode = '1234567'
  8. ORDER BY salecode limit 20) a
  9. LEFT JOIN
  10. (
  11. SELECT resourcesid sum(ifnull(allocation, 0) * 12345) allocated
  12. FROM my_resources r,
  13. (
  14. SELECT resourceid
  15. FROM my_distribute d
  16. WHERE isdelete = 0
  17. AND cusmanagercode = '1234567'
  18. ORDER BY salecode limit 20) a
  19. WHERE r.resourcesid = a.resourcesid
  20. GROUP BY resourcesid) c
  21. ON a.resourceid = c.resourcesid

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:

  1. WITH a AS
  2. (
  3. SELECT resourceid
  4. FROM my_distribute d
  5. WHERE isdelete = 0
  6. AND cusmanagercode = '1234567'
  7. ORDER BY salecode limit 20)
  8. SELECT a.*,
  9. c.allocated
  10. FROM a
  11. LEFT JOIN
  12. (
  13. SELECT resourcesid sum(ifnull(allocation, 0) * 12345) allocated
  14. FROM my_resources r,
  15. a
  16. WHERE r.resourcesid = a.resourcesid
  17. GROUP BY resourcesid) c
  18. ON a.resourceid = c.resourcesid

AliSQL即将推出WITH语法,敬请期待。

总结

  1. 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。
  2. 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。
  3. 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 ^^。
  4. 使用云上数据库遇到难点(不局限于SQL问题),随时寻求阿里云原厂专家服务的帮助。