使用过oracle或者其他关系数据库的DBA或者开发人员都有这样的经验,在子查询上都认为数据库已经做过优化,能够很好的选择驱动表执行,然后在把该经验移植到mysql数据库上,但是不幸的是,mysql在子查询的处理上有可能会让你大失所望,在我们的生产系统上就碰到过一些案例,例如:
SELECT i_id, sum(i_sell) AS i_sellFROM table_dataWHERE i_id IN(SELECT i_id FROM table_data WHERE Gmt_create >= '2011-10-07 00:00:00')GROUP BY i_id;
(备注:sql的业务逻辑可以打个比方:先查询出10-07号新卖出的100本书,然后在查询这新卖出的100本书在全年的销量情况)。
这条sql之所以出现的性能问题在于mysql优化器在处理子查询的弱点
,mysql优化器在处理子查询的时候,会将将子查询改写。通常情况下,我们希望由内到外,先完成子查询的结果,然后在用子查询来驱动外查询的表,完成查询;但是mysql处理为将会先扫描外面表中的所有数据,每条数据将会传到子查询中与子查询关联,如果外表很大的话,那么性能上将会出现问题;
针对上面的查询,由于table_data这张表的数据有70W的数据,同时子查询中的数据较多,有大量是重复的,这样就需要关联近70W次,大量的关联导致这条sql执行了几个小时也没有执行完成,所以我们需要改写sql:
SELECT t2.i_id, SUM(t2.i_sell) AS soldFROM(SELECT DISTINCT i_id FROM table_data WHERE gmt_create >= '2011-10-07 00:00:00') t1,table_data t2WHERE t1.i_id = t2.i_idGROUP BY t2.i_id;
我们将子查询改为了关联,同时在子查询中加上distinct,减少t1关联t2的次数;
改造后,sql的执行时间降到100ms以内。
mysql的子查询的优化一直不是很友好,一直有受业界批评比较多,也是我在sql优化中遇到过最多的问题之一,mysql在处理子查询的时候,会将子查询改写,通常情况下,我们希望由内到外,也就是先完成子查询的结果,然后在用子查询来驱动外查询的表,完成查询,但是恰恰相反,子查询不会先被执行;今天希望通过介绍一些实际的案例来加深对mysql子查询的理解。下面将介绍一个完整的案例及其分析、调优的过程与思路。
用户反馈数据库响应较慢,许多业务动更新被卡住;登录到数据库中观察,发现长时间执行的sql;
| 10437 | usr0321t9m9 | 10.242.232.50:51201 | oms | Execute | 1179 | SendingSql为:SELECT tradedto0_.*FROM a1 tradedto0_WHERE tradedto0_.tradestatus='1'AND (tradedto0_.tradeoid IN (SELECT orderdto1_.tradeoidFROM a2 orderdto1_WHERE orderdto1_.proname LIKE '%??%'OR orderdto1_.procode LIKE '%??%'))AND tradedto0_.undefine4='1'AND tradedto0_.invoicetype='1'AND tradedto0_.tradestep='0'AND (tradedto0_.orderCompany LIKE '0002%')ORDER BY tradedto0_.tradesign ASC, tradedto0_.makertime DESC LIMIT 15;
UPDATE a1SET tradesign='DAB67634-795C-4EAC-B4A0-78F0D531D62F',markColor=' #CD5555',memotime='2012-09- 22', markPerson='??'WHERE tradeoid IN ('gy2012092204495100032') ;
为了尽快恢复应用,将其长时间执行的sql kill掉后,应用恢复正常;
db@3306 :explainSELECT tradedto0_.*FROM a1 tradedto0_WHERE tradedto0_.tradestatus='1'AND (tradedto0_.tradeoid IN (SELECT orderdto1_.tradeoid FROM a2 orderdto1_ WHERE orderdto1_.proname LIKE '%??%' OR orderdto1_.procode LIKE '%??%'))AND tradedto0_.undefine4='1'AND tradedto0_.invoicetype='1'AND tradedto0_.tradestep='0'AND (tradedto0_.orderCompany LIKE '0002%')ORDER BY tradedto0_.tradesign ASC, tradedto0_.makertime DESC LIMIT 15;+----+--------------------+------------+------+---------------+------+---------+------+-------+-----| id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |+----+--------------------+------------+------+---------------+------+---------+------+-------+-----| 1 | PRIMARY | tradedto0_ | ALL | NULL | NULL | NULL | NULL | 27454 | Using where; Using filesort || 2 | DEPENDENT SUBQUERY | orderdto1_ | ALL | NULL | NULL | NULL | NULL | 40998 | Using where |+----+--------------------+------------+------+---------------+------+---------+------+-------+-----
从执行计划上,我们开始一步一步地进行优化:
首先,我们看看执行计划的第二行,也就是子查询的那部分,orderdto1_进行了全表的扫描,我们看看能不能添加适当的索引:
db@3306:alter table a2 add index ind_a2(proname,procode,tradeoid);ERROR 1071 (42000): Specified key was too long; max key length is 1000 bytes
添加组合索引超过了最大key length限制:
db@3306 :DESCa2 ;+---------------------+---------------+------+-----+---------+-------+| FIELD | TYPE| NULL | KEY | DEFAULT | Extra |+---------------------+---------------+------+-----+---------+-------+| OID | VARCHAR(50) | NO | PRI | NULL| || TRADEOID| VARCHAR(50) | YES| | NULL| || PROCODE | VARCHAR(50) | YES| | NULL| || PRONAME | VARCHAR(1000) | YES| | NULL| || SPCTNCODE | VARCHAR(200)| YES| | NULL| |
db@3306 :SELECT MAX(LENGTH(PRONAME)),avg(LENGTH(PRONAME)) FROM a2;+----------------------+----------------------+| MAX(LENGTH(PRONAME)) | avg(LENGTH(PRONAME)) |+----------------------+----------------------+|95| 24.5588 |
ALTER TABLE MODIFY COLUMN PRONAME VARCHAR(156);
再进行执行计划分析:
db@3306 :explainSELECT tradedto0_.*FROM a1 tradedto0_WHERE tradedto0_.tradestatus='1'AND (tradedto0_.tradeoid IN (SELECT orderdto1_.tradeoid FROM a2 orderdto1_ WHERE orderdto1_.proname LIKE '%??%' OR orderdto1_.procode LIKE '%??%'))AND tradedto0_.undefine4='1'AND tradedto0_.invoicetype='1'AND tradedto0_.tradestep='0'AND (tradedto0_.orderCompany LIKE '0002%')ORDER BY tradedto0_.tradesign ASC, tradedto0_.makertime DESC LIMIT 15;+----+--------------------+------------+-------+-----------------+----------------------+---------+| id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |+----+--------------------+------------+-------+-----------------+----------------------+---------+| 1 | PRIMARY | tradedto0_ | ref | ind_tradestatus | ind_tradestatus | 345 | const,const,const,const | 8962 | Using where; Using filesort || 2 | DEPENDENT SUBQUERY | orderdto1_ | index | NULL | ind_a2 | 777 | NULL | 41005 | Using where; Using index |+----+--------------------+------------+-------+-----------------+----------------------+---------+
发现性能还是上不去,关键在两个表扫描的行数并没有减小(8962*41005),上面添加的索引没有太大的效果,现在查看t表的执行结果:
db@3306 :SELECT orderdto1_.tradeoidFROM t orderdto1_WHERE orderdto1_.proname LIKE '%??%'OR orderdto1_.procode LIKE '%??%'; EmptySET (0.05 sec)
结果集为空,所以需要将t表的结果集做作为驱动表;
通过上面测试验证,普通的mysql子查询写法性能上是很差的,为mysql的子查询天然的弱点,需要将sql进行改写为关联的写法:
SELECT tradedto0_.*FROM a1 tradedto0_ ,(SELECT orderdto1_.tradeoid FROM a2 orderdto1_ WHERE orderdto1_.proname LIKE '%??%' OR orderdto1_.procode LIKE '%??%')t2WHERE tradedto0_.tradestatus='1'AND (tradedto0_.tradeoid=t2.tradeoid)AND tradedto0_.undefine4='1'AND tradedto0_.invoicetype='1'AND tradedto0_.tradestep='0'AND (tradedto0_.orderCompany LIKE '0002%')ORDER BY tradedto0_.tradesign ASC, tradedto0_.makertime DESC LIMIT 15;
db@3306 :explainSELECT tradedto0_.*FROM a1 tradedto0_ ,(SELECT orderdto1_.tradeoid FROM a2 orderdto1_ WHERE orderdto1_.proname LIKE '%??%' OR orderdto1_.procode LIKE '%??%')t2WHERE tradedto0_.tradestatus='1'AND (tradedto0_.tradeoid=t2.tradeoid)AND tradedto0_.undefine4='1'AND tradedto0_.invoicetype='1'AND tradedto0_.tradestep='0'AND (tradedto0_.orderCompany LIKE '0002%')ORDER BY tradedto0_.tradesign ASC, tradedto0_.makertime DESC LIMIT 15;+----+-------------+------------+-------+---------------+----------------------+---------+------+| id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |+----+-------------+------------+-------+---------------+----------------------+---------+------+| 1 | PRIMARY | NULL | NULL | NULL | NULL | NULL | NULL | NULL | Impossible WHERE noticed after reading const tables || 2 | DERIVED | orderdto1_ | index | NULL | ind_a2 | 777 | NULL | 41005 | Using where; Using index |+----+-------------+------------+-------+---------------+----------------------+---------+------+
db@3306 :SELECT tradedto0_.*FROM a1 tradedto0_ ,(SELECT orderdto1_.tradeoid FROM a2 orderdto1_ WHERE orderdto1_.proname LIKE '%??%' OR orderdto1_.procode LIKE '%??%')t2WHERE tradedto0_.tradestatus='1'AND (tradedto0_.tradeoid=t2.tradeoid)AND tradedto0_.undefine4='1'AND tradedto0_.invoicetype='1'AND tradedto0_.tradestep='0'AND (tradedto0_.orderCompany LIKE '0002%')ORDER BY tradedto0_.tradesign ASC, tradedto0_.makertime DESC LIMIT 15; EmptySET (0.03 sec)
缩短到了毫秒;
1. mysql子查询在执行计划上有着明显的弱点,需要将子查询进行改写
可以参考:
a. 生产库中遇到mysql的子查询:http://hidba.org/?p=412
b. 内建的builtin InnoDB,子查询阻塞更新:http://hidba.org/?p=456
2. 在表结构设计上,不要随便使用varchar(N)的大字段,导致无法使用索引
可以参考:
a. JDBC内存管理—varchar2(4000)的影响:http://hidba.org/?p=31
b. innodb中大字段的限制:http://hidba.org/?p=144
c. innodb使用大字段text,blob的一些优化建议: http://hidba.org/?p=551
[1] 生产库中遇到mysql的子查询 http://hidba.org/?p=412
[2] 浅谈mysql的子查询 http://hidba.org/?p=624
[3] mysql子查询的弱点 http://hidba.org/?p=260