爱编程的太阳 2019-06-13 11:42:19

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;

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;

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'

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);

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 |
+----+--------------------+-------+-------+---------------+---------+---------+-------+------+-----------------------------------------------------+

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

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 |
+----+-------------+-----------+------+---------------+-------+---------+-------+------+--------------------+

4. 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

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) |
+----+-------------+-------+--------+---------------+---------+---------+-----------------+--------+----------------------------------------------------+

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

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 |
+----+-------------+------------+--------+---------------+---------+---------+-------+--------+----------------------------------------------------+

5. 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

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

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

6、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。