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 |
+----+-------------+-----------+------+---------------+-------+---------+-------+------+--------------------+
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
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 |
+----+-------------+------------+--------+---------------+---------+---------+-------+--------+----------------------------------------------------+
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
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
9、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 。