作/译者:叶金荣(imysql#imysql.com>),来源:http://imysql.com,欢迎转载。
起因:在innodb表上做count(*)统计实在是太慢了,因此想办法看能不能再快点。
现象:先来看几个测试案例,如下
一、 sbtest 表上的测试
show create table sbtest\G
*************************** 1. row ***************************
Table: sbtest
Create Table: CREATE TABLE `sbtest` (
`aid` bigint(20) unsigned NOT NULL auto_increment,
`id` int(10) unsigned NOT NULL default '0',
`k` int(10) unsigned NOT NULL default '0',
`c` char(120) NOT NULL default '',
`pad` char(60) NOT NULL default '',
PRIMARY KEY (`aid`),
KEY `k` (`k`),
KEY `id` (`id`)
) ENGINE=InnoDB AUTO_INCREMENT=1000001 DEFAULT CHARSET=latin1
show index from sbtest;
+--------+------------+----------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+
| Table | Non_unique | Key_name | Seq_in_index | Column_name | Collation | Cardinality | Sub_part | Packed | Null | Index_type | Comment |
+--------+------------+----------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+
| sbtest | 0 | PRIMARY | 1 | aid | A | 1000099 | NULL | NULL | | BTREE | |
| sbtest | 1 | k | 1 | k | A | 18 | NULL | NULL | | BTREE | |
| sbtest | 1 | id | 1 | id | A | 1000099 | NULL | NULL | | BTREE | |
+--------+------------+----------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+
填充了 1000万条 记录。
1、 直接 count(*)
explain SELECT COUNT(*) FROM sbtest;
+----+-------------+--------+-------+---------------+---------+---------+------+---------+-------------+
| id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
+----+-------------+--------+-------+---------------+---------+---------+------+---------+-------------+
| 1 | SIMPLE | sbtest | index | NULL | PRIMARY | 8 | NULL | 1000099 | Using index |
+----+-------------+--------+-------+---------------+---------+---------+------+---------+-------------+
SELECT COUNT(*) FROM sbtest;
+----------+
| COUNT(*) |
+----------+
| 1000000 |
+----------+
1 row in set (1.42 sec)
可以看到,如果不加任何条件,那么优化器优先采用 primary key 来进行扫描。
2、count(*) 使用 primary key 字段做条件
explain SELECT COUNT(*) FROM sbtest WHERE aid>=0;
+----+-------------+--------+-------+---------------+---------+---------+------+--------+--------------------------+
| id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
+----+-------------+--------+-------+---------------+---------+---------+------+--------+--------------------------+
| 1 | SIMPLE | sbtest | range | PRIMARY | PRIMARY | 8 | NULL | 485600 | Using where; Using index |
+----+-------------+--------+-------+---------------+---------+---------+------+--------+--------------------------+
SELECT COUNT(*) FROM sbtest WHERE aid>=0;
+----------+
| COUNT(*) |
+----------+
| 1000000 |
+----------+
1 row in set (1.39 sec)
可以看到,尽管优化器认为只需要扫描 485600 条记录(其实是索引),比刚才少多了,但其实仍然要做全表(索引)扫描。因此耗时和第一种相当。
3、 count(*) 使用 secondary index 字段做条件
explain SELECT COUNT(*) FROM sbtest WHERE id>=0;
+----+-------------+--------+-------+---------------+------+---------+------+--------+--------------------------+
| id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
+----+-------------+--------+-------+---------------+------+---------+------+--------+--------------------------+
| 1 | SIMPLE | sbtest | range | id | id | 4 | NULL | 500049 | Using where; Using index |
+----+-------------+--------+-------+---------------+------+---------+------+--------+--------------------------+
SELECT COUNT(*) FROM sbtest WHERE id>=0;
+----------+
| COUNT(*) |
+----------+
| 1000000 |
+----------+
1 row in set (0.43 sec)
可以看到,采用这种方式查询会非常快。
有人也许会问了,会不会是因为 id 字段的长度比 aid 字段的长度来的小,导致它扫描起来比较快呢?先不着急下结论,咱们来看看下面的测试例子。
二、 sbtest1 表上的测试
show create table sbtest1\G
*************************** 1. row ***************************
Table: sbtest1
Create Table: CREATE TABLE `sbtest1` (
`aid` int(10) unsigned NOT NULL AUTO_INCREMENT,
`id` bigint(20) unsigned NOT NULL DEFAULT '0',
`k` int(10) unsigned NOT NULL DEFAULT '0',
`c` char(120) NOT NULL DEFAULT '',
`pad` char(60) NOT NULL DEFAULT '',
PRIMARY KEY (`aid`),
KEY `k` (`k`),
KEY `id` (`id`)
) ENGINE=InnoDB AUTO_INCREMENT=1000001 DEFAULT CHARSET=latin1
show index from sbtest1;
+---------+------------+----------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+
| Table | Non_unique | Key_name | Seq_in_index | Column_name | Collation | Cardinality | Sub_part | Packed | Null | Index_type | Comment |
+---------+------------+----------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+
| sbtest1 | 0 | PRIMARY | 1 | aid | A | 1000099 | NULL | NULL | | BTREE | |
| sbtest1 | 1 | k | 1 | k | A | 18 | NULL | NULL | | BTREE | |
| sbtest1 | 1 | id | 1 | id | A | 1000099 | NULL | NULL | | BTREE | |
+---------+------------+----------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+
这个表里,把 aid 和 id 的字段长度调换了一下,也填充了 1000万条 记录。
1、 直接 count(*)
explain SELECT COUNT(*) FROM sbtest1;
+----+-------------+---------+-------+---------------+---------+---------+------+---------+-------------+
| id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
+----+-------------+---------+-------+---------------+---------+---------+------+---------+-------------+
| 1 | SIMPLE | sbtest1 | index | NULL | PRIMARY | 4 | NULL | 1000099 | Using index |
+----+-------------+---------+-------+---------------+---------+---------+------+---------+-------------+
SELECT COUNT(*) FROM sbtest1;
+----------+
| COUNT(*) |
+----------+
| 1000000 |
+----------+
1 row in set (1.42 sec)
可以看到,如果不加任何条件,那么优化器优先采用 primary key 来进行扫描。
2、count(*) 使用 primary key 字段做条件
explain SELECT COUNT(*) FROM sbtest1 WHERE aid>=0;
+----+-------------+---------+-------+---------------+---------+---------+------+--------+--------------------------+
| id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
+----+-------------+---------+-------+---------------+---------+---------+------+--------+--------------------------+
| 1 | SIMPLE | sbtest1 | range | PRIMARY | PRIMARY | 4 | NULL | 316200 | Using where; Using index |
+----+-------------+---------+-------+---------------+---------+---------+------+--------+--------------------------+
1 row in set (0.00 sec)
SELECT COUNT(*) FROM sbtest1 WHERE aid>=0;
+----------+
| COUNT(*) |
+----------+
| 1000000 |
+----------+
1 row in set (1.42 sec)
可以看到,尽管优化器认为只需要扫描 485600 条记录(其实是索引),比刚才少多了,但其实仍然要做全表(索引)扫描。因此耗时和第一种相当。
3、 count(*) 使用 secondary index 字段做条件
explain SELECT COUNT(*) FROM sbtest1 WHERE id>=0;
+----+-------------+---------+-------+---------------+------+---------+------+--------+--------------------------+
| id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
+----+-------------+---------+-------+---------------+------+---------+------+--------+--------------------------+
| 1 | SIMPLE | sbtest1 | range | id | id | 8 | NULL | 500049 | Using where; Using index |
+----+-------------+---------+-------+---------------+------+---------+------+--------+--------------------------+
1 row in set (0.00 sec)
SELECT COUNT(*) FROM sbtest1 WHERE id>=0;
+----------+
| COUNT(*) |
+----------+
| 1000000 |
+----------+
1 row in set (0.45 sec)
可以看到,采用这种方式查询会非常快。
上面的所有测试,均在 mysql 5.1.24 环境下通过,并且每次查询前都重启了 mysqld。
可以看到,把 aid 和 id 的长度调换之后,采用 secondary index 查询仍然是要比用 primary key 查询来的快很多。看来主要不是字段长度引起的索引扫描快慢,而是采用 primary key 以及 secondary index 引起的区别。那么,为什么用 secondary index 扫描反而比 primary key 扫描来的要快呢?我们就需要了解innodb的 clustered index 和secondary index 之间的区别了。
innodb 的 clustered index 是把 primary key 以及 row data 保存在一起的,而 secondary index 则是单独存放,然后有个指针指向 primary key。因此,需要进行 count(*) 统计表记录总数时,利用 secondary index 扫描起来,显然更快。而primary key则主要在扫描索引,同时要返回结果记录时的作用较大,例如:
SELECT * FROM sbtest WHERE aid = xxx;
那既然是使用 secondary index 会比 primary key 更快,为何优化器却优先选择 primary key 来扫描呢,Heikki Tuuri 的回答是:
in the example table, the secondary index is inserted into in a perfect order! That is
very unusual. Normally the secondary index would be fragmented, causing random disk I/O,
and the scan would be slower than in the primary index.
I am changing this to a feature request: keep 'clustering ratio' statistics on a secondary
index and do the scan there if the order is almost the same as in the primary index. I
doubt this feature will ever be implemented, though.
详情请看:这个 bug,以及这个文章:InnoDB Row Counting using Indexes。
最后感谢老杨的帮助。