MySQL分表优化试验

http://blog.chinaunix.net/u/29134/showart_1332349.html

MySQL分表优化试验

我们的项目中有好多不等于的情况。今天写这篇文章简单的分析一下怎么个优化法。
这里的分表逻辑是根据t_group表的user_name组的个数来分的。
因为这种情况单独user_name字段上的索引就属于烂索引。起不了啥名明显的效果。

1、试验PROCEDURE.
DELIMITER $$
DROP PROCEDURE `t_girl`.`sp_split_table`$$
CREATE  PROCEDURE `t_girl`.`sp_split_table`()
BEGIN
  declare done int default 0;
  declare v_user_name varchar(20) default '';
  declare v_table_name varchar(64) default '';
  -- Get all users' name.
  declare cur1 cursor for select user_name from t_group group by user_name;
  -- Deal with error or warnings.
  declare continue handler for 1329 set done = 1;
  -- Open cursor.
  open cur1;
  while done <> 1
  do
    fetch cur1 into v_user_name;
    if not done then
      -- Get table name.
      set v_table_name = concat('t_group_',v_user_name);
      -- Create new extra table.
      set @stmt = concat('create table ',v_table_name,' like t_group');
      prepare s1 from @stmt;
      execute s1;
      drop prepare s1;
      -- Load data into it.
      set @stmt = concat('insert into ',v_table_name,' select * from t_group where user_name = ''',v_user_name,'''');
      prepare s1 from @stmt;
      execute s1;
      drop prepare s1;
    end if;
  end while;
  -- Close cursor.
  close cur1;
  -- Free variable from memory.
  set @stmt = NULL;
END$$

DELIMITER ;
2、试验表。
我们用一个有一千万条记录的表来做测试。

mysql> select count(*) from t_group;
+----------+
| count(*) |
+----------+
| 10388608 |
+----------+
1 row in set (0.00 sec)

表结构。
mysql> desc t_group;
+-------------+------------------+------+-----+-------------------+----------------+
| Field       | Type             | Null | Key | Default           | Extra          |
+-------------+------------------+------+-----+-------------------+----------------+
| id          | int(10) unsigned | NO   | PRI | NULL              | auto_increment |
| money       | decimal(10,2)    | NO   |     |                   |                |
| user_name   | varchar(20)      | NO   | MUL |                   |                |
| create_time | timestamp        | NO   |     | CURRENT_TIMESTAMP |                |
+-------------+------------------+------+-----+-------------------+----------------+
4 rows in set (0.00 sec)

索引情况。

mysql> show index from t_group;
+---------+------------+------------------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+
| Table   | Non_unique | Key_name         | Seq_in_index | Column_name | Collation | Cardinality | Sub_part | Packed | Null | Index_type | Comment |
+---------+------------+------------------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+
| t_group |          0 | PRIMARY          |            1 | id          | A         |    10388608 |     NULL | NULL   |      | BTREE      |         |
| t_group |          1 | idx_user_name    |            1 | user_name   | A         |           8 |     NULL | NULL   |      | BTREE      |         |
| t_group |          1 | idx_combination1 |            1 | user_name   | A         |           8 |     NULL | NULL   |      | BTREE      |         |
| t_group |          1 | idx_combination1 |            2 | money       | A         |        3776 |     NULL | NULL   |      | BTREE      |         |
+---------+------------+------------------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+
4 rows in set (0.00 sec)

PS:
idx_combination1 这个索引是必须的，因为要对user_name来GROUP BY。此时属于松散索引扫描！当然完了后你可以干掉她。
idx_user_name 这个索引是为了加快单独执行constant这种类型的查询。
我们要根据用户名来分表。

mysql> select user_name from t_group where 1 group by user_name;
+-----------+
| user_name |
+-----------+
| david     |
| leo       |
| livia     |
| lucy      |
| sarah     |
| simon     |
| sony      |
| sunny     |
+-----------+
8 rows in set (0.00 sec)

所以结果表应该是这样的。
mysql> show tables like 't_group_%';
+------------------------------+
| Tables_in_t_girl (t_group_%) |
+------------------------------+
| t_group_david                |
| t_group_leo                  |
| t_group_livia                |
| t_group_lucy                 |
| t_group_sarah                |
| t_group_simon                |
| t_group_sony                 |
| t_group_sunny                |
+------------------------------+
8 rows in set (0.00 sec)

3、对比结果。


mysql> select count(*) from t_group where user_name = 'david';
+----------+
| count(*) |
+----------+
|  1298576 |
+----------+
1 row in set (1.71 sec)

执行了将近2秒。

mysql> select count(*) from t_group_david;
+----------+
| count(*) |
+----------+
|  1298576 |
+----------+
1 row in set (0.00 sec)
几乎是瞬间的。

mysql> select count(*) from t_group where user_name <> 'david';
+----------+
| count(*) |
+----------+
|  9090032 |
+----------+
1 row in set (9.26 sec)
执行了将近10秒，可以想象，这个是实际的项目中是不能忍受的。
mysql> select (select count(*) from t_group) - (select count(*) from t_group_david) as total;
+---------+
| total   |
+---------+
| 9090032 |
+---------+
1 row in set (0.00 sec)
几乎是瞬间的。


我们来看看聚集函数。
对于原表的操作。

mysql> select min(money),max(money) from t_group where user_name = 'david';
+------------+------------+
| min(money) | max(money) |
+------------+------------+
|      -6.41 |     500.59 |
+------------+------------+
1 row in set (0.00 sec)
最小，最大值都是FULL INDEX SCAN。所以是瞬间的。
mysql> select sum(money),avg(money) from t_group where user_name = 'david';
+--------------+------------+
| sum(money)   | avg(money) |
+--------------+------------+
| 319992383.84 | 246.417910 |
+--------------+------------+
1 row in set (2.15 sec)
其他聚集函数的结果就不是FULL INDEX SCAN了。耗时2.15秒。

对于小表的操作。
mysql> select min(money),max(money) from t_group_david;
+------------+------------+
| min(money) | max(money) |
+------------+------------+
|      -6.41 |     500.59 |
+------------+------------+
1 row in set (1.50 sec)
最大最小值完全是FULL TABLE SCAN,耗时1.50秒，不划算。以此看来。
mysql> select sum(money),avg(money) from t_group_david;
+--------------+------------+
| sum(money)   | avg(money) |
+--------------+------------+
| 319992383.84 | 246.417910 |
+--------------+------------+
1 row in set (1.68 sec)

取得这两个结果也是花了快2秒，快了一点。

我们来看看这个小表的结构。
mysql> desc t_group_david;
+-------------+------------------+------+-----+-------------------+----------------+
| Field       | Type             | Null | Key | Default           | Extra          |
+-------------+------------------+------+-----+-------------------+----------------+
| id          | int(10) unsigned | NO   | PRI | NULL              | auto_increment |
| money       | decimal(10,2)    | NO   |     |                   |                |
| user_name   | varchar(20)      | NO   | MUL |                   |                |
| create_time | timestamp        | NO   |     | CURRENT_TIMESTAMP |                |
+-------------+------------------+------+-----+-------------------+----------------+
4 rows in set (0.00 sec)

明显的user_name属性是多余的。那么就干掉它。
mysql> alter table t_group_david drop user_name;
Query OK, 1298576 rows affected (7.58 sec)
Records: 1298576  Duplicates: 0  Warnings: 0

现在来重新对小表运行查询

mysql> select min(money),max(money) from t_group_david;
+------------+------------+
| min(money) | max(money) |
+------------+------------+
|      -6.41 |     500.59 |
+------------+------------+
1 row in set (0.00 sec)

此时是瞬间的。
mysql> select sum(money),avg(money) from t_group_david;
+--------------+------------+
| sum(money)   | avg(money) |
+--------------+------------+
| 319992383.84 | 246.417910 |
+--------------+------------+
1 row in set (0.94 sec)

这次算是控制在一秒以内了。

mysql> Aborted

小总结一下：分出的小表的属性尽量越少越好。大胆的去干吧。