一个left join SQL 简单优化分析

有个关联查询的sql,需要2秒多，于是进行查看一番：

SELECT
	a.id,
	a.brand_id,
	a.series_id,
	a.product_id,
	a.material_id,
	a.custom_cateGory_id,
	a.price,
	a.product_url,
	a.organ_id,
	.....
FROM
	pm_brand_xxxx a
LEFT JOIN pm_brand_yyyyy d ON a.series_id = d.id
WHERE
	a.is_delete = 0
AND d.is_delete = 0
AND a.organ_id = 'Cxxx'
AND a.brand_id = 6491603
AND d.brand_id = 6491603
AND a.model_flag = 14;

Mysql> show profile for query 4;
+----------------------+----------+
| Status               | Duration |
+----------------------+----------+
| starting             | 0.000072 |
| checking permissions | 0.000002 |
| checking permissions | 0.000002 |
| Opening tables       | 0.000011 |
| init                 | 0.000026 |
| System lock          | 0.000007 |
| optimizing           | 0.000016 |
| statistics           | 0.000142 |
| preparing            | 0.000018 |
| executing            | 0.000002 |
| Sending data         | 2.281192 |<<<<<<<执行的主要时间消耗
| end                  | 0.000007 |
| query end            | 0.000011 |
| closing tables       | 0.000011 |
| freeing items        | 0.000030 |
| logging slow query   | 0.000003 |
| logging slow query   | 0.000102 |
| cleaning up          | 0.000022 |
+----------------------+----------+

+----+-------------+-------+------------+-------------+---------------------------------------------------------------------------------------------------------------+---------------------------------------------------------------------------+---------+-------+-------+----------+------------------------------------------------------------------------------------------------------------------------------------------------+
| id | select_type | table | partitions | type        | possible_keys                                                                                                 | key                                                                       | key_len | ref   | rows  | filtered | Extra                                                                                                                                          |
+----+-------------+-------+------------+-------------+---------------------------------------------------------------------------------------------------------------+---------------------------------------------------------------------------+---------+-------+-------+----------+------------------------------------------------------------------------------------------------------------------------------------------------+
|  1 | SIMPLE      | d     | NULL       | ref         | PRIMARY,idx_pm_yyyy_bid                                                                                        | idx_pm_yyyyy_bid                                                        | 9       | const |     1 |    10.00 | Using where                                                                                                                                    |
|  1 | SIMPLE      | a     | NULL       | index_merge | idx_pm_xxxx_sid,idx_pm_xxx_bid,idx_pm_brand_xxxx_organ                                                         | idx_pm_xxx_organ,idx_pm_brand_xxxx_bid                                   | 99,9    | NULL  | 11314 |     0.04 | Using intersect(idx_pm_xxxxx_organ,idx_pm_xxxx_bid); Using where; Using join buffer (Block Nested Loop)                                        |
+----+-------------+-------+------------+-------------+---------------------------------------------------------------------------------------------------------------+---------------------------------------------------------------------------+---------+-------+-------+----------+------------------------------------------------------------------------------------------------------------------------------------------------+
2 rows in set, 1 warning (0.00 sec)

从执行计划来看，d表是做了驱动表，a做了被驱动表

d表 type = ref ，使用非唯一性索引或者唯一索引的前缀扫描，返回匹配某个单独值的记录行,这里使用了索引idx_pm_yyyyy_bid，该索引正是brand_id上的索引，

即是说，在和a表的关联中d先通过brand_id来查找记录行，再通过相应记录的id去和a表的series_id做匹配。

我查看相应的记录数，发现a表145万的大表,d表是4075的小表。

a表

mysql> select count(*) from pm_xxxxxx;

+----------+

| count(*) |

+----------+

|  1459777 |

+----------+

1 row in set (0.27 sec)

d表：

mysql> select count(*) from pm_yyyyyy;

+----------+

| count(*) |

+----------+

|     4075 |

+----------+

1 row in set (0.00 sec)

而 a表是type=index_merge 索引合并,这里走了idx_pm_xxx_organ(organ_id),idx_pm_brand_xxxx_bid(brand_id) ,extra 是

Using intersect(idx_pm_xxxxx_organ,idx_pm_xxxx_bid); Using where; Using join buffer (Block Nested Loop) 

Using intersect正说明了这里使用了(idx_pm_xxxxx_organ,idx_pm_xxxx_bid)的交集

Using where 是用model_flag等这些其他条件的过滤

Using join buffer (Block Nested Loop) 说明使用BNL的算法进行匹配

 BNL 算法是将外层循环的行/结果集（驱动表）存入join buffer, 内层循环的每一行与整个buffer中的记录做比较，从而减少内层循环的次数.

举例来说，外层循环的结果集是100行，使用NLJ 算法需要扫描内部表100次，如果使用BNL算法，先把对Outer Loop表(外部表)每次读取的10行记录放到join buffer,然后在InnerLoop表(内部表)中直接匹配这10行数据，内存循环就可以一次与这10行进行比较, 这样只需要比较10次，对内部表的扫描减少了9/10。所以BNL算法就能够显著减少内层循环表扫描的次数.

在这里就是d表中取得结果集分批放入buffer中与a表进行匹配。

而这个语句无论如何都要2秒中，也在我们的认识中小表驱动大表并没错，我的猜想应该就是在进行BNL时消耗了时间，表现到过程中就是 Sending data 的时间消耗增多。

吐槽的是mysql中貌似没有什么办法来多方面看查询消耗了。

我想到的是如果该表现有sql关联的顺序是否性能能改善，在该sql中，我发现了两个条件：

AND a.brand_id = 6491603

AND d.brand_id = 6491603

在业务逻辑上这两个表的字段应该是一致的，如果我将d表的d.brand_id = 6491603去掉，以上的执行计划应该会改变，于是去掉之后执行，执行时间非常小。

mysql> show profile for query 1;
+----------------------+----------+
| Status               | Duration |
+----------------------+----------+
| starting             | 0.000080 |
| checking permissions | 0.000002 |
| checking permissions | 0.000002 |
| Opening tables       | 0.000012 |
| init                 | 0.000030 |
| System lock          | 0.000006 |
| optimizing           | 0.000014 |
| statistics           | 0.000130 |
| preparing            | 0.000016 |
| executing            | 0.000001 |
| Sending data         | 0.027325 |
| end                  | 0.000003 |
| query end            | 0.000015 |
| closing tables       | 0.000005 |
| freeing items        | 0.000014 |
| cleaning up          | 0.000009 |
+----------------------+----------+
16 rows in set, 1 warning (0.00 sec)

看其执行计划：
+----+-------------+-------+------------+-------------+---------------------------------------------------------------------------------------------------------------+---------------------------------------------------------------------------+---------+-------------------------+-------+----------+---------------------------------------------------------------------------------------------------------+
| id | select_type | table | partitions | type        | possible_keys                                                                                                 | key                                                                       | key_len | ref                     | rows  | filtered | Extra                                                                                                   |
+----+-------------+-------+------------+-------------+---------------------------------------------------------------------------------------------------------------+---------------------------------------------------------------------------+---------+-------------------------+-------+----------+---------------------------------------------------------------------------------------------------------+
|  1 | SIMPLE      | a     | NULL       | index_merge | idx_pm_xxxxx_sid,idx_pm_xxxxx_bid,idx_pm_xxxx_organ                                                           | idx_pm_xxxxx_organ,idx_pm_xxxx_bid                                        | 99,9    | NULL                    | 11315 |     1.00 | Using intersect(idx_pm_xxxxx_organ,idx_pm_xxxx_bid); Using where                                        |
|  1 | SIMPLE      | d     | NULL       | eq_ref      | PRIMARY                                                                                                       | PRIMARY                                                                   | 8       | xxxx.a.series_id |     1 |    10.00 | Using where                                                                                             |
+----+-------------+-------+------------+-------------+---------------------------------------------------------------------------------------------------------------+---------------------------------------------------------------------------+---------+-------------------------+-------+----------+---------------------------------------------------------------------------------------------------------+
2 rows in set, 1 warning (0.00 sec)

发现变成了a表做驱动表，d表做被驱动表,从extra列看

a表是Using intersect(idx_pm_xxxxx_organ,idx_pm_xxxx_bid); Using where   依然是使用索引合并，where条件来取结果，使用了idx_pm_xxxxx_organ,idx_pm_xxxx_bid 连个索引。

d表走PRIMARY 主键索引，从ref列来看是通过a表的series_id 来关联，这样效率表提升了。

需要说的一点是，小结果集并不代表就是小表，大表也可以有小结果集，当大表用来被匹配并被扫描多次，自然效率并不高.