SQL Tuning 基础概述06 - 表的关联方式：Nested Loops Join，Merge Sort Join & Hash Join

nested loops join（嵌套循环）  

驱动表返回几条结果集，被驱动表访问多少次，有驱动顺序，无须排序，无任何限制。

驱动表限制条件有索引，被驱动表连接条件有索引。

hints：use_nl()

merge sort join（排序合并）  

驱动表和被驱动表都是最多访问1次，无驱动顺序，需要排序（SORT_AREA_SIZE），连接条件是<>或like导致无法使用。

在连接条件上建立索引可以消除一张表的排序。

hints：use_merge()

hash join（哈希连接）  

驱动表和被驱动表都是最多访问1次，有驱动顺序，无须排序（HASH_AREA_SIZE但是会消耗内存用于建HASH表），连接条件是<> > < 或like导致无法使用。

索引列在表连接中无特殊要求，与单表情况相同。

hints：use_hash()

实验验证：

1.不同表连接的表访问次数验证

2.不同表连接的驱动顺序区别

3.不同表连接的排序情况分析

4.不同表连接的限制场景对比

5.不同表连接和索引的关系

首先，准备两张表t1,t2,分别初始化随机插入100条和100,000条数据：

drop table t1 cascade constraints purge;

drop table t2 cascade constraints purge;

create table t1( id number not null, n number, contents varchar2(4000) );

create table t2( id number not null, t1_id number not null, n number, contents varchar2(4000) );

execute dbms_random.seed(0);

insert into t1  select rownum, rownum, dbms_random.string('a',50)   from dual   connect by level <= 100   order by dbms_random.random;

commit;  

insert into t2  select rownum, rownum, rownum, dbms_random.string('b',50)  from dual  connect by level <= 100000  order by dbms_random.random;

commit;

select count(1) from t1;

select count(1) from t2;

1.不同表连接的表访问次数验证：

set linesize 1000 pagesize 200
alter session set statistics_level = all;

select * from table(dbms_xplan.display_cursor(null,null,'allstats last'));

1.1 nested loops join:

select /*+ leading(t1)use_nl(t2) */ * from t1, t2 where t1.id = t2.t1_id;

select /*+ leading(t1)use_nl(t2) */ * from t1, t2 where t1.id = t2.t1_id and t1.n in(17,19);

select /*+ leading(t1)use_nl(t2) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 19;

select /*+ leading(t1)use_nl(t2) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 99999999;

1.2 hash join:

select /*+ leading(t1)use_hash(t2) */ * from t1, t2 where t1.id = t2.t1_id;

select /*+ leading(t1)use_hash(t2) */ * from t1, t2 where t1.id = t2.t1_id and t1.n in(17,19);

select /*+ leading(t1)use_hash(t2) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 19;

select /*+ leading(t1)use_hash(t2) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 99999999;

select /*+ leading(t1)use_hash(t2) */ * from t1, t2 where t1.id = t2.t1_id and 1 = 2;

1.3 merge sort join

select /*+ ordered use_merge(t2) */ * from t1, t2 where t1.id = t2.t1_id; 

2.不同表连接的驱动顺序区别：

2.1 nested loops join

select /*+ leading(t1)use_nl(t2) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 19;

select /*+ leading(t2)use_nl(t1) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 19; 

2.2 hash join

select /*+ leading(t1)use_hash(t2) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 19;

select /*+ leading(t2)use_hash(t1) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 19; 

2.3 merge sort join

select /*+ leading(t1)use_merge(t2) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 19;

select /*+ leading(t2)use_merge(t1) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 19;

说到不同表连接表的驱动顺序，网上也有一个普遍流行的观点，就是小表作为驱动表。其实通过上面的实验可以发现这样的描述是不准确的。

正确地描述应该是：对于nested loops join和hash join来说，小的结果集先访问,大的结果集后访问（即与表的大小没有关系，与具体sql返回的结果集大小有关）；而对于merge sort join 来说，先访问谁效率都是一样的。

3.不同表连接的排序情况分析：

嵌套循环，不排序；

hash连接，消耗内存建立hash表；

排序合并，需要排序。

开发人员需要注意不要取多余的字段参与排序：

select /*+ leading(t2)use_merge(t1) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 19;

select /*+ leading(t2)use_merge(t1) */ t1.id from t1, t2 where t1.id = t2.t1_id and t1.n = 19;

select * from table(dbms_xplan.display_cursor(null,null,'allstats last'));

4.不同表连接的限制场景对比：

4.1 hash join不支持<> > < like连接条件

select /*+ leading(t1)use_hash(t2) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 19;

select /*+ leading(t1)use_hash(t2) */ * from t1, t2 where t1.id <> t2.t1_id and t1.n = 19;

select /*+ leading(t1)use_hash(t2) */ * from t1, t2 where t1.id > t2.t1_id and t1.n = 19;

select /*+ leading(t1)use_hash(t2) */ * from t1, t2 where t1.id < t2.t1_id and t1.n = 19;

select /*+ leading(t1)use_hash(t2) */ * from t1, t2 where t1.id like t2.t1_id and t1.n = 19;

select * from table(dbms_xplan.display_cursor(null,null,'allstats last'));

4.2 merge sort join不支持<> like 连接方式

select /*+ leading(t1)use_merge(t2) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 19;

select /*+ leading(t1)use_merge(t2) */ * from t1, t2 where t1.id <> t2.t1_id and t1.n = 19;

select /*+ leading(t1)use_merge(t2) */ * from t1, t2 where t1.id > t2.t1_id and t1.n = 19;

select /*+ leading(t1)use_merge(t2) */ * from t1, t2 where t1.id < t2.t1_id and t1.n = 19;

select /*+ leading(t1)use_merge(t2) */ * from t1, t2 where t1.id like t2.t1_id and t1.n = 19;

select * from table(dbms_xplan.display_cursor(null,null,'allstats last'));

4.3 nested loops join 所有都支持

从上面实验结果来看，不能走hash和merge的表连接条件，都会走nested loops join。

5.不同表连接和索引的关系：

5.1 nested loops join

驱动表的限制条件建立索引，被驱动表的连接条件建立索引。

create index idx_t1_n on t1(n);

create index idx_t2_t1_id on t2(t1_id);

select /*+ leading(t1)use_nl(t2) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 19;

select * from table(dbms_xplan.display_cursor(null,null,'allstats last'));

5.2 hash join

select /*+ leading(t1)use_hash(t2) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 19;

select * from t1, t2 where t1.id = t2.t1_id and t1.n = 19;

一般查询没有合适索引，Oracle都会选择用hash join的表连接。

5.3 merge sort join

create index idx_t1_id on t1(id);

select /*+ ordered use_merge(t2) */ * from t1, t2 where t1.id = t2.t1_id and t1.n = 19;

Oracle 10g版本，在连接条件建立索引可以消除merge sort join表连接的一张表的排序操作。(虽然在两张表的连接条件都建立了索引，却只能消除一张表的排序操作)

注：本文为《收获，不止Oracle》表连接一章的总结笔记。