Oracle OR条件的优化与改写

一、同一字段谓词条件的or

1. 构造测试表

create table t_do as select * from dba_objects;
create index t_do_n1 on t_do(object_name);
create index t_do_n2 on t_do(owner);create table t_dt as select * from dba_tables;
create index t_dt_n1 on t_dt(table_name);create table t_di as select * from dba_indexes;
create index t_di_n1 on t_di(index_name);

2. 原SQL

SELECT object_type,object_name
FROM t_do
WHERE owner = 'SYS'
AND
( object_name = 'T3'
OR object_name IN (SELECT table_name FROM t_dt WHERE table_name LIKE 'TRIGER%')
OR object_name IN (SELECT index_name FROM t_di WHERE index_name LIKE 'IND_OBJ%')
)
order by 1;

由于or的写法，t3表虽然object_name字段有索引，选择率也高，但用不了。

3. 改写方法

对于同一个字段多个or条件的，可以直接改为字段 in (select xxx union all select xxx)

SELECT object_type,object_name
FROM t_do
WHERE owner = 'SYS'
and object_name in
(
select 'T3' from dual
union all
SELECT table_name FROM t_dt WHERE table_name LIKE 'TRIGER%'
union all
SELECT index_name FROM t_di WHERE index_name LIKE 'IND_OBJ%'
)
order by 1;

4. 补充

12.2版本中对 or 进行了加强，可以尝试添加 /*+ or_expand */ hint来优化SQL。

SELECT /*+ or_expand */ object_type,object_name
FROM t_do
WHERE owner = 'SYS'
AND
( object_name = 'T3'
OR object_name IN (SELECT table_name FROM t_dt WHERE table_name LIKE 'TRIGER%')
OR object_name IN (SELECT index_name FROM t_di WHERE index_name LIKE 'IND_OBJ%')
)
order by 1;

从执行计划可以看到，它是把三部分分别做关联后再union all，明显复杂很多，效果没有改写的好。

二、两个字段谓词条件的or

select object_name,object_type,object_id from t1 where object_name='T1' or object_id<=10;

优化方法：如果字段选择率高，两个字段分别加索引即可

三、高选择度的or改写

上面这种情况，如果两个字段的选择性可以，而且都存在索引，不论是oracle还是mysql，优化器都是会自动改写的。如果要手工改写，可以这样改：

select object_name,object_type,object_id from t1 where object_name='T1'
union all
select object_name,object_type,object_id from t1 where object_id<=10 and lnnvl(object_name='T1') ;

等价改写是用union all，而不是union，既不等价，效率又低

这里跟第一部分“同一字段谓词条件的or” 的区别是：or的两边是两个不同的字段，两个条件数据可能会有交集，因此不能只用 union all，还需要加 lnnvl(object_name='T1') 去除查询中的重复值。注意也不要写成object_name<>'T1'，两种不等价。

lnnvl用于某个语句的where子句中的条件，如果条件为true就返回false；如果条件为UNKNOWN或者false就返回true。非oracle库没有lnnvl函数，其实 lnnvl(object_name='T1') 就相当于(object_name<>'T1' or object_name is null)。

四、 or与半连接 in/exists

select object_name,object_type,object_id from t1 where object_name='T1' or exists (select 1 from t2 where t1.object_id=t2.object_id and t2.object_id<=10);

--------------------------------------------------------------------------------------------------------------------------
| Id  | Operation          | Name      | Starts | E-Rows |E-Bytes| Cost (%CPU)| E-Time  | A-Rows |   A-Time   | Buffers |
--------------------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT   |           |      1 |        |       |   112 (100)|         |       9 |00:00:00.01 |     424 |
|*  1 |  FILTER            |           |      1 |        |       |            |         |       9 |00:00:00.01 |     424 |
|   2 |   TABLE ACCESS FULL| T1        |      1 |  23754 |  2134K|   112   (1)| 00:00:01 |  22943 |00:00:00.01 |     410 |
|*  3 |   FILTER           |           |  22943 |        |       |            |         |       9 |00:00:00.01 |      14 |
|*  4 |    INDEX RANGE SCAN| IND_T2_N2 |      9 |      1 |    13 |     1   (0)| 00:00:01 |      9 |00:00:00.01 |      14 |
--------------------------------------------------------------------------------------------------------------------------

这种情况优化器就不会自动帮你改写了，需要手动改写（12.2版本及以上可以自动使用or_expand做查询转换)。

本质上改写的思想还是：小结果用nest loop，大结果集用hash join。

1. 小结果集的改写

object_name='T1' 结果集小，还是按照上面的思路：

select object_name,object_type,object_id from t1 where object_name='T1'
union all
select object_name,object_type,object_id from t1 where exists (select 1 from t2 where t1.object_id=t2.object_id and t2.object_id<=10) and lnnvl(object_name='T1');

-----------------------------------------------------------------------------------------------------------------------------------------------------
| Id  | Operation                            | Name      | Starts | E-Rows |E-Bytes| Cost (%CPU)| E-Time   | A-Rows |   A-Time   | Buffers | Reads  |
-----------------------------------------------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT                     |           |      1 |        |       |    6 (100)|           |      9 |00:00:00.01 |      15 |      1 |
|   1 |  UNION-ALL                           |           |      1 |        |       |      |       9 |00:00:00.01 |      15 |      1 |
|   2 |   TABLE ACCESS BY INDEX ROWID BATCHED| T1        |      1 |      1 |    92 |    1   (0)| 00:00:01 |       0 |00:00:00.01 |       2 |      0 |
|*  3 |    INDEX RANGE SCAN                  | IND_T1_N1 |      1 |      1 |       |    1   (0)| 00:00:01 |       0 |00:00:00.01 |       2 |      0 |
|   4 |   NESTED LOOPS                       |           |      1 |      1 |   105 |    5  (20)| 00:00:01 |       9 |00:00:00.01 |      13 |      1 |
|   5 |    NESTED LOOPS                      |           |      1 |      1 |   105 |    5  (20)| 00:00:01 |       9 |00:00:00.01 |      11 |      1 |
|   6 |     VIEW                             | VW_SQ_1   |      1 |      9 |   117 |    2   (0)| 00:00:01 |       9 |00:00:00.01 |       2 |      0 |
|   7 |      HASH UNIQUE                     |           |      1 |      1 |   117 |      |       9 |00:00:00.01 |       2 |      0 |
|*  8 |       INDEX RANGE SCAN               | IND_T2_N2 |      1 |      9 |   117 |    2   (0)| 00:00:01 |       9 |00:00:00.01 |       2 |      0 |
|*  9 |     INDEX RANGE SCAN                 | IND_T1_N2 |      9 |      1 |       |    1   (0)| 00:00:01 |       9 |00:00:00.01 |       9 |      1 |
|* 10 |    TABLE ACCESS BY INDEX ROWID       | T1        |      9 |      1 |    92 |    2   (0)| 00:00:01 |       9 |00:00:00.01 |       2 |      0 |
-----------------------------------------------------------------------------------------------------------------------------------------------------

2. 大结果集的改写

object_name='T1' 结果集大

原SQL

select count(owner) from t1 where object_name LIKE 'W%' or exists (select 1 from t2 where t1.object_id=t2.object_id and owner='SYS');

COUNT(OWNER)
------------
14866

---------------------------------------------------------------------------------------------------------------------------------------------
| Id  | Operation                             | Name      | Starts | E-Rows |E-Bytes| Cost (%CPU)| E-Time   | A-Rows |   A-Time   | Buffers |
---------------------------------------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT                      |           |      1 |        |       |  112 (100)|           |      1 |00:00:00.06 |   43900 |
|   1 |  SORT AGGREGATE                       |           |      1 |      1 |   145 |   |           |      1 |00:00:00.06 |   43900 |
|*  2 |   FILTER                              |           |      1 |        |       |   |           |  14866 |00:00:00.06 |   43900 |
|   3 |    TABLE ACCESS FULL                  | T1        |      1 |  23754 |  3363K|  112   (1)| 00:00:01 |  22943 |00:00:00.01 |      409 |
|*  4 |    TABLE ACCESS BY INDEX ROWID BATCHED| T2        |  21528 |    132 | 10428 |    2   (0)| 00:00:01 |  13451 |00:00:00.04 |   43491 |
|*  5 |     INDEX RANGE SCAN                  | IND_T2_N2 |  21528 |     80 |       |    1   (0)| 00:00:01 |  21527 |00:00:00.02 |   21964 |
---------------------------------------------------------------------------------------------------------------------------------------------

改写思路

大结果集用hash join --> 改写为表关联
inner join 还是 left join？ --> left join，因为有or条件，必须保证 object_name LIKE 'W%' 部分的条件不被关联条件过滤掉（这跟单纯的半连接改写是不一样的）。因此t1表必须是主表，关联后返回其所有数据。
如何实现exists条件 --> 关联后t2.object_id is not null，就是关联后有匹配上的（ t1和t2表中均存在的）
半关联自带去重 --> t2表的子查询要加 distinct

改写后语句

select count(owner)
from t1
left join (select distinct object_id from t2 where owner='SYS') v_t2
on t1.object_id=v_t2.object_id
where ( t1.object_name='EMP' or v_t2.object_id is not null);

SQL> select count(owner)2  from t13  left join (select distinct object_id from t2 where owner='SYS') v_t24  on t1.object_id=v_t2.object_id5  where ( t1.object_name like 'W%' or v_t2.object_id is not null);COUNT(OWNER)
------------14866----------------------------------------------------------------------------------------------------------------------------------
| Id  | Operation               | Name | Starts | E-Rows |E-Bytes|E-Temp | Cost (%CPU)| E-Time   | A-Rows |   A-Time   | Buffers |
----------------------------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT        |      |      1 |        |       |       |   474 (100)| |       1 |00:00:00.01 |     840 |
|   1 |  SORT AGGREGATE         |      |      1 |      1 |   158 |       |            | |       1 |00:00:00.01 |     840 |
|*  2 |   FILTER                |      |      1 |        |       |       |            | |  14866 |00:00:00.01 |     840 |
|*  3 |    HASH JOIN RIGHT OUTER|      |      1 |  23754 |  3665K|       |   474   (1)| 00:00:01 |  22943 |00:00:00.01 |     840 |
|   4 |     VIEW                |      |      1 |  13210 |   167K|       |   362   (1)| 00:00:01 |  14556 |00:00:00.01 |     431 |
|   5 |      HASH UNIQUE        |      |      1 |  13210 |  1019K|  1160K|   362   (1)| 00:00:01 |  14556 |00:00:00.01 |     431 |
|*  6 |       TABLE ACCESS FULL | T2   |      1 |  13210 |  1019K|       |   118   (1)| 00:00:01 |  14556 |00:00:00.01 |     431 |
|   7 |     TABLE ACCESS FULL   | T1   |      1 |  23754 |  3363K|       |   112   (1)| 00:00:01 |  22943 |00:00:00.01 |     409 |
----------------------------------------------------------------------------------------------------------------------------------

3. 实例验证

下面我们通过一个简化的例子，看看大结果改写中间每一步的结果输出，加深理解。

构造测试表

SQL> create table ta(id int,name varchar2(10),owner varchar2(10));Table created.SQL> create table tb(id int,name varchar2(10),owner varchar2(10));Table created.SQL> select * from ta;ID NAME       OWNER
---------- ---------- ----------1 A          SYS2 B          TMP3 W          SYSSQL> select * from tb;ID NAME       OWNER
---------- ---------- ----------1 A          SYS2 B          TMP3 W          TMP4 C          TMP5 D          SYS

原SQL

select * from ta where name LIKE 'W%'
or exists (select 1 from tb where ta.id=tb.id and owner='SYS');       ID NAME       OWNER
---------- ---------- ----------1 A          SYS3 W          SYS

name LIKE 'W%'：ta表id=3这行一定要保留

tb.owner='SYS'：tb表保留id=1和5

ta.id=tb.id：只有 id=1 符合

最后ta表返回id=1和3的数据

改写为 inner join 会怎么样？

下面就是当做普通的半连接来改写：

SQL> select * from ta join (select distinct id from tb where owner='SYS') v_tb on ta.id=v_tb.id where (ta.name like 'W%' or v_tb.id is not null);

ID NAME OWNER ID
---------- ---------- ---------- ----------
1 A SYS 1

子查询：b表返回 id=1,5

inner join：on优先级比where高，按照ta.id=v_tb.id ，a表一下只剩id=1（其实就错在这步，ta有or条件，应该作为主表，不论能不能匹配到tb，都应该返回所有数据）

where 条件：匹配到 v_tb.id is not null 该行返回；没有 ta.name like 'W%' 的行了，忽略

改写为 left join 会怎么样？

首先我们不加 v_tb.id is not null，可以看到根据left join含义 ta表的所有行都被保留了下来。但是 id=2 这一行，既不符合tb表中存在，又不符合 ta.name like 'W%'，所以应该是要过滤掉的。

SQL> select * from ta left join (select distinct id from tb where owner='SYS') v_tb on ta.id=v_tb.id;

ID NAME OWNER ID
---------- ---------- ---------- ----------
1 A SYS 1
2 B TMP
3 W SYS

增加条件 v_tb.id is not null 仅保留tb匹配行，最终返回结果与原sql一致

SQL> select * from ta left join (select distinct id from tb where owner='SYS') v_tb on ta.id=v_tb.id where (ta.name like 'W%' or v_tb.id is not null);

ID NAME OWNER ID
---------- ---------- ---------- ----------
1 A SYS 1
3 W SYS

五、 update语句中的or

原SQL是这样的，expiretime字段有索引，但OR的写法导致无法用到，绝大部分数据只能在回表时做filter，效率很低。

1. 测试案例

我们做一个简化版的例子：

create table t3(id int,name varchar2(20),state int);insert into t3 values(1,'aaa',0);
insert into t3 values(2,'bbb',0);
insert into t3 values(3,'ccc',0);
commit;create index ind_t3_id on t3(id);select * from t3;

测试update

update t3 set id=id+10 where id=1 or (id>=10 and state=0);

查看更新后结果

2. 直接但不等价的改写

根据前面select or的改写，最直接的想法就是把它改成union all + lnnvl形式，又因为update没有union all的语法，这里拆成了2个update。

update t3 set id=id+10 where id=1
-- union all
update t3 set id=id+10 where (id>=10 and state=0) and lnnvl(id=1);

这种写法在把id更新为固定值（比如100）时没有问题，但如果像上面更新成id+10或者时间字段更新成sysdate，第二条update就可能会重复更新数据。

3. 等价改写

可以改写为merge语句，把union all + lnnvl 格式语句放在using部分，比如这样

merge into t3 b
using (
select * from t3 where id=1
union all
select * from t3 where (id>=10 and state=0) and lnnvl(id=1)
) a
on (b.id=a.id)
when matched then
update set b.id = b.id+10;

但是一执行发现会报错

我们在之前的文章 Oracle 关联更新 update_Hehuyi_In的博客-CSDN博客中学习过这种“更新关联字段的merge写法”，需要将关联字段改为rowid。这里因为只涉及t3一个表，直接在using部分加上rowid字段即可。

merge into t3 b
using (
select rowid as rid,t3.* from t3 where id=1
union all
select rowid as rid,t3.* from t3 where (id>=10 and state=0) and lnnvl(id=1)
) a
on (b.rowid=a.rid)
when matched then
update set b.id = b.id+10;

六、关联条件的or改写

1. 构造测试表

我们再加两个测试表

create table t4(id int,name varchar2(20),state int);insert into t4 values(1,'ddd',0);
insert into t4 values(10,'eee',0);
insert into t4 values(20,'fff',0);
commit;create table t5(id int,name varchar2(20),state int);insert into t5 values(1,'ggg',0);
insert into t5 values(5,'hhh',0);
insert into t5 values(20,'iii',0);
commit;

2. 关联条件中用or

select t3.id,t4.id,t5.id from t3
full join t4
on t3.id = t4.id
left join t5
on t3.id = t5.id or t4.id = t5.id;

查看执行计划（11g版本），外层只能用nest loop join，全关联返回结果集大时效率低。

3. 改写方法

还是按照之前的思路，结果集大时，通过改写让它走hash join，这里巧妙地把or条件改为了nvl。

select t3.id,t4.id,t5.id from t3
full join t4
on t3.id = t4.id
left join t5
on nvl(t3.id,t4.id) = t5.id;

改写后的SQL走上了hash join

这里补充一下，在19c版本测试时，发现原SQL外层查询走的也不再是nest loop join，而是merge join，应该是优化器又增强了。

参考

Oracle 表关联、半关联、反关联_Hehuyi_In的博客-CSDN博客

时隔4年，重新分析并修正一个update SQL的优化方法

Oracle 关联更新 update_Hehuyi_In的博客-CSDN博客

《SQL写法与改写-第二期》