Oracle "脑残" CBO 优化案例

今天晚上下班回来才有空看群，群友发了一条很简单的慢SQL问怎么优化。

非常简单，我自己模拟的数据。

表结构：

-- auto-generated definition
CREATE TABLE HHHHHH
(
    ID           NUMBER NOT NULL
        PRIMARY KEY,
    NAME         VARCHAR2(20),
    PARAGRAPH_ID NUMBER
)
/

CREATE INDEX IDX_1_2_PARAGRAPH_HIST_RULE
    ON HHHHHH (PARAGRAPH_ID)
/

CREATE INDEX IDX_1_2_NAME_HIST_RULE
    ON HHHHHH (NAME)
/

数据量：

SQL> select count(1) from HHHHHH;

  COUNT(1)
----------
    200002

Elapsed: 00:00:00.00

慢SQL：

SELECT a.* FROM hhhhhh a
WHERE   a.name IN (
    SELECT   name from hhhhhh b
    GROUP BY b.name HAVING count(DISTINCT b.paragraph_id) = 1
);

Plan hash value: 1063187735

------------------------------------------------------------------------------------
| Id  | Operation              | Name      | Rows  | Bytes | Cost (%CPU)| Time     |
------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT       |           |     1 |    38 |     5  (20)| 00:00:01 |
|*  1 |  FILTER                |           |       |       |            |          |
|   2 |   TABLE ACCESS FULL    | HHHHHH    |     1 |    38 |     2   (0)| 00:00:01 |
|*  3 |   FILTER               |           |       |       |            |          |
|   4 |    HASH GROUP BY       |           |     1 |    25 |     3  (34)| 00:00:01 |
|   5 |     VIEW               | VM_NWVW_1 |     1 |    25 |     3  (34)| 00:00:01 |
|   6 |      SORT GROUP BY     |           |     1 |    25 |     3  (34)| 00:00:01 |
|   7 |       TABLE ACCESS FULL| HHHHHH    |     1 |    25 |     2   (0)| 00:00:01 |
------------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

"   1 - filter( EXISTS (SELECT 0 FROM  (SELECT ""B"".""PARAGRAPH_ID"" "
"              ""$vm_col_1"",""B"".""NAME"" ""$vm_col_2"" FROM ""HHHHHH"" ""B"" GROUP BY "
"              ""B"".""NAME"",""B"".""PARAGRAPH_ID"") ""VM_NWVW_1"" GROUP BY ""$vm_col_2"" HAVING "
"              ""$vm_col_2""=:B1 AND COUNT(""$vm_col_1"")=1))"
"   3 - filter(""$vm_col_2""=:B1 AND COUNT(""$vm_col_1"")=1)"

 跑了24秒没出结果我就干掉了，正常来说Oracle 这种遥遥领先的数据库，不能100毫秒以内出结果都有问题。

简单看了下上面的计划 Predicate Information 谓词信息，里面信息很复杂，懒得解释（其实我也不懂为啥CBO为啥这样乱分组过滤），并没啥卵用，感觉很SB。

一句话就是CBO等价改写了 EXISTS 还有 :B1这种变量，每次都是传个值到:B1 然后进行filter , 重点是每次。反正各位读者以后在计划中看到这种 :B1 变量都是每次每次，就是一次一次的传值，比较完一个数据继续传。

这种按照 PG 的说法就是复杂的子连接无法提升，  GROUP BY b.name HAVING count(DISTINCT b.paragraph_id) = 1 惹得锅。

复杂的子连接无法提升参考 <<PostgreSQL技术内幕：查询优化深度探索 >>这本书 3.2篇章。

加个HINT：

SELECT a.* FROM hhhhhh a
WHERE   a.name IN (
    SELECT  /*+ unnest */ name from hhhhhh b
    GROUP BY b.name HAVING count(DISTINCT b.paragraph_id) = 1
  5  );

    ID NAME         PARAGRAPH_ID
---------- -------------------- ------------
    200002 aaaaa            10000001

Elapsed: 00:00:00.05

Plan hash value: 3353221841

-------------------------------------------------------------------------------------
| Id  | Operation               | Name      | Rows  | Bytes | Cost (%CPU)| Time     |
-------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT        |           |     1 |    50 |     5  (20)| 00:00:01 |
|*  1 |  HASH JOIN SEMI         |           |     1 |    50 |     5  (20)| 00:00:01 |
|   2 |   TABLE ACCESS FULL     | HHHHHH    |     1 |    38 |     2   (0)| 00:00:01 |
|   3 |   VIEW                  | VW_NSO_1  |     1 |    12 |     3  (34)| 00:00:01 |
|*  4 |    FILTER               |           |       |       |            |          |
|   5 |     HASH GROUP BY       |           |     1 |    25 |     3  (34)| 00:00:01 |
|   6 |      VIEW               | VM_NWVW_2 |     1 |    25 |     3  (34)| 00:00:01 |
|   7 |       HASH GROUP BY     |           |     1 |    25 |     3  (34)| 00:00:01 |
|   8 |        TABLE ACCESS FULL| HHHHHH    |     1 |    25 |     2   (0)| 00:00:01 |
-------------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

"   1 - access(""A"".""NAME""=""NAME"")"
"   4 - filter(COUNT(""$vm_col_1"")=1)"

使用HINT将子链接强行提升（展开）以后，秒出。

但是使用HINT容易将执行计划固定住，非必要情况下不推荐。

等价改写该SQL 方式1：

SELECT A.*
FROM HHHHHH A
         INNER JOIN (SELECT COUNT(1) BB, NAME
                     FROM HHHHHH B
  5                       GROUP BY NAME) B ON A.NAME = B.NAME AND B.BB = 1;

    ID NAME         PARAGRAPH_ID
---------- -------------------- ------------
    200002 aaaaa            10000001

Elapsed: 00:00:00.03

Plan hash value: 3909860973

--------------------------------------------------------------------------------
| Id  | Operation             | Name   | Rows  | Bytes | Cost (%CPU)| Time     |
--------------------------------------------------------------------------------
|   0 | SELECT STATEMENT      |        |     1 |    50 |     5  (20)| 00:00:01 |
|*  1 |  HASH JOIN            |        |     1 |    50 |     5  (20)| 00:00:01 |
|   2 |   TABLE ACCESS FULL   | HHHHHH |     1 |    38 |     2   (0)| 00:00:01 |
|   3 |   VIEW                |        |     1 |    12 |     3  (34)| 00:00:01 |
|*  4 |    FILTER             |        |       |       |            |          |
|   5 |     HASH GROUP BY     |        |     1 |    12 |     3  (34)| 00:00:01 |
|   6 |      TABLE ACCESS FULL| HHHHHH |     1 |    12 |     2   (0)| 00:00:01 |
--------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

"   1 - access(""A"".""NAME""=""B"".""NAME"")"
   4 - filter(COUNT(*)=1)

改写成 join 以后也是秒出。

等价改写该SQL 方式2：

SELECT X.ID,
       X.NAME,
       X.PARAGRAPH_ID
FROM (SELECT A.*, COUNT(DISTINCT PARAGRAPH_ID) OVER (PARTITION BY NAME) CNT FROM HHHHHH A) X
  5  WHERE X.CNT = 1;

    ID NAME         PARAGRAPH_ID
---------- -------------------- ------------
    200002 aaaaa            10000001

Elapsed: 00:00:00.07

Plan hash value: 2750561680

------------------------------------------------------------------------------
| Id  | Operation           | Name   | Rows  | Bytes | Cost (%CPU)| Time     |
------------------------------------------------------------------------------
|   0 | SELECT STATEMENT    |        |     1 |    51 |     3  (34)| 00:00:01 |
|*  1 |  VIEW               |        |     1 |    51 |     3  (34)| 00:00:01 |
|   2 |   WINDOW SORT       |        |     1 |    38 |     3  (34)| 00:00:01 |
|   3 |    TABLE ACCESS FULL| HHHHHH |     1 |    38 |     2   (0)| 00:00:01 |
------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

"   1 - filter(""X"".""CNT""=1)"

改写成开窗函数以后也是秒出。

<<PostgreSQL技术内幕：查询优化深度探索 >> 这本书是真的不错，偷偷刷了好几次，每次看完都有新的理解。