相关列问题

 先来创建一个表T

create table t as select level as id ,level||'a' as a,level||level||'b' as b from dual connect by level<100;

这里A列的值能够确定B列的值，

insert into t  select * from t; 
.............................. 一直重复插入数据

SQL> select count(*) from t;

  COUNT(*)
----------
   3244032

create index idx1 on t(a);  
create index idx2 on t(a,b);   
  
BEGIN
  DBMS_STATS.GATHER_TABLE_STATS(ownname          => 'SCOTT',
                                tabname          => 'T',
                                estimate_percent => 100,
                                method_opt       => 'for all columns size skewonly',
                                no_invalidate    => FALSE,
                                degree           => 8,
                                cascade          => TRUE);
END;
/

SQL> select * from t where a='1a' and b='11b';

已选择32768行。

已用时间:  00: 00: 03.98

执行计划
----------------------------------------------------------
Plan hash value: 2303463401

------------------------------------------------------------------------------------
| Id  | Operation                   | Name | Rows  | Bytes | Cost (%CPU)| Time     |
------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT            |      |   331 |  3972 |    84   (0)| 00:00:02 |
|   1 |  TABLE ACCESS BY INDEX ROWID| T    |   331 |  3972 |    84   (0)| 00:00:02 |
|*  2 |   INDEX RANGE SCAN          | IDX2 |   331 |       |     3   (0)| 00:00:01 |
------------------------------------------------------------------------------------

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

   2 - access("A"='1a' AND "B"='11b')


统计信息
----------------------------------------------------------
          1  recursive calls
          0  db block gets
      11838  consistent gets
       7943  physical reads
          0  redo size
     441749  bytes sent via SQL*Net to client
      24424  bytes received via SQL*Net from client
       2186  SQL*Net roundtrips to/from client
          0  sorts (memory)
          0  sorts (disk)
      32768  rows processed

因为CBO不知道A与B关系，所以计算基数等于331，

SQL> select 1/99/99*3244032 from dual; ----这个其实就是 a选择性*b选择性 =(1/99)*(1/99)

1/99/99*3244032
---------------
     330.989899

但是实际上它要返回32768条记录

SQL> select * from t where a='1a';

已选择32768行。

已用时间:  00: 00: 01.38

执行计划
----------------------------------------------------------
Plan hash value: 1601196873

--------------------------------------------------------------------------
| Id  | Operation         | Name | Rows  | Bytes | Cost (%CPU)| Time     |
--------------------------------------------------------------------------
|   0 | SELECT STATEMENT  |      | 32768 |   384K|  1874   (8)| 00:00:23 |
|*  1 |  TABLE ACCESS FULL| T    | 32768 |   384K|  1874   (8)| 00:00:23 |
--------------------------------------------------------------------------

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

   1 - filter("A"='1a')


统计信息
----------------------------------------------------------
          0  recursive calls
          0  db block gets
      10120  consistent gets
       6312  physical reads
          0  redo size
     441749  bytes sent via SQL*Net to client
      24424  bytes received via SQL*Net from client
       2186  SQL*Net roundtrips to/from client
          0  sorts (memory)
          0  sorts (disk)
      32768  rows processed

如果where条件单独是 where a='1a' CBO 就能够算对基数，它的基数是这样计算的

SQL> select 3244032/99 from dual;

3244032/99
----------
     32768 

很显然，这个SQL select * from t where a='1a' and b='11b' 的执行计划走错了，它应该走全表扫描，但是因为计算基数错误，导致它走 IDX2这个索引

相关列的解决办法在Oracle中有2个，一个是动态采样，另外一个就是Oracle11g，对相关列收集扩展统计

SQL> ALTER SESSION SET optimizer_dynamic_sampling=6;

会话已更改。

SQL> set lines 200
SQL> set pages 200
SQL> set timi on
SQL> explain plan for select * from t where a='1a' and b='11b';

已解释。

已用时间:  00: 00: 00.86
SQL> select * from table(dbms_xplan.display);

PLAN_TABLE_OUTPUT
---------------------------------------------------------------------------------------
Plan hash value: 1601196873

--------------------------------------------------------------------------
| Id  | Operation         | Name | Rows  | Bytes | Cost (%CPU)| Time     |
--------------------------------------------------------------------------
|   0 | SELECT STATEMENT  |      | 32776 |   384K|  1885   (8)| 00:00:23 |
|*  1 |  TABLE ACCESS FULL| T    | 32776 |   384K|  1885   (8)| 00:00:23 |
--------------------------------------------------------------------------

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

   1 - filter("A"='1a' AND "B"='11b')

Note
-----
   - dynamic sampling used for this statement

已选择17行。

设置动态采样之后 Oracle评估基数就基本正确了，关于11g扩展统计这里就不做了，有兴趣的请自己做一下。

我对相关列的建议就是，能否在程序里拼接？如果A能确定B，那么做DB 设计的时候就不要创建B列了 直接在程序里根据A列的值生成B的值 这样减少DB的存储空间。

如果非要在DB里设置B列，写SQL的时候就不要把2个列都写进去，也就是说不要写成
select * from t where a='1a' and b='11b';

直接写成
select * from t where a='1a'  或者 select * from t where b='11b'

这样能尽量避免CBO计算基数出错，如果这个表要与多表关联，基数一旦算错，必然导致整个SQL的执行计划全部出错，从而导致SQL性能下降。

动态采样和扩展统计虽然是解决办法，但是如果产品要考虑兼容性呢？我的产品要同时支持ORACLE,DB2，SQLSERVER，甚至以后的国产数据库达梦，他们没有动态采样怎么办。