AdaptiveCursorSharing(第二篇)

选择率和硬解析

我们上面提到了,在v$sql_cs_histogram视图中,如果此游标的3个桶中出现了两个桶中的count都有非0值,那么此后的解析都要窥探绑定变量的值计算谓词选择率,如果计算选择率不在现有的游标的选择率范围内,就会基于窥探到的绑定变量的值重新硬解析产生一个新的游标,当然这个新游标的执行计划可能与之前是一样的。我们还是来看一个例子就会非常明白这种机制了。

成都创新互联公司从2013年成立,是专业互联网技术服务公司,拥有项目做网站、成都做网站网站策划,项目实施与项目整合能力。我们以让每一个梦想脱颖而出为使命,1280元莎车做网站,已为上家服务,为莎车各地企业和个人服务,联系电话:13518219792

SQL>create table t as select 1 id,a.* from dba_objects a,dba_objects b where rownum<10;

 

Table created.

 

SQL>create index t_ind on t(id);

 

Index created.

 

SQL>insert into t select 2,a.* from dba_objects a,dba_objects b where rownum<1000;

 

999 rows created.

 

SQL>insert into t select 3 ,a.* from dba_objects a,dba_objects b where rownum<10000;

 

9999 rows created.

 

SQL>insert into t select 4 ,a.* from dba_objects a,dba_objects b where rownum<100000;

 

99999 rows created.

 

SQL>insert into t select 5 ,a.* from dba_objects a,dba_objects b where rownum<1000000;

 

999999 rows created.

 

SQL>commit;

 

Commit complete.

SQL>begin

  2    dbms_stats.gather_table_stats(user,

  3                                  't',

  4                                  method_opt => 'for columns status size 5',

  5                                   cascade    => true);

  6    

  7  end;

  8  /

 

SQL>select id,count(*) from t group by id order by id;

 

        ID   COUNT(*)

---------- ----------

         1          9

         2        999

         3       9999

         4      99999

         5     999999

上面的代码精心构造了一个例子,表t上的id字段一共有5个唯一值,每个值的数量都不一样,id字段上有索引,分析了直方图。在这种情况下,如果我们直接使用字符变量不使用绑定变量的话,id在对1,2,3,4做查询的时候,都会使用索引扫描,这种情况下,索引扫描的成本要比全表扫描的成本低,id在对5做查询时,会使用全表扫描,这种情况下全表扫描的成本要比索引扫描成本低。如下表格,我是通过explain工具,使用文本变量后,得出的每个执行计划的cost,可以看到全表扫描的cost为2911,在查询id<5的情况下,由于索引扫描的cost都小于全表扫描的cost因此执行计划都选择了走索引扫描,只有在查询id等于5的时,才选择了走全表扫描。

ID

执行计划

COST

选择率

1

索引扫描

4

0.0000081

2

索引扫描

16

0.000899186

3

索引扫描

139

0.008999959

4

索引扫描

1370

0.090007696

5

索引扫描

13690

0. 900085058

5

全表扫描

2911

0. 900085058

上面的表格最后一列提供了谓词的选择率,此处选择率的计算公式为:

选择率=id=?的值在表中的数量/总数量

根据上面表格的cost我们可以知道,谓词的选择率在0.0000081到0.090007696之间都应该选择索引扫描,在0. 900085058的时候应该选择全表扫描,因为id在5的时候,索引扫描的成本13690已经远远大于了全表扫描的成本2911。我们看看下面的例子:

SQL>var a number;

SQL>exec :a :=1;

 

PL/SQL procedure successfully completed.

 

SQL>select count(object_id) from t where id=:a;

 

COUNT(OBJECT_ID)

----------------

               9

 

SQL>exec :a :=5;

 

PL/SQL procedure successfully completed.

 

SQL>select count(object_id) from t where id=:a;

 

COUNT(OBJECT_ID)

----------------

          999999

 

SQL>select count(object_id) from t where id=:a;

 

COUNT(OBJECT_ID)

----------------

          999999

 

SQL>col PREDICATE for a10

SQL>-- 选择率

SQL>SELECT   hash_value, sql_id, child_number, predicate, range_id, low, high

  2      FROM v$sql_cs_selectivity

  3     WHERE sql_id='56g5zg95hcxc1'

ORDER BY sql_id, child_number;  4  

 

SQL_ID             CHILD_NUMBER PREDICATE    RANGE_ID LOW                  HIGH

-----------------  ------------ ---------- -------------------- --------------------

56g5zg95hcxc1                 1 =A                  0 0.810076             0.990093

经过上面的一系列的操作后我们已经让这个cursor变得bind aware,如何让SQL变得bind aware我们上面已经论述过,这里不再做详细说明。经过这些步骤后,优化器已经产生出了一个child_number为1的新游标,这个游标基于绑定变量为5的值生成,谓词的选择率范围是:0.810076到0.990093。这个选择率跟我们上面表格里提供的选择率的关系是:(0.810076+0.990093)/2约等于我们上面表格里提供的选择率0. 900085058,Oracle为选择率稍微的预留了一些余地,这样很好。我们再执行id为1的查询看看:

SQL>exec :a :=1;

 

PL/SQL procedure successfully completed.

 

SQL>select count(object_id) from t where id=:a;

 

COUNT(OBJECT_ID)

----------------

               9

 

SQL>-- 选择率

SQL>SELECT   hash_value, sql_id, child_number, predicate, range_id, low, high

  2      FROM v$sql_cs_selectivity

  3     WHERE sql_id='56g5zg95hcxc1'

ORDER BY sql_id, child_number;   

 

SQL_ID         CHILD_NUMBER PREDICATE    RANGE_ID LOW                  HIGH

-------------- ------------ ---------- ---------- -------------------- --------------------

56g5zg95hcxc1             1 =A                  0 0.810076             0.990093

56g5zg95hcxc1             2 =A                  0 0.000007             0.000009

已经产生了child_number为2的子游标,是基于id为1的值产生的,选择率范围为:0.000007到0.000009。下面就到了本节关键的时刻了,我们再次查询id为4看看会出现什么情况。

SQL>exec :a :=4

 

PL/SQL procedure successfully completed.

 

SQL>select count(object_id) from t where id=:a;

 

COUNT(OBJECT_ID)

----------------

           99999

 

SQL>

SQL>col PREDICATE for a10

SQL>-- 选择率

SQL>SELECT   hash_value, sql_id, child_number, predicate, range_id, low, high

  2      FROM v$sql_cs_selectivity

  3     WHERE sql_id='56g5zg95hcxc1'

  4  ORDER BY sql_id, child_number;

 

SQL_ID          CHILD_NUMBER PREDICATE    RANGE_ID LOW                  HIGH

--------------- ------------ ---------- ---------- -------------------- --------------------

56g5zg95hcxc1              1 =A                  0 0.810076             0.990093

56g5zg95hcxc1              2 =A                  0 0.000007             0.000009

56g5zg95hcxc1              3 =A                  0 0.000007             0.099008

 

SQL>SELECT child_number, executions, buffer_gets, is_bind_sensitive,  

  2        is_bind_aware,IS_SHAREABLE                                     

  3   FROM v$sql                                                          

  4  WHERE sql_id='56g5zg95hcxc1';                                        

 

CHILD_NUMBER EXECUTIONS BUFFER_GETS IS IS IS

------------ ---------- ----------- -- -- --

           0          2       13690 Y  N  N

           1          1       13162 Y  Y  Y

           2          1           4 Y  Y  N

           3          1        1495 Y  Y  Y

优化器已经重新生成了一个child_numer为3的子游标,同时选择率的范围已经扩大了,从0.000007到0.099008,也就是现在从id为1到4都被包含在child_number为3的子游标里了。child_number为2的子游标已经被标记为不能共享失效了,如果共享池有紧缺这块内存就可以被清除出去。那是不是意味着我们查询id为3的值时,将不用重新产生新游标,直接可以使用child_number为3的子游标了。我们来看看:

SQL>exec :a :=3

 

PL/SQL procedure successfully completed.

 

SQL>col PREDICATE for a10

SQL>-- 选择率

SQL>SELECT   hash_value, sql_id, child_number, predicate, range_id, low, high

  2      FROM v$sql_cs_selectivity

  3     WHERE sql_id='56g5zg95hcxc1'

  4  ORDER BY sql_id, child_number;

 

SQL_ID          CHILD_NUMBER PREDICATE    RANGE_ID LOW                  HIGH

--------------- ------------ ---------- ---------- -------------------- --------------------

56g5zg95hcxc1              1 =A                  0 0.810076             0.990093

56g5zg95hcxc1              2 =A                  0 0.000007             0.000009

56g5zg95hcxc1              3 =A                  0 0.000007             0.099008

 

SQL>select count(object_id) from t where id=:a;

 

COUNT(OBJECT_ID)

----------------

            9999

 

SQL>-- 选择率

SQL>SELECT   hash_value, sql_id, child_number, predicate, range_id, low, high

  2      FROM v$sql_cs_selectivity

  3     WHERE sql_id='56g5zg95hcxc1'

  4  ORDER BY sql_id, child_number;

 

SQL_ID         CHILD_NUMBER PREDICATE    RANGE_ID LOW                  HIGH

-------------- ------------ ---------- ---------- -------------------- --------------------

56g5zg95hcxc1             1 =A                  0 0.810076             0.990093

56g5zg95hcxc1             2 =A                  0 0.000007             0.000009

56g5zg95hcxc1             3 =A                  0 0.000007             0.099008

 

SQL>SELECT child_number, executions, buffer_gets, is_bind_sensitive,  

  2        is_bind_aware,IS_SHAREABLE                                     

  3   FROM v$sql                                                          

  4  WHERE sql_id='56g5zg95hcxc1';                                        

 

CHILD_NUMBER EXECUTIONS BUFFER_GETS IS IS IS

------------ ---------- ----------- -- -- --

           0          2       13690 Y  N  N

           1          1       13162 Y  Y  Y

           2          1           4 Y  Y  N

           3          2        1495 Y  Y  Y

没有再生成新的子游标了,同时v$sql中的child_number为3的子游标的执行次数已经加1了。

从上面的示例我们可以知道,在v$sql_cs_histogram视图中,如果此游标的3个桶中出现了两个桶中的count都有非0值,那么此后的解析都要窥探绑定变量的值计算谓词选择率,如果计算选择率不在现有的游标的选择率范围内,就会基于窥探到的绑定变量的值重新硬解析产生一个新的游标,记录此游标的可以代表的选择率范围,当然就像我们例子看到的,新游标的执行计划可能跟之前是一样的,只不过是选择率的范围更广了。


新闻名称:AdaptiveCursorSharing(第二篇)
分享路径:http://scyanting.com/article/gpjdoj.html