distinct关键字对执行计划的影响

一、前言

最近看到一段话，"count(distinct 列名)若列上有索引，且有非空约束或在where子句中使用is not null，则会选择索引快速全扫描。其余情况则选择全表扫描"，对其中的原理不理解，因此有了以下的实验。

二、准备工作

1. 准备t1表

SQL> create table t1 as select * from dba_objects;
SQL> insert into t1 select * from t1;
SQL> insert into t1 select * from t1;
SQL> commit;

2. 将object_name列弄出少量的空值

SQL> update t1 set object_name = null where owner = 'SCOTT';

3. 在object_name列上创建普通索引

SQL> create index idx_t1_name on t1(object_name);

4. 收集t1表和t1表上索引的统计信息

SQL> begin
   2  dbms_stats.gather_table_stats(ownname => 'SCOTT',
   3  tabname => 'T1',
   4  estimate_percent => 100,
   5  cascade => true, 
   6  no_invalidate => false,
   7  degree => 4);
   8  end;
   9  /

5. 统计t1表的总行数，object_name的行数

SQL> select count(*), count(object_name), count(distinct object_name) from t1;

  COUNT(*) COUNT(OBJECT_NAME) COUNT(DISTINCTOBJECT_NAME)
---------- ------------------ --------------------------
     54068              54060                      10472

至此，准备工作已经完成。t1表有54068行，object_name列有54060行，之所以这个值比总行数少，是因为count(列)的时候不统计该列上的空值。

三、查看执行计划

分别执行下面四条sql，观察执行计划
a. select count(object_name) from t1;    
b. select count(object_name) from t1 where object_name is not null;        
c. select count(distinct object_name) from t1 where object_name is not null;        
d. select count(distinct object_name) from t1;

1. 执行sql(a)

SQL> set autot on
SQL> select count(object_name) from t1;

COUNT(OBJECT_NAME)
------------------
             54060

-------------------------------------------------------------------------------------
| Id  | Operation             | Name        | Rows  | Bytes | Cost (%CPU)| Time     |
-------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT      |             |     1 |    19 |    63   (0)| 00:00:01 |
|   1 |  SORT AGGREGATE       |             |     1 |    19 |            |          |
|   2 |   INDEX FAST FULL SCAN| IDX_T1_NAME | 54068 |  1003K|    63   (0)| 00:00:01 |
-------------------------------------------------------------------------------------

2. 执行sql(b)

SQL> select count(object_name) from t1 where object_name is not null;

COUNT(OBJECT_NAME)
------------------
             54060

-------------------------------------------------------------------------------------
| Id  | Operation             | Name        | Rows  | Bytes | Cost (%CPU)| Time     |
-------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT      |             |     1 |    19 |    63   (0)| 00:00:01 |
|   1 |  SORT AGGREGATE       |             |     1 |    19 |            |          |
|*  2 |   INDEX FAST FULL SCAN| IDX_T1_NAME | 54060 |  1003K|    63   (0)| 00:00:01 |
-------------------------------------------------------------------------------------

可以看到sql(a)和sql(b)的执行结果和执行计划都一样，执行结果一样很好理解，count(object_name)本来就不会统计object_name为空的行，所以后面有没有where object_name is not null对结果都没有影响。
执行计划一样，也很好理解，都是走的索引快速全扫描，毕竟我只是想得到object_name有多少个值，空值我根本不管，而btree索引刚好也不存储空值，所以只需要统计object_name上的索引有多少行就行了。

3. 执行sql(c)

SQL> select count(distinct object_name) from t1 where object_name is not null;

COUNT(DISTINCTOBJECT_NAME)
--------------------------
                     10472

-----------------------------------------------------------------------------------------------
| Id  | Operation               | Name        | Rows  | Bytes |TempSpc| Cost (%CPU)| Time     |
-----------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT        |             |     1 |    66 |       |   220   (2)| 00:00:03 |
|   1 |  SORT AGGREGATE         |             |     1 |    66 |       |            |          |
|   2 |   VIEW                  | VW_DAG_0    | 10472 |   674K|       |   220   (2)| 00:00:03 |
|   3 |    HASH GROUP BY        |             | 10472 |   194K|  1496K|   220   (2)| 00:00:03 |
|*  4 |     INDEX FAST FULL SCAN| IDX_T1_NAME | 54060 |  1003K|       |    63   (0)| 00:00:01 |
-----------------------------------------------------------------------------------------------

可以看到sql(c)比sql(b)多了一个distinct关键字，执行计划仍然采用的是索引快速全扫描。

4. 执行sql(d)

SQL> select count(distinct object_name) from t1;

COUNT(DISTINCTOBJECT_NAME)
--------------------------
                     10472

-----------------------------------------------------------------------------------------
| Id  | Operation            | Name     | Rows  | Bytes |TempSpc| Cost (%CPU)| Time     |
-----------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT     |          |     1 |    66 |       |   349   (1)| 00:00:05 |
|   1 |  SORT AGGREGATE      |          |     1 |    66 |       |            |          |
|   2 |   VIEW               | VW_DAG_0 | 10472 |   674K|       |   349   (1)| 00:00:05 |
|   3 |    HASH GROUP BY     |          | 10472 |   194K|  1496K|   349   (1)| 00:00:05 |
|   4 |     TABLE ACCESS FULL| T1       | 54068 |  1003K|       |   192   (0)| 00:00:03 |
-----------------------------------------------------------------------------------------

可以看到sql(d)在sql(c)的基础上，删掉了where object_name is not null，执行结果没有变，但是执行计划由索引快速全扫描变成了全表扫描。照道理来讲，sql(d)依然可以使用索引的快速全扫描就可以得出结果，但是却选择了cost更大的全表扫描，这个是为什么呢？

四、问题

a. select count(object_name) from t1;    
b. select count(object_name) from t1 where object_name is not null;        
c. select count(distinct object_name) from t1 where object_name is not null;        
d. select count(distinct object_name) from t1;

sql(a)与sql(b)，都走索引INDEX FAST FULL SCAN，在它的上层是SORT AGGREGATE。也就是扫个索引，统计下索引行数就行了。
sql(c)，也走索引INDEX FAST FULL SCAN，它的上层是HASH GROUP BY，然后是VIEW，最后才是SORT AGGREGATE。
sql(d)，走的是全表扫描，它的上层是HASH GROUP BY，然后是VIEW，最后才是SORT AGGREGATE。

count(object_name)，oracle知道空值对结果没有什么影响，所以不管加不加where条件，都能走索引。
count(distinct object_name)，oracle估计就懵了，它会在sql中先看看有没有过滤条件。如果将空值踢掉了，开开心心走索引，没踢掉，老老实实全表扫描。
这是为啥？