对Group By 语句的一次优化过程

对Group By 语句的一次优化过程

作者： fuyuncat

来源： www.HelloDBA.com

生产环境中发现一条语句很慢，拿回来一看，其实是一个简单的Group By语句:

表CCMMT的数据量比较大，5M多条记录。

1、

SQL> select CDE, CID

2 from CCMMT

3 GROUP BY CDE, CID

4 having max(ADT) < sysdate - 180;

707924 rows selected.

Elapsed: 00:06:17.49

Execution Plan

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

0 SELECT STATEMENT Optimizer=CHOOSE (Cost=414 Card=238583 Bytes=4771660)

1 0 FILTER

2 1 SORT (GROUP BY NOSORT) (Cost=414 Card=238583 Bytes=4771660)

3 2 TABLE ACCESS (BY INDEX ROWID) OF 'CCMMT' (Cost=414 Card=57969096 Bytes=1159381920)

4 3 INDEX (FULL SCAN) OF 'CCMMT_TEMP_IDX' (NON-UNIQUE) (Cost=26 Card=57969096)

Statistics

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

0 recursive calls

0 db block gets

2769177 consistent gets

1089991 physical reads

0 redo size

23926954 bytes sent via SQL*Net to client

519785 bytes received via SQL*Net from client

47196 SQL*Net roundtrips to/from client

0 sorts (memory)

0 sorts (disk)

707924 rows processed

要6min多返回。尝试调整语句写法，用minus代替Group By:

2、

SQL> select DISTINCT CDE, CID

2 from CCMMT

3 where ADT < sysdate - 180

4 minus

5 select DISTINCT CDE, CID

6 from CCMMT

7 where ADT >= sysdate - 180;

707924 rows selected.

Elapsed: 00:00:21.53

Execution Plan

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

0 SELECT STATEMENT Optimizer=CHOOSE (Cost=190624 Card=2794940

Bytes=111797600)

1 0 MINUS

2 1 SORT* (UNIQUE) (Cost=95312 Card=2794940 Bytes=55898800) :Q13049001

3 2 INDEX* (FAST FULL SCAN) OF 'CCMMT_UQ1' (UNIQUE) (Cost=77305 Card=2898455 Bytes=57969100) :Q13049000

4 1 SORT* (UNIQUE) (Cost=95312 Card=2794940 Bytes=55898800) :Q13050001

5 4 INDEX* (FAST FULL SCAN) OF 'CCMMT_UQ1' (UNIQUE) (Cost=77305 Card=2898455 Bytes=57969100) :Q13050000

2 PARALLEL_TO_SERIAL SELECT DISTINCT C0 C0,C1 C1 FROM :Q13049000 ORDER BY C0,C1

3 PARALLEL_TO_PARALLEL SELECT /*+ INDEX_RRS(A1 "CCMMT_UQ1")*/ A1."CDE" C0,A1."CA

4 PARALLEL_TO_SERIAL SELECT DISTINCT C0 C0,C1 C1 FROM :Q13050000 ORDER BY C0,C1

5 PARALLEL_TO_PARALLEL SELECT /*+ INDEX_RRS(A1 "CCMMT_UQ1")*/ A1."CDE" C0,A1."CA

Statistics

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

0 recursive calls

33 db block gets

126566 consistent gets

129243 physical reads

0 redo size

18461368 bytes sent via SQL*Net to client

519785 bytes received via SQL*Net from client

47196 SQL*Net roundtrips to/from client

4 sorts (memory)

2 sorts (disk)

707924 rows processed

效果不错，Consistent gets 和 Physical Reads都下降了，同时只需要21s就返回了。但从查询计划看，用到了并行查询，因此会消耗更多的CPU。

在(ADT, CDE, CID )上创建索引，再次执行：

3、

SQL> select DISTINCT CDE, CID

2 from CCMMT

3 where ADT < sysdate - 180

4 minus

5 select DISTINCT CDE, CID

6 from CCMMT

7 where ADT >= sysdate - 180;

707924 rows selected.

Elapsed: 00:00:26.94

Execution Plan

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

0 SELECT STATEMENT Optimizer=CHOOSE (Cost=36018 Card=2794940 Bytes=111797600)

1 0 MINUS

2 1 SORT (UNIQUE) (Cost=18009 Card=2794940 Bytes=55898800)

3 2 INDEX (RANGE SCAN) OF 'CCMMT_IDX3' (NON-UNIQUE) (Cost=2 Card=2898455 Bytes=57969100)

4 1 SORT (UNIQUE) (Cost=18009 Card=2794940 Bytes=55898800)

5 4 INDEX (RANGE SCAN) OF 'CCMMT_IDX3' (NON-UNIQUE) (Cost=2 Card=2898455 Bytes=57969100)

Statistics

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

0 recursive calls

118 db block gets

22565 consistent gets

31604 physical reads

0 redo size

18461368 bytes sent via SQL*Net to client

519785 bytes received via SQL*Net from client

47196 SQL*Net roundtrips to/from client

1 sorts (memory)

1 sorts (disk)

707924 rows processed

效果也比较理想，consistent gets和physical reads再次大大下降，返回时间和上面差不多，在一个数量级上，但是不再使用并行查询了。

用NOT Exists代替minus:

4、

SQL> select DISTINCT CDE, CID

2 from CCMMT a

3 where ADT < sysdate - 180

4 AND NOT EXISTS

5 (SELECT CDE, CID FROM

6 (select DISTINCT CDE, CID

7 from CCMMT

8 where ADT >= sysdate - 180) b

9 WHERE a.CDE = b.CDE

10 AND a.CID = b.CID);

707924 rows selected.

Elapsed: 00:10:35.70

Execution Plan

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

0 SELECT STATEMENT Optimizer=CHOOSE (Cost=600 Card=144923 Bytes=2898460)

1 0 SORT (UNIQUE) (Cost=600 Card=144923 Bytes=2898460)

2 1 INDEX (RANGE SCAN) OF 'CCMMT_IDX3' (NON-UNIQUE)(Cost=2 Card=144923 Bytes=2898460)

3 2 TABLE ACCESS (BY INDEX ROWID) OF 'CCMMT' (Cost=2 Card=1 Bytes=20)

4 3 INDEX (RANGE SCAN) OF 'CCMMT_TEMP_IDX' (NON-UNIQUE) (Cost=1 Card=9)

Statistics

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

5 recursive calls

118 db block gets

40535587 consistent gets

3157604 physical reads

0 redo size

18461368 bytes sent via SQL*Net to client

519785 bytes received via SQL*Net from client

47196 SQL*Net roundtrips to/from client

2 sorts (memory)

1 sorts (disk)

707924 rows processed

FT! consistent gets和physical reads爆涨，10min才返回结果！

用Not In换掉Not Exists:

5、

SQL> select DISTINCT CDE, CID

2 from CCMMT a

3 where ADT < sysdate - 180

4 AND (CDE, CID) NOT IN

5 (select DISTINCT CDE, CID

6 from CCMMT

7 where ADT >= sysdate - 180);

707924 rows selected.

Elapsed: 00:01:00.70

Execution Plan

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

0 SELECT STATEMENT Optimizer=CHOOSE (Cost=36425 Card=1 Bytes=40)

1 0 SORT (UNIQUE NOSORT) (Cost=36425 Card=1 Bytes=40)

2 1 MERGE JOIN (ANTI) (Cost=36423 Card=1 Bytes=40)

3 2 SORT (JOIN) (Cost=18212 Card=2898455 Bytes=57969100)

4 3 INDEX (RANGE SCAN) OF 'CCMMT_IDX3' (NON-UNIQUE) (Cost=2 Card=2898455 Bytes=57969100)

5 2 SORT (UNIQUE) (Cost=18212 Card=2898455 Bytes=57969100)

6 5 INDEX (RANGE SCAN) OF 'CCMMT_IDX3' (NON-UNIQUE) (Cost=2 Card=2898455 Bytes=57969100)

Statistics

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

0 recursive calls

419 db block gets

22565 consistent gets

98692 physical reads

0 redo size

18461368 bytes sent via SQL*Net to client

519785 bytes received via SQL*Net from client

47196 SQL*Net roundtrips to/from client

1 sorts (memory)

1 sorts (disk)

707924 rows processed

恩，consistent gets和建了索引时的minus方式一样，但是physical reads太大，返回时间太长---1min。同时用到了刚才建的索引。（呵呵，所以说，NOT EXISTS并不是什么情况下都比NOT IN更优啊）

在尝试用left join + is null代替not in:

6、

SQL> SELECT a.CDE, a.CID

2 FROM

3 (select DISTINCT CDE, CID

4 from CCMMT

5 where ADT < sysdate - 180) a,

6 (select DISTINCT CDE, CID

7 from CCMMT

8 where ADT >= sysdate - 180) b

9 WHERE a.CDE = b.CDE(+)

10 AND a.CID = b.CID(+)

11 AND b.CDE IS NULL;

707924 rows selected.

Elapsed: 00:00:25.46

Execution Plan

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

0 SELECT STATEMENT Optimizer=CHOOSE (Cost=54675 Card=2794940 Bytes=117387480)

1 0 FILTER

2 1 MERGE JOIN (OUTER)

3 2 VIEW (Cost=18009 Card=2794940 Bytes=58693740)

4 3 SORT (UNIQUE) (Cost=18009 Card=2794940 Bytes=55898800)

5 4 INDEX (RANGE SCAN) OF 'CCMMT_IDX3' (NON-UNIQUE) (Cost=2 Card=2898455 Bytes=57969100)

6 2 SORT (JOIN) (Cost=36667 Card=2794940 Bytes=58693740)

7 6 VIEW (Cost=18009 Card=2794940 Bytes=58693740)

8 7 SORT (UNIQUE) (Cost=18009 Card=2794940 Bytes=55898800)

9 8 INDEX (RANGE SCAN) OF 'CCMMT_IDX3' (NON-UNIQUE) (Cost=2 Card=2898455 Bytes=57969100)

Statistics

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

10 recursive calls

118 db block gets

22569 consistent gets

31300 physical reads

0 redo size

18461368 bytes sent via SQL*Net to client

519785 bytes received via SQL*Net from client

47196 SQL*Net roundtrips to/from client

6 sorts (memory)

1 sorts (disk)

707924 rows processed

效果不错，和有索引时使用minus在同一数量级上。

总结，以上几种方式中，效果最好的应该是第3种和第6种，buffer gets、磁盘IO和CPU消耗都比较少，返回时间大大减少，但是需要新建一个索引，消耗更多磁盘空间，并存在影响其它语句的正常查询计划的风险。而第2种方式应该是次好的。在返回时间上，和上面两种差不多，不需要新的索引，但是会消耗更多的内存、磁盘和CPU资源。

出于综合考虑，采用了第2种方式对生产库进行了优化。

（以上例子中的对象名进行了替换，其他都是原版）