窗口函数的成本很高，我们能克服它吗？

Question

我的需求：找出a的前10个账户DEPT_NUM，按账号升序排列。

询问：

SELECT * FROM
(
  select acctnum,dept_num,row_number() OVER (PARTITION BY DEPT_NUM ORDER BY ACCTNUM) as row_identifier
   FROM MYTABLE
)
WHERE row_identifier between 1 and 10;

痕迹：

    7532 rows selected.


Execution Plan
----------------------------------------------------------
Plan hash value: 1480074522

--------------------------------------------------------------------------------------------
| Id  | Operation                | Name    | Rows  | Bytes |TempSpc| Cost (%CPU)| Time     |
--------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT         |         |   577K|    15M|       |  3855   (1)| 00:00:47 |
|*  1 |  VIEW                    |         |   577K|    15M|       |  3855   (1)| 00:00:47 |
|*  2 |   WINDOW SORT PUSHED RANK|         |   577K|  7890K|    13M|  3855   (1)| 00:00:47 |
|   3 |    INDEX FAST FULL SCAN  | IMTAB05 |   577K|  7890K|       |   987   (1)| 00:00:12 |
--------------------------------------------------------------------------------------------

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

   1 - filter("ROW_IDENTIFIER">=1 AND "ROW_IDENTIFIER"<=5)
   2 - filter(ROW_NUMBER() OVER ( PARTITION BY "DEPT_NUM" ORDER BY "ACCTNUM")<=5)


Statistics
----------------------------------------------------------
          0  recursive calls
          2  db block gets
       4298  consistent gets
          0  physical reads
          0  redo size
     144367  bytes sent via SQL*Net to client
        486  bytes received via SQL*Net from client
          3  SQL*Net roundtrips to/from client
          1  sorts (memory)
          0  sorts (disk)
       7532  rows processed

指数：

上面index scan说，INDEX STORAGE在一个柱子上DEPT_NUM。

强制Full Table扫描成本从 3855 降到 11092

表中总行数为632667；

---

以上均为测试区结果。产量实际上是原来的两倍。

我的数据库是Exadata，Quarter RAC。运行 Oracle 11g R2。数据库足够强大，可以立即执行，但DBA不愿意使用13M的tempSpc。据商业报道，该报告的频率为每小时 4 次。主要的是，这个表进行了大量的实时插入/更新

我们可以临时改进这个过程吗？
1）增加会话的 PGA？（不确定，是否真的可能？）
2）额外的索引会有帮助吗？

只是希望用不同的眼光来看待这个问题，因为我们的团队完全只关注 DBA 参数。

感谢您的任何建议！

Answer 1

分析函数的性能可能取决于索引列的顺序。将索引从 更改为(ACCTNUM,DEPT_NUM)可能(DEPT_NUM,ACCTNUM)会降低成本并消除对临时表空间的需要。

partition by COL_2 order by COL_1 => INDEX FAST FULL SCAN|WINDOW SORT PUSHED RANK
partition by COL_1 order by COL_2 => INDEX FULL SCAN|WINDOW NOSORT

INDEX FAST FULL SCAN 使用更快的多块 IO，但它还需要对数据进行排序，并且可能需要为排序区域提供临时表空间。

INDEX FULL SCAN 使用较慢的单块 IO，但它按顺序返回数据并避免排序。

示例架构和数据

--drop table mytable;
create table mytable(dept_num number not null, acctnum number not null
    ,a number, b number, c number, d number, e number);
insert into mytable
select 1 dept_num, 1 acctnum, 0,0,0,0,0 from dual union all
select 1 dept_num, 2 acctnum, 0,0,0,0,0 from dual union all
select 1 dept_num, 3 acctnum, 0,0,0,0,0 from dual union all
select 2 dept_num, 1 acctnum, 0,0,0,0,0 from dual union all
select 2 dept_num, 2 acctnum, 0,0,0,0,0 from dual union all
select 3 dept_num, 1 acctnum, 0,0,0,0,0 from dual;
--Create 600K similar rows.
insert into mytable
  select dept_num + rownumber*3, acctnum, a,b,c,d,e
  from mytable
  cross join (select level rownumber from dual connect by level <= 100000);
begin
    dbms_stats.gather_table_stats(user, 'mytable');
end;
/

(ACCTNUM,DEPT_NUM) = 窗口排序推送排名

create index mytable_idx on mytable(acctnum, dept_num);

explain plan for
select dept_num, acctnum from
(
    select dept_num, acctnum
        ,row_number() over (partition by dept_num order by acctnum) as row_identifier
    from mytable
)
where row_identifier between 1 and 10;

select * from table(dbms_xplan.display);

Plan hash value: 952182109

------------------------------------------------------------------------------------------------
| Id  | Operation                | Name        | Rows  | Bytes |TempSpc| Cost (%CPU)| Time     |
------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT         |             |   600K|    22M|       |  1625   (3)| 00:00:23 |
|*  1 |  VIEW                    |             |   600K|    22M|       |  1625   (3)| 00:00:23 |
|*  2 |   WINDOW SORT PUSHED RANK|             |   600K|  4687K|  9424K|  1625   (3)| 00:00:23 |
|   3 |    INDEX FAST FULL SCAN  | MYTABLE_IDX |   600K|  4687K|       |   239   (3)| 00:00:04 |
------------------------------------------------------------------------------------------------

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

   1 - filter("ROW_IDENTIFIER">=1 AND "ROW_IDENTIFIER"<=10)
   2 - filter(ROW_NUMBER() OVER ( PARTITION BY "DEPT_NUM" ORDER BY "ACCTNUM")<=10)

(DEPT_NUM,ACCTNUM) = 窗口不排序

drop index mytable_idx;
create index mytable_idx on mytable(dept_num, acctnum);

explain plan for
select dept_num, acctnum from
(
    select dept_num, acctnum
        ,row_number() over (partition by dept_num order by acctnum) as row_identifier
    from mytable
)
where row_identifier between 1 and 10;

select * from table(dbms_xplan.display);

Plan hash value: 1773829932

---------------------------------------------------------------------------------
| Id  | Operation         | Name        | Rows  | Bytes | Cost (%CPU)| Time     |
---------------------------------------------------------------------------------
|   0 | SELECT STATEMENT  |             |   600K|    22M|   792   (2)| 00:00:12 |
|*  1 |  VIEW             |             |   600K|    22M|   792   (2)| 00:00:12 |
|*  2 |   WINDOW NOSORT   |             |   600K|  4687K|   792   (2)| 00:00:12 |
|   3 |    INDEX FULL SCAN| MYTABLE_IDX |   600K|  4687K|   792   (2)| 00:00:12 |
---------------------------------------------------------------------------------

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

   1 - filter("ROW_IDENTIFIER">=1 AND "ROW_IDENTIFIER"<=10)
   2 - filter(ROW_NUMBER() OVER ( PARTITION BY "DEPT_NUM" ORDER BY 
              "ACCTNUM")<=10)