Oracle 12c 中的子选择性能不佳答案

【问题标题】：Subselect in Oracle 12c with bad performanceOracle 12c 中的子选择性能不佳
【发布时间】：2017-02-27 08:25:50
【问题描述】：

从 Oracle 11g 迁移到 12c 后，我们遇到了一个巨大的性能问题，例如以下 select 语句。同样的语句在 11g 中也能正常工作。

表和索引

-- Create table
create table PS_CS_ADRART_TB
(
  cs_adressmandant   VARCHAR2(5) not null,
  cs_person_id       VARCHAR2(20) not null,
  cs_beziehung_id    VARCHAR2(20) not null,
  seqnum             INTEGER not null,
  effdt              DATE,
  eff_status         VARCHAR2(1) not null,
  cs_adrart_cd       VARCHAR2(20) not null,
  cs_adress_id       VARCHAR2(20) not null,
  cs_kdnr_as400      VARCHAR2(8) not null,
  cs_plzgk           VARCHAR2(11) not null,
  cs_plz_pf          VARCHAR2(15) not null,
  cs_aendgr_cd       VARCHAR2(20) not null,
  cs_datasource_cd   VARCHAR2(20) not null,
  cs_betrag          NUMBER(14,4) not null,
  cs_belegdat        DATE,
  cs_adrtyp_xl       VARCHAR2(2) not null,
  cs_checked         VARCHAR2(1) not null,
  cs_journal_xl      VARCHAR2(4) not null,
  address2           VARCHAR2(55) not null,
  row_added_dttm     TIMESTAMP(6),
  row_added_oprid    VARCHAR2(30) not null,
  row_lastmant_dttm  TIMESTAMP(6),
  row_lastmant_oprid VARCHAR2(30) not null,
  cs_recstat_xl      VARCHAR2(4) not null,
  cs_update_count    NUMBER(10) not null
)
tablespace CS_APP
  pctfree 10
  initrans 1
  maxtrans 255
  storage
  (
    initial 102416K
    next 1M
    minextents 1
    maxextents unlimited
  );
-- Create/Recreate indexes 
create unique index PSACS_ADRART_TB on PS_CS_ADRART_TB (CS_ADRESSMANDANT, CS_KDNR_AS400, EFFDT)
  tablespace PSINDEX
  pctfree 10
  initrans 2
  maxtrans 255
  storage
  (
    initial 20M
    next 1M
    minextents 1
    maxextents unlimited
  );
create index PSBCS_ADRART_TB on PS_CS_ADRART_TB (CS_PERSON_ID)
  tablespace PSINDEX
  pctfree 10
  initrans 2
  maxtrans 255
  storage
  (
    initial 20M
    next 1M
    minextents 1
    maxextents unlimited
  );
create index PSCCS_ADRART_TB on PS_CS_ADRART_TB (CS_BEZIEHUNG_ID)
  tablespace PSINDEX
  pctfree 10
  initrans 2
  maxtrans 255
  storage
  (
    initial 20M
    next 1M
    minextents 1
    maxextents unlimited
  );
create unique index PS_CS_ADRART_TB on PS_CS_ADRART_TB (CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT)
  tablespace PSINDEX
  pctfree 10
  initrans 2
  maxtrans 255
  storage
  (
    initial 20M
    next 1M
    minextents 1
    maxextents unlimited
  );
create index PSDCS_ADRART_TB on PS_CS_ADRART_TB (CS_PLZ_PF)
  tablespace PSINDEX
  pctfree 10
  initrans 2
  maxtrans 255
  storage
  (
    initial 40K
    next 1M
    minextents 1
    maxextents unlimited
  );
create index PS0CS_ADRART_TB on PS_CS_ADRART_TB (CS_ADRESS_ID, CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT)
  tablespace PSINDEX
  pctfree 10
  initrans 2
  maxtrans 255
  storage
  (
    initial 20M
    next 1M
    minextents 1
    maxextents unlimited
  );
create index PS1CS_ADRART_TB on PS_CS_ADRART_TB (CS_KDNR_AS400, CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT)
  tablespace PSINDEX
  pctfree 10
  initrans 2
  maxtrans 255
  storage
  (
    initial 20M
    next 1M
    minextents 1
    maxextents unlimited
  );
create index PS2CS_ADRART_TB on PS_CS_ADRART_TB (ROW_ADDED_DTTM, CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT)
  tablespace PSINDEX
  pctfree 10
  initrans 2
  maxtrans 255
  storage
  (
    initial 40K
    next 1M
    minextents 1
    maxextents unlimited
  );
create index PS3CS_ADRART_TB on PS_CS_ADRART_TB (ROW_ADDED_OPRID, CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT)
  tablespace PSINDEX
  pctfree 10
  initrans 2
  maxtrans 255
  storage
  (
    initial 40K
    next 1M
    minextents 1
    maxextents unlimited
  );
create index PS4CS_ADRART_TB on PS_CS_ADRART_TB (ROW_LASTMANT_DTTM, CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT)
  tablespace PSINDEX
  pctfree 10
  initrans 2
  maxtrans 255
  storage
  (
    initial 40K
    next 1M
    minextents 1
    maxextents unlimited
  );
create index PS5CS_ADRART_TB on PS_CS_ADRART_TB (ROW_LASTMANT_OPRID, CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT)
  tablespace PSINDEX
  pctfree 10
  initrans 2
  maxtrans 255
  storage
  (
    initial 40K
    next 1M
    minextents 1
    maxextents unlimited
  );
create index PS6CS_ADRART_TB on PS_CS_ADRART_TB (CS_RECSTAT_XL, CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT)
  tablespace PSINDEX
  pctfree 10
  initrans 2
  maxtrans 255
  storage
  (
    initial 40K
    next 1M
    minextents 1
    maxextents unlimited
  );

表格大小

select count(*) from ps_cs_adrart_tb a 
--> 41367270

选择语句

    SELECT A.CS_ADRESS_ID, A.SEQNUM, TO_CHAR(A.EFFDT, 'YYYY-MM-DD')
      from PS_CS_ADRART_TB A
    where A.CS_ADRESSMANDANT = '001'
       and a.cs_kdnr_as400 = '63916917'
       and a.effdt =
           (select max(b.effdt)
              from ps_cs_adrart_tb b
             where b.cs_adressmandant = a.cs_adressmandant 
               and b.cs_person_id = a.cs_person_id 
               and b.cs_beziehung_id = a.cs_beziehung_id
               and b.seqnum = a.seqnum
               and b.effdt  <= trunc(sysdate)
               )

说明 Oracle 11g 计划

    --------------------------------------------------------------------------------------------
    | Id  | Operation                       | Name            | Rows | Bytes | Cost | Time     |
    --------------------------------------------------------------------------------------------
    |   0 | SELECT STATEMENT                |                 |    1 |    89 |    2 | 00:00:01 |
    | * 1 |   FILTER                        |                 |      |       |      |          |
    |   2 |    TABLE ACCESS BY INDEX ROWID  | PS_CS_ADRART_TB |    1 |    89 |    1 | 00:00:01 |
    | * 3 |     INDEX RANGE SCAN            | PSACS_ADRART_TB |    1 |       |    1 | 00:00:01 |
    |   4 |    SORT AGGREGATE               |                 |    1 |    59 |      |          |
    |   5 |     FIRST ROW                   |                 |    1 |    59 |    1 | 00:00:01 |
    | * 6 |      INDEX RANGE SCAN (MIN/MAX) | PS_CS_ADRART_TB |    1 |    59 |    1 | 00:00:01 |
    --------------------------------------------------------------------------------------------

    Predicate Information (identified by operation id):
    ------------------------------------------
    * 1 - filter("A"."EFFDT"= (SELECT MAX("B"."EFFDT") FROM "PS_CS_ADRART_TB" "B" WHERE "B"."EFFDT"<=TRUNC(SYSDATE@!) AND "B"."SEQNUM"=:B1 AND "B"."CS_BEZIEHUNG_ID"=:B2 AND "B"."CS_PERSON_ID"=:B3 AND
      "B"."CS_ADRESSMANDANT"=:B4))
    * 3 - access("A"."CS_ADRESSMANDANT"='001' AND "A"."CS_KDNR_AS400"='63916917')
    * 6 - access("B"."CS_ADRESSMANDANT"=:B1 AND "B"."CS_PERSON_ID"=:B2 AND "B"."CS_BEZIEHUNG_ID"=:B3 AND "B"."SEQNUM"=:B4 AND "B"."EFFDT"<=TRUNC(SYSDATE@!))

说明 Oracle 12c 计划

    ------------------------------------------------------------------------------------------------------------
    | Id  | Operation                              | Name            | Rows     | Bytes     | Cost  | Time     |
    ------------------------------------------------------------------------------------------------------------
    |   0 | SELECT STATEMENT                       |                 |        1 |       140 | 34366 | 00:00:02 |
    | * 1 |   HASH JOIN                            |                 |        1 |       140 | 34366 | 00:00:02 |
    |   2 |    TABLE ACCESS BY INDEX ROWID BATCHED | PS_CS_ADRART_TB |        1 |        89 |     1 | 00:00:01 |
    | * 3 |     INDEX RANGE SCAN                   | PS1CS_ADRART_TB |        1 |           |     1 | 00:00:01 |
    |   4 |    VIEW                                | VW_SQ_1         |    41889 |   2136339 | 34365 | 00:00:02 |
    | * 5 |     FILTER                             |                 |          |           |       |          |
    |   6 |      HASH GROUP BY                     |                 |    41889 |   2471451 | 34365 | 00:00:02 |
    | * 7 |       INDEX RANGE SCAN                 | PS_CS_ADRART_TB | 12746381 | 752036479 | 34365 | 00:00:02 |
    ------------------------------------------------------------------------------------------------------------

    Predicate Information (identified by operation id):
    ------------------------------------------
    * 1 - access("A"."EFFDT"="MAX(B.EFFDT)" AND "ITEM_1"="A"."CS_ADRESSMANDANT" AND "ITEM_2"="A"."CS_PERSON_ID" AND "ITEM_3"="A"."CS_BEZIEHUNG_ID" AND "ITEM_4"="A"."SEQNUM")
    * 3 - access("A"."CS_KDNR_AS400"='63916917' AND "A"."CS_ADRESSMANDANT"='001')
    * 5 - filter('001'='001')
    * 7 - access("B"."CS_ADRESSMANDANT"='001' AND "B"."EFFDT"<=TRUNC(SYSDATE@!))
    * 7 - filter("B"."EFFDT"<=TRUNC(SYSDATE@!))


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

如您所见，从索引PS_CS_ADRART_TB 中，只有CS_ADRESSMANDANT 和EFFDT 用于过滤数据，这非常糟糕。

对于以下稍有不同的选择语句，Oracle 12c 使用索引来确定子选择的数据。

改变外部选择的where条件

  SELECT  A.CS_ADRESS_ID, A.SEQNUM, TO_CHAR(A.EFFDT, 'YYYY-MM-DD')
    from PS_CS_ADRART_TB A
  where a.cs_kdnr_as400 = '53916917' -- without CS_ADRESSMANDANT condition
     and a.effdt =
         (select  max(b.effdt)
            from ps_cs_adrart_tb b
           where b.cs_adressmandant = a.cs_adressmandant and 
                 b.cs_person_id = a.cs_person_id 
             and b.cs_beziehung_id = a.cs_beziehung_id 
             and b.seqnum = a.seqnum 
             and b.effdt <= trunc(sysdate)
             )

    --------------------------------------------------------------------------------------------------
    | Id  | Operation                             | Name            | Rows | Bytes | Cost | Time     |
    --------------------------------------------------------------------------------------------------
    |   0 | SELECT STATEMENT                      |                 |    1 |    89 |    2 | 00:00:01 |
    |   1 |   TABLE ACCESS BY INDEX ROWID BATCHED | PS_CS_ADRART_TB |    1 |    89 |    1 | 00:00:01 |
    | * 2 |    INDEX RANGE SCAN                   | PS1CS_ADRART_TB |    1 |       |    1 | 00:00:01 |
    |   3 |     SORT AGGREGATE                    |                 |    1 |    59 |      |          |
    |   4 |      FIRST ROW                        |                 |    1 |    59 |    1 | 00:00:01 |
    | * 5 |       INDEX RANGE SCAN (MIN/MAX)      | PS_CS_ADRART_TB |    1 |    59 |    1 | 00:00:01 |
    --------------------------------------------------------------------------------------------------

    Predicate Information (identified by operation id):
    ------------------------------------------
    * 2 - access("A"."CS_KDNR_AS400"='53916917')
    * 2 - filter("A"."EFFDT"= (SELECT MAX("B"."EFFDT") FROM "PS_CS_ADRART_TB" "B" WHERE "B"."EFFDT"<=TRUNC(SYSDATE@!) AND "B"."SEQNUM"=:B1 AND "B"."CS_BEZIEHUNG_ID"=:B2 AND "B"."CS_PERSON_ID"=:B3 AND
      "B"."CS_ADRESSMANDANT"=:B4))
    * 5 - access("B"."CS_ADRESSMANDANT"=:B1 AND "B"."CS_PERSON_ID"=:B2 AND "B"."CS_BEZIEHUNG_ID"=:B3 AND "B"."SEQNUM"=:B4 AND "B"."EFFDT"<=TRUNC(SYSDATE@!))


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

在子选择中使用min() 而不是max()

    SELECT A.CS_ADRESS_ID, A.SEQNUM, TO_CHAR(A.EFFDT, 'YYYY-MM-DD'), a.cs_person_id 
      from PS_CS_ADRART_TB A
    where a.cs_kdnr_as400 = '63916917' and a.cs_adressmandant = '001'
       and a.effdt =
           (select  min(b.effdt)
              from ps_cs_adrart_tb b
             where b.cs_adressmandant = a.cs_adressmandant
               and b.cs_person_id = a.cs_person_id
               and b.cs_beziehung_id = a.cs_beziehung_id
               and b.seqnum = a.seqnum
               and b.effdt  <= sysdate
               );

    ---------------------------------------------------------------------------------------------------
    | Id  | Operation                              | Name            | Rows | Bytes | Cost | Time     |
    ---------------------------------------------------------------------------------------------------
    |   0 | SELECT STATEMENT                       |                 |    1 |   109 |    2 | 00:00:01 |
    |   1 |   NESTED LOOPS                         |                 |    1 |   109 |    2 | 00:00:01 |
    |   2 |    TABLE ACCESS BY INDEX ROWID BATCHED | PS_CS_ADRART_TB |    1 |    89 |    1 | 00:00:01 |
    | * 3 |     INDEX RANGE SCAN                   | PS1CS_ADRART_TB |    1 |       |    1 | 00:00:01 |
    | * 4 |    VIEW PUSHED PREDICATE               | VW_SQ_1         |    1 |    20 |    1 | 00:00:01 |
    |   5 |     SORT GROUP BY                      |                 |    1 |    59 |    1 | 00:00:01 |
    | * 6 |      INDEX RANGE SCAN                  | PS_CS_ADRART_TB |    1 |    59 |    1 | 00:00:01 |
    ---------------------------------------------------------------------------------------------------

    Predicate Information (identified by operation id):
    ------------------------------------------
    * 3 - access("A"."CS_KDNR_AS400"='63916917' AND "A"."CS_ADRESSMANDANT"='001')
    * 4 - filter("A"."EFFDT"="MIN(B.EFFDT)" AND "ITEM_1"="A"."CS_ADRESSMANDANT")
    * 6 - access("B"."CS_ADRESSMANDANT"='001' AND "B"."CS_PERSON_ID"="A"."CS_PERSON_ID" AND "B"."CS_BEZIEHUNG_ID"="A"."CS_BEZIEHUNG_ID" AND "B"."SEQNUM"="A"."SEQNUM" AND "B"."EFFDT"<=SYSDATE@!)


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

重建索引（如上所述）并更新统计数据

EXEC DBMS_STATS.delete_table_stats('SYSADM', 'PS_CS_ADRART_TB');
-- EXEC DBMS_STATS.gather_table_stats('SYSADM', 'PS_CS_ADRART_TB');
EXEC DBMS_STATS.gather_table_stats(ownname => 'SYSADM', tabname =>  'PS_CS_ADRART_TB', cascade => true );

没用。

这里发生了什么？任何建议都将受到高度赞赏。提前致谢。

【问题讨论】：

请让我知道你为什么不投这个问题。
问题很清楚，标签很好，包含构建所需结构的所有语句，提供有关表大小的信息，包含许多尝试和每个语句的计划......我可以'不知道这种问题还应该包含什么... +1
12c 计划中的“用于此语句的动态采样 似乎表明尽管有gather_table_stats()，但统计信息不是最新的。如果你打电话给gather_table_stats() 和cascade => true？（为了排除显而易见的问题：SYSADM 真的是这些表的所有者吗？）
@a_horse_with_no_name 我们使用选项cascade => true 调用DBMS_STATS.gather_table_stats，结果相同。是的，SYSADM 是该表的所有者。该数据库是我们的 Oracle 11g 数据库的升级副本。
两个数据库中参数OPTIMIZER_DYNAMIC_SAMPLING的值是多少？

标签： oracle oracle11g oracle12c

【解决方案1】：

我认为您应该将参数_unnest_subquery 设置为FALSE。

你可以试试

alter session set "_unnest_subquery" = FALSE;

【讨论】：

就是这样！ Oracle 支持建议将隐藏参数 _unnest_subquery 设置为 FALSE 以提高 Oracle 数据库上 PeopleSoft 应用程序的性能。这与bobbydurrettdba.com/2012/11/07/… 中描述的行为相同

【解决方案2】：

弗兰克。据我了解，您希望通过任何 cs_person_id 和 cs_beziehung_id 获得最后一行，其中 b.effdt 低于或等于今天。首先，您可以收集有关选择性信息的统计数据。它帮助 CBO 做出选择。

BEGIN 
 dbms_stats.Gather_table_stats('SYSADM', 'PS_CS_ADRART_TB',
 method_opt => 'FOR COLUMNS SIZE 254 CS_KDNR_AS400 CS_ADRESSMANDANT'); 
END; 
/

另一个人认为我发现了什么。您有很多索引，可能会被创建为唯一的。它已经包含唯一索引列。

create unique index PS1CS_ADRART_TB on PS_CS_ADRART_TB (CS_KDNR_AS400, CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT)  ;
create unique index PS0CS_ADRART_TB on PS_CS_ADRART_TB (CS_ADRESS_ID, CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT);
create unique index PS1CS_ADRART_TB on PS_CS_ADRART_TB (CS_KDNR_AS400, CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT);
create unique index PS2CS_ADRART_TB on PS_CS_ADRART_TB (ROW_ADDED_DTTM, CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT);
create unique index PS3CS_ADRART_TB on PS_CS_ADRART_TB (ROW_ADDED_OPRID, CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT);
create unique index PS4CS_ADRART_TB on PS_CS_ADRART_TB (ROW_LASTMANT_DTTM, CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT);
create unique index PS5CS_ADRART_TB on PS_CS_ADRART_TB (ROW_LASTMANT_OPRID, CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT);
create unique index PS6CS_ADRART_TB on PS_CS_ADRART_TB (CS_RECSTAT_XL, CS_ADRESSMANDANT, CS_PERSON_ID, CS_BEZIEHUNG_ID, SEQNUM, EFFDT);

最后我尝试重写查询以排除子查询

SELECT cs_adress_id,seqnum, effdt_chr 
FROM   (SELECT /* */ a.cs_adress_id 
          ,a.seqnum 
          ,to_char(a.effdt, 'YYYY-MM-DD') AS effdt_chr 
          ,dense_rank() over(PARTITION BY a.cs_adressmandant, a.cs_person_id
                                        , a.cs_beziehung_id, a.seqnum 
                                 ORDER BY a.effdt DESC) AS  rnk 
    FROM   ps_cs_adrart_tb a 
    WHERE  a.cs_adressmandant = '001' 
      AND  a.cs_kdnr_as400 = '63916917' 
      AND  a.effdt <= trunc(sysdate) 
) 
WHERE  rnk = 1;

【讨论】：

是的，该语句选择给定人员的当前数据。解释计划中的统计值是最新的（CS_ADRESSMANDANT=001 和 EFFDT=sysdate 的 12746381 行是一个足够的值），因此问题是为什么不使用该索引的其他字段 CS_PERSON_ID、CS_BEZIEHUNG_ID 和 SEQNUM。是的，我们可以更改 select 语句或索引，但我们会尽量减少对应用程序的更改，以避免工作量和副作用带来的风险。
尽管如此，我已经用method_opt => 'FOR COLUMNS SIZE 254 CS_KDNR_AS400 CS_ADRESSMANDANT' 尝试过 - 但没有成功。
它看起来像 hash_group_by 的问题。能否请您检查带有提示/*+ opt_param('_optimizer_aggr_groupby_elim', 'true')*/ 的计划查询？
这个提示完全不影响执行计划。