2014-06-07 07:19:30
来 源
中存储网
Oracle教程
这次接到同事求助,说有一个select query,在Oracle上要跑一分多钟,他希望能在5s内出结果,以下就是解决这个问题的方法,相信会对各位网友学习和提高Oracle水平有帮助。

该sql如下:
相关操作代码如下:

Select  /*+ parallel(src, 8) */ distinct
  src.systemname as systemname
  ,  src.databasename as databasename
  ,  src.tablename as tablename
  ,  src.username as username
from  <STRONG>meta_dbql_table_usage_exp_hst</STRONG> src
 inner <STRONG>join DR_QRY_LOG_EXP_HST</STRONG> rl on
  <STRONG>src.acctstringdate = rl.acctstringdate
  and src.queryid = rl.queryid</STRONG>
  And Src.Systemname = Rl.Systemname
  and src.acctstringdate > sysdate - 30
  And Rl.Acctstringdate > Sysdate - 30
 inner join  <STRONG>meta_dr_qry_log_tgt_all_hst </STRONG>tgt on
  upper(tgt.systemname) = upper('MOZART')
  And Upper(tgt.Databasename) = Upper('GDW_TABLES')
  And Upper(tgt.Tablename) = Upper('SSA_SLNG_LSTG_MTRC_SD')
  <STRONG>AND src.acctstringdate = tgt.acctstringdate
  and rl.statement_id = tgt.statement_id</STRONG>
  and rl.systemname = tgt.systemname
  And Tgt.Acctstringdate > Sysdate - 30
  And Not(
    Upper(Tgt.Systemname)=Upper(src.systemname)
    And
    Upper(Tgt.Databasename) = Upper(Src.Databasename)
    And
    Upper(Tgt.Tablename) = Upper(Src.Tablename)
    )
  And   tgt.Systemname is not null
  And   tgt.Databasename Is Not Null
  And   tgt.tablename is not null


SQL的简单分析
总得来看,这个SQL就是三个表(meta_dbql_table_usage_exp_hst,DR_QRY_LOG_EXP_HST,meta_dr_qry_log_tgt_all_hst)的INNER JOIN,这三个表数据量都在百万级别,且都是分区表(以acctstringdate为分区键),执行计划如下:
相关操作代码如下:

------------------------------------------------------------------------------------------------------------------------
| Id  | Operation                              | Name                          | Rows  | Bytes | Cost  | Pstart| Pstop |
------------------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT                       |                               |     1 |   159 |  8654 |       |       |
|   1 |  PX COORDINATOR                        |                               |       |       |       |       |       |
|   2 |   PX SEND QC (RANDOM)                  | :TQ10002                      |     1 |   159 |  8654 |       |       |
|   3 |    SORT UNIQUE                         |                               |     1 |   159 |  8654 |       |       |
|   4 |     PX RECEIVE                         |                               |     1 |    36 |     3 |       |       |
|   5 |      PX SEND HASH                      | :TQ10001                      |     1 |    36 |     3 |       |       |
|*  6 |       TABLE ACCESS BY LOCAL INDEX ROWID| DR_QRY_LOG_EXP_HST            |     1 |    36 |     3 |       |       |
|   7 |        NESTED LOOPS                    |                               |     1 |   159 |  8633 |       |       |
|   8 |         NESTED LOOPS                   |                               |  8959 |  1076K|  4900 |       |       |
|   9 |          BUFFER SORT                   |                               |       |       |       |       |       |
|  10 |           PX RECEIVE                   |                               |       |       |       |       |       |
|  11 |            PX SEND BROADCAST           | :TQ10000                      |       |       |       |       |       |
|  12 |             PARTITION RANGE ITERATOR   |                               |     1 |    56 |  4746 |   KEY |    14 |
|* 13 |              TABLE ACCESS FULL         | META_DR_QRY_LOG_TGT_ALL_HST   |     1 |    56 |  4746 |   KEY |    14 |
|  14 |          PX BLOCK ITERATOR             |                               |  8959 |   586K|   154 |   KEY |   KEY |
|* 15 |           TABLE ACCESS FULL            | META_DBQL_TABLE_USAGE_EXP_HST |  8959 |   586K|   154 |   KEY |   KEY |
|  16 |         PARTITION RANGE ITERATOR       |                               |     1 |       |     2 |   KEY |   KEY |
|* 17 |          INDEX RANGE SCAN              | DR_QRY_LOG_EXP_HST_IDX        |     1 |       |     2 |   KEY |   KEY |
------------------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
   6 - filter("RL"."STATEMENT_ID"="TGT"."STATEMENT_ID" AND "RL"."SYSTEMNAME"="TGT"."SYSTEMNAME" AND "SRC"."SYSTEMNAME"="RL"."SYSTEMNAME")
  13 - filter(UPPER("TGT"."SYSTEMNAME")='MOZART' AND UPPER("TGT"."DATABASENAME")='GDW_TABLES' AND
              UPPER("TGT"."TABLENAME")='SSA_SLNG_LSTG_MTRC_SD' AND "TGT"."ACCTSTRINGDATE">SYSDATE@!-30 AND "TGT"."SYSTEMNAME" IS NOT NULL
              "TGT"."DATABASENAME" IS NOT NULL AND "TGT"."TABLENAME" IS NOT NULL)
  15 - filter("SRC"."ACCTSTRINGDATE"="TGT"."ACCTSTRINGDATE" AND (UPPER("TGT"."SYSTEMNAME")<>UPPER("SRC"."SYSTEMNAME") OR
              UPPER("TGT"."DATABASENAME")<>UPPER("SRC"."DATABASENAME") OR UPPER("TGT"."TABLENAME")<>UPPER("SRC"."TABLENAME")) AND
              "SRC"."ACCTSTRINGDATE">SYSDATE@!-30)
  17 - access("SRC"."QUERYID"="RL"."QUERYID" AND "SRC"."ACCTSTRINGDATE"="RL"."ACCTSTRINGDATE")
       filter("RL"."ACCTSTRINGDATE">SYSDATE@!-30)


定位问题
从上面执行计划中的表连接方式可以知道,这三个表之间进行了两次NESTED LOOP,问题出现在最里层的NESTED LOOP(对两个表都做了TABLE FULL SCAN),因为表都是百万级别的(即时过滤后的数据量也不小),性能问题就出现在内表(即被驱动表)META_DBQL_TABLE_USAGE_EXP_HST做了太多次的全表扫描。如果能把全表扫描转换成索引,则性能可以大幅度提高。

下面是NESTED LOOP的介绍:
嵌套连接把要处理的数据集分为外部循环(驱动数据源)和内部循环(被驱动数据源),外部循环只执行一次,内部循环执行的次数等于外部循环执行返回的数据个数。
这种连接的好处是内存使用非常少。
如果驱动数据源有限,且被驱动表在连接列上有相应的索引,则这种连接方式才是高效的。

下面是这三个表上索引的情况:
相关操作代码如下:

SQL> select index_name, table_name from user_indexes where table_name in ('DR_QRY_LOG_EXP_HST',upper('meta_dbql_table_usage_exp_hst'), upper('meta_dr_qry_log_tgt_all_hs
INDEX_NAME                                                   TABLE_NAME
------------------------------------------------------------ ------------------------------------------------------------
META_DR_QRY_LOG_TGT_ALL_IDX                                  META_DR_QRY_LOG_TGT_ALL_HST
META_DBQL_TUSAGE_EHST_IDX                                    META_DBQL_TABLE_USAGE_EXP_HST
DR_QRY_LOG_EXP_HST_IDX                                       DR_QRY_LOG_EXP_HST
CREATE INDEX "GV"."META_DR_QRY_LOG_TGT_ALL_IDX" ON "GV"."META_DR_QRY_LOG_TGT_ALL_HST" ("STATEMENT_ID", "ACCTSTRINGDATE")
CREATE INDEX "GV"."META_DBQL_TUSAGE_EHST_IDX" ON "GV"."META_DBQL_TABLE_USAGE_EXP_HST" ("QUERYID", "ACCTSTRINGDATE")
CREATE INDEX "GV"."DR_QRY_LOG_EXP_HST_IDX" ON "GV"."DR_QRY_LOG_EXP_HST" ("QUERYID", "ACCTSTRINGDATE")


这三个索引都是本地分区索引(都包含分区键acctstringdate),很显然,DR_QRY_LOG_EXP_HST表少了个索引,因为它与表meta_dr_qry_log_tgt_all_hst 在statement_id上做join,因此应该在它的statement_id上也创建本地分区索引如下:
相关操作代码如下:

create index DR_QRY_LOG_EXP_HST_IDX2 on gv.DR_QRY_LOG_EXP_HST (statement_id,ACCTSTRINGDATE) local;


性能对比
新的执行计划如下:
相关操作代码如下:

------------------------------------------------------------------------------------------------------------------------
| Id  | Operation                              | Name                          | Rows  | Bytes | Cost  | Pstart| Pstop |
------------------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT                       |                               |     1 |   159 |  4838 |       |       |
|   1 |  SORT UNIQUE                           |                               |     1 |   159 |  4838 |       |       |
|*  2 |   TABLE ACCESS BY LOCAL INDEX ROWID    | META_DBQL_TABLE_USAGE_EXP_HST |     1 |    67 |     3 |       |       |
|   3 |    NESTED LOOPS                        |                               |     1 |   159 |  4816 |       |       |
|   4 |     NESTED LOOPS                       |                               |    18 |  1656 |  4762 |       |       |
|   5 |      PARTITION RANGE ITERATOR          |                               |     1 |    56 |  4746 |   KEY |    14 |
|*  6 |       TABLE ACCESS FULL                | META_DR_QRY_LOG_TGT_ALL_HST   |     1 |    56 |  4746 |   KEY |    14 |
|   7 |      PARTITION RANGE ITERATOR          |                               |    18 |   648 |    16 |   KEY |    14 |
|*  8 |       TABLE ACCESS BY LOCAL INDEX ROWID| DR_QRY_LOG_EXP_HST            |    18 |   648 |    16 |   KEY |    14 |
|*  9 |        <STRONG>INDEX RANGE SCAN                | DR_QRY_LOG_EXP_HST_IDX2</STRONG>       |    31 |       |    15 |   KEY |    14 |
|  10 |     PARTITION RANGE ITERATOR           |                               |     1 |       |     2 |   KEY |   KEY |
|* 11 |      INDEX RANGE SCAN                  | META_DBQL_TUSAGE_EHST_IDX     |     1 |       |     2 |   KEY |   KEY |
------------------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
   2 - filter((UPPER("TGT"."SYSTEMNAME")<>UPPER("SRC"."SYSTEMNAME") OR
              UPPER("TGT"."DATABASENAME")<>UPPER("SRC"."DATABASENAME") OR UPPER("TGT"."TABLENAME")<>UPPER("SRC"."TABLENAME"))
              AND "SRC"."SYSTEMNAME"="RL"."SYSTEMNAME")
   6 - filter(UPPER("TGT"."SYSTEMNAME")='MOZART' AND UPPER("TGT"."DATABASENAME")='GDW_TABLES' AND
              UPPER("TGT"."TABLENAME")='SSA_SLNG_LSTG_MTRC_SD' AND "TGT"."ACCTSTRINGDATE">SYSDATE@!-30 AND "TGT"."SYSTEMNAME"
              IS NOT NULL AND "TGT"."DATABASENAME" IS NOT NULL AND "TGT"."TABLENAME" IS NOT NULL)
   8 - filter("RL"."SYSTEMNAME"="TGT"."SYSTEMNAME")
   9 - access("RL"."STATEMENT_ID"="TGT"."STATEMENT_ID" AND "RL"."ACCTSTRINGDATE">SYSDATE@!-30 AND
              "RL"."ACCTSTRINGDATE" IS NOT NULL)
  11 - access("SRC"."QUERYID"="RL"."QUERYID" AND "SRC"."ACCTSTRINGDATE"="RL"."ACCTSTRINGDATE")
       filter("SRC"."ACCTSTRINGDATE"="TGT"."ACCTSTRINGDATE" AND "SRC"."ACCTSTRINGDATE">SYSDATE@!-30)


从新的的执行计划可以看出,它的第一个NESTED LOOP果然用了最新创建的索引。
下面是执行时间:
相关操作代码如下:

已用时间:  00: 00: 02.16


两秒种搞定,远远超出他期望的5s :)
方法总结
NESTED LOOP高效的条件:驱动数据源有限,且被驱动表在连接列上有相应的索引。

声明: 此文观点不代表本站立场;转载须要保留原文链接;版权疑问请联系我们。