一个通过添加本地分区索引提高SQL性能的案例

今天接到同事求助，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，该sql如下：

[sql] view plain copy print ?

Select /*+ parallel(src, 8) */ distinct
src.systemname as systemname
, src.databasename as databasename
, src.tablename as tablename
, src.username as username
from meta_dbql_table_usage_exp_hst src
inner join DR_QRY_LOG_EXP_HST rl on
src.acctstringdate = rl.acctstringdate
and src.queryid = rl.queryid
And Src.Systemname = Rl.Systemname
and src.acctstringdate > sysdate - 30
And Rl.Acctstringdate > Sysdate - 30
inner join meta_dr_qry_log_tgt_all_hst tgt on
upper(tgt.systemname) = upper('MOZART')
And Upper(tgt.Databasename) = Upper('GDW_TABLES')
And Upper(tgt.Tablename) = Upper('SSA_SLNG_LSTG_MTRC_SD')
AND src.acctstringdate = tgt.acctstringdate
and rl.statement_id = tgt.statement_id
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：

[sql] view plain copy print ?

------------------------------------------------------------------------------------------------------------------------
| 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：

[sql] view plain copy print ?

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：

[sql] view plain copy print ?

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：

[sql] view plain copy print ?

------------------------------------------------------------------------------------------------------------------------
| 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 | INDEX RANGE SCAN | DR_QRY_LOG_EXP_HST_IDX2 | 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：

[plain] view plain copy print ?

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。