Oracle 有表連線的connect by 的優化
說明
有1個有表連線，還有connect by 的SQL，整了好幾天才優化成功，感覺像遊戲中打死了只大boss一樣。現將這個過程整理一下。
優化前：
SELECT r.OUT_VER_BEGIN_IDdataID
FROM DMS_DATA_RELA r, DMS_OBJ o
WHERE r.DELETE_FLAG = '0'
AND r.RELA_TYPE_CODE = 'parent'
AND r.OUT_OBJ_CODE = o.OBJ_CODE
AND o.DELETE_FLAG = '0'
AND o.OPEN_STATE = '1'
AND r.IN_OBJ_CODE != 'o_in'
START WITH r.IN_DATA_ID in
(SELECT d.OUT_DATA_ID
FROM DMS_DATA_RELA d
where d.OUT_VER_BEGIN_ID = :1
and d.last_curent_flag = '1')
CONNECT BY r.IN_VER_BEGIN_ID = PRIOR r.OUT_VER_BEGIN_ID
優化後：
我把這個SQL先做了connect by 迴圈，然後再與另1個表做了連線，效果超好，我從李華值 《海量資料庫解決方案》3.2.5 找到相關例子，並有這樣的說明 ： ”如果查詢條件中的列位於同一表中時，並沒有必要優先執行表連線“
select dataID
from(SELECT r.OUT_VER_BEGIN_ID dataID, r.out_obj_code
FROM DMS_DATA_RELA r
WHERE r.DELETE_FLAG = '0'
AND r.RELA_TYPE_CODE = 'parent'
AND r.IN_OBJ_CODE != 'o_in'
STARTWITH r.IN_DATA_ID in
(SELECT d.OUT_DATA_ID
FROM DMS_DATA_RELA d
where d.OUT_VER_BEGIN_ID=:1
and d.last_curent_flag= '1')
CONNECT BY r.IN_VER_BEGIN_ID = PRIOR r.OUT_VER_BEGIN_ID) e,
DMS_OBJ o
wheree.OUT_OBJ_CODE = o.OBJ_CODE
ANDo.DELETE_FLAG = '0'
ANDo.OPEN_STATE = '1'
優化過程：
曾中途一籌莫展時，到劉大的論壇求助過，下面是地址。
http://t.askmaclean.com/thread-3381-1-1.html
下面過程是基於以上的整理。有基本資訊和試過的方法
基本資訊
基本環境
作業系統：windows server 2008 r2 enterprise
SQL> select * from V$version;
BANNER
----------------------------------------------------------------
Oracle Database 10g Enterprise EditionRelease 10.2.0.4.0 - 64bi
PL/SQL Release 10.2.0.4.0 - Production
CORE 10.2.0.4.0 Production
TNS for 64-bit Windows: Version 10.2.0.4.0- Production
NLSRTL Version 10.2.0.4.0 - Production
開發中的庫，在做測試時，發現有1SQL比較消耗資源，我想把他優化一下。
SQL如下：
SELECT r.OUT_VER_BEGIN_ID dataID
FROM DMS_DATA_RELA r, DMS_OBJ o
WHERE r.DELETE_FLAG = '0'
AND r.RELA_TYPE_CODE ='parent'
AND r.OUT_OBJ_CODE =o.OBJ_CODE
AND o.DELETE_FLAG = '0'
AND o.OPEN_STATE = '1'
AND r.IN_OBJ_CODE != 'o_in'
START WITH r.IN_DATA_ID in
(SELECT d.OUT_DATA_ID
FROM DMS_DATA_RELA d
where d.OUT_VER_BEGIN_ID = :1
and d.last_curent_flag = '1')
CONNECT BY r.IN_VER_BEGIN_ID = PRIORr.OUT_VER_BEGIN_ID
執行計劃： （來自awrsqrpt)
0 SELECT STATEMENT 8045 (100)
1 FILTER
2 CONNECT BY WITHFILTERING
3 FILTER
4 COUNT
5 HASH JOIN 717K 111M 8045 (1)00:01:37
6 TABLE ACCESSFULL DMS_OBJ 41 656 3 (0)00:00:01
7 TABLE ACCESSFULL DMS_DATA_RELA 717K 100M 8037 (1)00:01:37
8 TABLE ACCESS BY INDEX ROWIDDMS_DATA_RELA 1 91 4 (0) 00:00:01
9 INDEX RANGE SCAN OUT_VER_BEGIN_ID_INDEX 1 3 (0) 00:00:01
10 HASH JOIN
11 CONNECT BY PUMP
12 COUNT
13 HASH JOIN 717K 111M 8045 (1)00:01:37
14 TABLE ACCESSFULL DMS_OBJ 41 656 3 (0)00:00:01
15 TABLE ACCESSFULL DMS_DATA_RELA 717K 100M 8037 (1)00:01:37
隨便找個變數單獨執行這條SQL試了下，大概能執行6秒。
表上記錄數
SQL> select count(*) from dms_data_rela;
COUNT(*)
----------
858470
SQL> select count(*) from dms_obj;
COUNT(*)
----------
41
表上索引：（索引較多，如果不影響此條SQL的查詢效能，先不考慮這些索引合不合理，因為有些非技術因素）
SQL> selectindex_name,column_name,table_name from dba_ind_columns where table_name='DMS_DATA_RELA' order by index_name;
INDEX_NAME COLUMN_NAME TABLE_NAME
-------------------------------------------------- ------------------------------
CREATE_TIME_INDEX SYS_NC00031$ DMS_DATA_RELA
DELETE_FLAG_INDEX DELETE_FLAG DMS_DATA_RELA
IN_DATA_ID_INDEX IN_DATA_ID DMS_DATA_RELA
IN_DATA_NAME_INDEX IN_DATA_NAME DMS_DATA_RELA
IN_DATA_SOURCE_CODE_INDEX IN_DATA_SOURCE_CODE DMS_DATA_RELA
IN_DATA_SOURCE_ID_INDEX IN_DATA_SOURCE_ID DMS_DATA_RELA
IN_OBJ_CODE_INDEX IN_OBJ_CODE DMS_DATA_RELA
IN_VER_BEGIN_ID_INDEX IN_VER_BEGIN_ID DMS_DATA_RELA
LAST_CURENT_FLAG_INDEX LAST_CURENT_FLAG DMS_DATA_RELA
OUT_DATA_ID_INDEX OUT_DATA_ID DMS_DATA_RELA
OUT_DATA_NAME_INDEX OUT_DATA_NAME DMS_DATA_RELA
OUT_DATA_SOURCE_ID_INDEX OUT_DATA_SOURCE_ID DMS_DATA_RELA
OUT_OBJ_CODE_INDEX OUT_OBJ_CODE DMS_DATA_RELA
OUT_VER_BEGIN_ID_INDEX OUT_VER_BEGIN_ID DMS_DATA_RELA
PK_DMS_DATA_RELA RELA_ID DMS_DATA_RELA
RELA_TYPE_CODE_INDEX RELA_TYPE_CODE DMS_DATA_RELA
SQL> selectindex_name,column_name,table_name from dba_ind_columns where table_name='DMS_OBJ' order by index_name;
INDEX_NAME COLUMN_NAME TABLE_NAME
-------------------------------------------------- ------------------------------
PK_DMS_OBJ OBJ_ID DMS_OBJ
上面SQL走全表，我找了一些資料說是在IN_VER_BEGIN_ID 列上有索���的話，會走索引，但是實際有索引，還是全表，加hint也還是全表。
另外表也收集過統計資訊了。
|
表上的列： |
另外也有人提到清理索引後，效果會好。 首先，索引有些非技術原因不讓清理。另外我弄了1個新環境，沒有過多索引，試過不同的列上建不同的索引，效果也是一樣的。單就此條SQL來說，過多的索引應該影響不大。
有人問索引狀態，也一併附上：
SQL> select table_name, index_name,index_type,statusfrom user_indexes where table_name='DMS_DATA_RELA';
TABLE_NAME INDEX_NAME INDEX_TYPE STATUS
------------------------------------------------------------ --------------------------- --------
DMS_DATA_RELA OUT_VER_BEGIN_ID_INDEX NORMAL VALID
DMS_DATA_RELA IN_DATA_ID_INDEX NORMAL VALID
DMS_DATA_RELA IN_DATA_NAME_INDEX NORMAL VALID
DMS_DATA_RELA IN_DATA_SOURCE_CODE_INDEX NORMAL VALID
DMS_DATA_RELA IN_DATA_SOURCE_ID_INDEX NORMAL VALID
DMS_DATA_RELA IN_OBJ_CODE_INDEX NORMAL VALID
DMS_DATA_RELA IN_VER_BEGIN_ID_INDEX NORMAL VALID
DMS_DATA_RELA OUT_DATA_NAME_INDEX NORMAL VALID
DMS_DATA_RELA OUT_DATA_SOURCE_ID_INDEX NORMAL VALID
DMS_DATA_RELA OUT_OBJ_CODE_INDEX NORMAL VALID
DMS_DATA_RELA RELA_TYPE_CODE_INDEX NORMAL VALID
DMS_DATA_RELA DELETE_FLAG_INDEX NORMAL VALID
DMS_DATA_RELA LAST_CURENT_FLAG_INDEX NORMAL VALID
DMS_DATA_RELA CREATE_TIME_INDEX FUNCTION-BASED NORMAL VALID
DMS_DATA_RELA OUT_DATA_ID_INDEX NORMAL VALID
DMS_DATA_RELA PK_DMS_DATA_RELA NORMAL VALID
16 rows selected
我試過物化檢視
CREATE MATERIALIZED VIEW mv_dms_ddr
REFRESH force
ON demand
WITH ROWID enable query rewrite AS
SELECT r.OUT_VER_BEGIN_ID, r.IN_DATA_ID,r.IN_VER_BEGIN_ID
FROM DMS_DATA_RELA r, DMS_OBJ o
WHERE r.DELETE_FLAG = '0'
AND r.RELA_TYPE_CODE = 'parent'
AND r.IN_OBJ_CODE != 'o_in'
AND r.OUT_OBJ_CODE =o.OBJ_CODE
AND o.DELETE_FLAG = '0'
AND o.OPEN_STATE = '1'
改寫SQL
SELECT OUT_VER_BEGIN_ID dataID
FROM mv_dms_ddr
START WITH IN_DATA_ID in
(SELECT d.OUT_DATA_ID
FROM DMS_DATA_RELA d
where d.OUT_VER_BEGIN_ID = :1
and d.last_curent_flag = '1')
CONNECT BY PRIOR OUT_VER_BEGIN_ID =IN_VER_BEGIN_ID
（沒用fast更新），帶入變數試了，第1次慢，後面快，執行時間大概是3秒多了。有提升，但不是很理想，而且表更新也頻繁，重新整理方式沒做on
commit。（我猜測在資料庫變動頻繁的情況下，是不是這個查詢還是會慢，這個想法還沒驗證）
管理
試過Nested Loop
有人提到nested loop，我前面測過了，效果不怎麼好，現在再把nl的執行計劃附上。執行計劃是代入變數測的。
代入變數值，不加nl的hint
set autot traceonly
SELECT
r.OUT_VER_BEGIN_ID dataID
FROM DMS_DATA_RELA r, DMS_OBJ o
WHERE r.DELETE_FLAG = '0'
AND r.RELA_TYPE_CODE = 'parent'
AND r.OUT_OBJ_CODE = o.OBJ_CODE
AND o.DELETE_FLAG = '0'
AND o.OPEN_STATE = '1'
AND r.IN_OBJ_CODE != 'o_in'
START WITH r.IN_DATA_ID in
(SELECT d.OUT_DATA_ID
FROM DMS_DATA_RELA d
where d.OUT_VER_BEGIN_ID = '20130131036703_syspro_o_wbs'
and d.last_curent_flag = '1')
CONNECT BY r.IN_VER_BEGIN_ID = PRIORr.OUT_VER_BEGIN_ID
已用時間: 00: 00: 06.45
執行計劃
----------------------------------------------------------
Plan hash value: 3423681500
----------------------------------------------------------------------------------------------------
| Id | Operation |Name | Rows | Bytes | Cost (%CPU)| Time |
----------------------------------------------------------------------------------------------------
| 0| SELECT STATEMENT | | 637K| 92M| 6643 (1)| 00:01:20 |
|* 1| FILTER | | | | | |
|* 2| CONNECT BY WITH FILTERING | | | | | |
|* 3| FILTER | | | | | |
| 4| COUNT | | | | | |
|* 5| HASH JOIN | | 637K| 92M| 6643 (1)| 00:01:20 |
| 6| TABLE ACCESS FULL | DMS_OBJ | 63 | 945 | 3 (0)| 00:00:01 |
| 7| TABLE ACCESS FULL | DMS_DATA_RELA | 637K| 83M| 6636 (1)| 00:01:20 |
|* 8| TABLE ACCESS BY INDEX ROWID|DMS_DATA_RELA | 1 | 82 | 4 (0)| 00:00:01 |
|* 9| INDEX RANGE SCAN | OUT_DATA_ID_INDEX | 1 | | 3 (0)| 00:00:01 |
|* 10 | HASH JOIN | | | | | |
| 11| CONNECT BY PUMP | | | | | |
| 12| COUNT | | | | | |
|* 13 | HASH JOIN | | 637K| 92M| 6643 (1)| 00:01:20 |
| 14| TABLE ACCESS FULL | DMS_OBJ | 63 | 945 | 3 (0)| 00:00:01 |
| 15| TABLE ACCESS FULL | DMS_DATA_RELA | 637K| 83M| 6636 (1)| 00:01:20 |
| 16| COUNT | | | | | |
|* 17 | HASH JOIN | | 637K| 92M| 6643 (1)| 00:01:20 |
| 18| TABLE ACCESS FULL | DMS_OBJ | 63 | 945 | 3 (0)| 00:00:01 |
| 19| TABLE ACCESS FULL | DMS_DATA_RELA | 637K| 83M| 6636 (1)| 00:01:20 |
|* 20 | TABLE ACCESS BY INDEX ROWID | DMS_DATA_RELA | 1 | 82 | 4 (0)| 00:00:01 |
|* 21 | INDEX RANGE SCAN |OUT_DATA_ID_INDEX | 1 | | 3 (0)| 00:00:01 |
----------------------------------------------------------------------------------------------------
Predicate Information (identified byoperation id):
---------------------------------------------------
1- filter("R"."DELETE_FLAG"='0' AND"R"."RELA_TYPE_CODE"='parent' AND
"O"."DELETE_FLAG"='0' AND"O"."OPEN_STATE"='1' AND"R"."IN_OBJ_CODE"<>'o_in')
2- filter( EXISTS (SELECT 0 FROM "DMS_DATA_RELA" "D" WHERE"D"."OUT_DATA_ID"=:B1 AND
"D"."OUT_VER_BEGIN_ID"='20130131036703_syspro_o_wbs'AND "D"."LAST_CURENT_FLAG"='1'))
3- filter( EXISTS (SELECT 0 FROM "DMS_DATA_RELA" "D" WHERE"D"."OUT_DATA_ID"=:B1 AND
"D"."OUT_VER_BEGIN_ID"='20130131036703_syspro_o_wbs'AND "D"."LAST_CURENT_FLAG"='1'))
5-access("R"."OUT_OBJ_CODE"="O"."OBJ_CODE")
8-filter("D"."OUT_VER_BEGIN_ID"='20130131036703_syspro_o_wbs'AND
"D"."LAST_CURENT_FLAG"='1')
9- access("D"."OUT_DATA_ID"=:B1)
10- access("R"."IN_VER_BEGIN_ID"=NULL)
13-access("R"."OUT_OBJ_CODE"="O"."OBJ_CODE")
17-access("R"."OUT_OBJ_CODE"="O"."OBJ_CODE")
20-filter("D"."OUT_VER_BEGIN_ID"='20130131036703_syspro_o_wbs'AND
"D"."LAST_CURENT_FLAG"='1')
21- access("D"."OUT_DATA_ID"=:B1)
統計資訊
----------------------------------------------------------
1 recursive calls
0 db block gets
2103709 consistent gets
0 physical reads
0 redo size
596 bytes sent via SQL*Net toclient
350 bytes received via SQL*Netfrom client
2 SQL*Net roundtrips to/fromclient
4 sorts (memory)
0 sorts (disk)
5 rows processed