選擇組合索引的前導列，必須根據具體的業務（SQL）寫法和列的資料分佈不同而不同，很多書或網上都說，前導列要選擇高選擇性的，但是，脫離具體的業務，這些是沒有意義的，本文就舉一些常見的例子來分析下如何正確選擇前導列,以拋磚引玉，實際應用中，有更多複雜的情況需要具體分析。

1.都是等值條件的列，誰做前導列都一樣

DROP TABLE t; CREATE TABLE t  AS SELECT * FROM dba_objects; CREATE INDEX idx1_t ON t(owner,object_id); CREATE INDEX idx2_t ON t(object_id,owner); BEGIN   dbms_stats.gather_table_stats(ownname => USER,tabname => 'T',estimate_percent => 100,cascade => TRUE); END;  / [email protected]>  SELECT COUNT(DISTINCT owner),COUNT(DISTINCT object_id),COUNT(*) FROM t; COUNT(DISTINCTOWNER) COUNT(DISTINCTOBJECT_ID)   COUNT(*) -------------------- ------------------------ ----------                   33                    75250      75251 1 row selected.

owner有33個不同的值，object_id有75250,顯然object_id的選擇性更好。但是下面的查詢，應用idx1_t與idx2_t的效能一樣(COST與CONSISTENT GETS一樣）。

[email protected]>  SELECT/*+index(t idx1_t)*/ * FROM t   2   WHERE owner='DINGJUN123' AND object_id=75677; 1 row selected. Elapsed: 00:00:00.00 Execution Plan ---------------------------------------------------------- Plan hash value: 2071967826 -------------------------------------------------------------------------------------- | Id  | Operation                   | Name   | Rows  | Bytes | Cost (%CPU)| Time     | -------------------------------------------------------------------------------------- |   0 | SELECT STATEMENT            |        |     1 |    97 |     2   (0)| 00:00:01 | |   1 |  TABLE ACCESS BY INDEX ROWID| T      |     1 |    97 |     2   (0)| 00:00:01 | |*  2 |   INDEX RANGE SCAN          | IDX1_T |     1 |       |     1   (0)| 00:00:01 | -------------------------------------------------------------------------------------- Predicate Information (identified by operation id): ---------------------------------------------------    2 - access("OWNER"='DINGJUN123' AND "OBJECT_ID"=75677) Statistics ----------------------------------------------------------           1  recursive calls           0  db block gets           4  consistent gets           0  physical reads           0  redo size        1403  bytes sent via SQL*Net to client         416  bytes received via SQL*Net from client           2  SQL*Net roundtrips to/from client           0  sorts (memory)           0  sorts (disk)           1  rows processed [email protected]>   SELECT/*+index(t idx2_t)*/ * FROM t   2   WHERE owner='DINGJUN123' AND object_id=75677; 1 row selected. Elapsed: 00:00:00.01 Execution Plan ---------------------------------------------------------- Plan hash value: 3787301248 -------------------------------------------------------------------------------------- | Id  | Operation                   | Name   | Rows  | Bytes | Cost (%CPU)| Time     | -------------------------------------------------------------------------------------- |   0 | SELECT STATEMENT            |        |     1 |    97 |     2   (0)| 00:00:01 | |   1 |  TABLE ACCESS BY INDEX ROWID| T      |     1 |    97 |     2   (0)| 00:00:01 | |*  2 |   INDEX RANGE SCAN          | IDX2_T |     1 |       |     1   (0)| 00:00:01 | -------------------------------------------------------------------------------------- Predicate Information (identified by operation id): ---------------------------------------------------    2 - access("OBJECT_ID"=75677 AND "OWNER"='DINGJUN123') Statistics ----------------------------------------------------------           1  recursive calls           0  db block gets           4  consistent gets           0  physical reads           0  redo size        1403  bytes sent via SQL*Net to client         416  bytes received via SQL*Net from client           2  SQL*Net roundtrips to/from client           0  sorts (memory)           0  sorts (disk)           1  rows processed

      雖然如此，但是要記住，這個索引不是僅僅給這2條SQL使用的，事實上可能我們有的查詢謂詞只有owner或object_id,這時候得考慮使用owner作為前導列還是使用object_id作為前導列。
    還有其他引用owner,object_id的情況，比如GROUP BY ,ORDER BY,甚至SELECT...都需要進行整體的分析，這樣才能建立最佳的索引。
 
2.有的列是大於(等於)或小於(等於）或者是like 模糊匹配等不等條件，有的列是等值的條件，等值的一般作為前導列更好
--做5次，增加幾十萬行SYS的進去
INSERT INTO t SELECT * FROM t WHERE owner='SYS';
COMMIT;
--重新收集統計資訊（省略）

[email protected]> SELECT * FROM t   2  WHERE owner='DINGJUN123'   3  AND object_id>=107889; 1 row selected. Elapsed: 00:00:00.01 Execution Plan ---------------------------------------------------------- Plan hash value: 2071967826 -------------------------------------------------------------------------------------- | Id  | Operation                   | Name   | Rows  | Bytes | Cost (%CPU)| Time     | -------------------------------------------------------------------------------------- |   0 | SELECT STATEMENT            |        |     1 |    96 |     4   (0)| 00:00:01 | |   1 |  TABLE ACCESS BY INDEX ROWID| T      |     1 |    96 |     4   (0)| 00:00:01 | |*  2 |   INDEX RANGE SCAN          | IDX1_T |     1 |       |     3   (0)| 00:00:01 | -------------------------------------------------------------------------------------- Predicate Information (identified by operation id): ---------------------------------------------------    2 - access("OWNER"='DINGJUN123' AND "OBJECT_ID">=107889 AND "OBJECT_ID" IS               NOT NULL) Statistics ----------------------------------------------------------           1  recursive calls           0  db block gets           5  consistent gets           0  physical reads           0  redo size        1399  bytes sent via SQL*Net to client         416  bytes received via SQL*Net from client           2  SQL*Net roundtrips to/from client           0  sorts (memory)           0  sorts (disk)           1  rows processed

      上面的SQL走idx1_t,注意觀察謂詞，只有access,說明索引完全被利用上，很顯然因為owner是前導列，而且是等值查詢，按照前導列查詢，然後只要分析索引的第2列object_id,當發現不滿足條件object_id>=107889之後就停止了，索引掃描沒有浪費。

[email protected]> SELECT/*+index(t idx2_t)*/ * FROM t   2  WHERE owner='DINGJUN123'   3  AND object_id>=107889; 1 row selected. Elapsed: 00:00:00.01 Execution Plan ---------------------------------------------------------- Plan hash value: 3787301248 -------------------------------------------------------------------------------------- | Id  | Operation                   | Name   | Rows  | Bytes | Cost (%CPU)| Time     | -------------------------------------------------------------------------------------- |   0 | SELECT STATEMENT            |        |     1 |    96 |     4   (0)| 00:00:01 | |   1 |  TABLE ACCESS BY INDEX ROWID| T      |     1 |    96 |     4   (0)| 00:00:01 | |*  2 |   INDEX RANGE SCAN          | IDX2_T |     1 |       |     3   (0)| 00:00:01 | -------------------------------------------------------------------------------------- Predicate Information (identified by operation id): ---------------------------------------------------    2 - access("OBJECT_ID">=107889 AND "OWNER"='DINGJUN123' AND "OBJECT_ID" IS               NOT NULL)        filter("OWNER"='DINGJUN123') Statistics ----------------------------------------------------------           1  recursive calls           0  db block gets           5  consistent gets           0  physical reads           0  redo size        1399  bytes sent via SQL*Net to client         416  bytes received via SQL*Net from client           2  SQL*Net roundtrips to/from client           0  sorts (memory)           0  sorts (disk)           1  rows processed

      強制使用idx2_t,object_id是前導列，謂詞有access,還有filter,說明索引沒有被完全利用上，這是因為object_id的不是等值查詢，滿足object_id>=107889的，按照順序搜尋
所以，中間可能有一些不滿足owner='DINGJUN123'的，還要filter掉。
這種查詢和不等值條件作為前導列的查詢，一旦object_id>=107889不滿足owner='DINGJUN123'的很多，那麼必然造成過多不必要的索引搜尋，COST與邏輯讀會上升很快，
從而效能急劇下降，因為本例子基本都滿足owner條件，所以沒有啥浪費。但是下面的例子：

[email protected]> SELECT * FROM t   2  WHERE owner='DINGJUN123'   3  AND object_id>=100; 2540 rows selected. Elapsed: 00:00:00.15 Execution Plan ---------------------------------------------------------- Plan hash value: 2071967826 -------------------------------------------------------------------------------------- | Id  | Operation                   | Name   | Rows  | Bytes | Cost (%CPU)| Time     | -------------------------------------------------------------------------------------- |   0 | SELECT STATEMENT            |        |  2539 |   238K|   499   (0)| 00:00:06 | |   1 |  TABLE ACCESS BY INDEX ROWID| T      |  2539 |   238K|   499   (0)| 00:00:06 | |*  2 |   INDEX RANGE SCAN          | IDX1_T |  2539 |       |    12   (0)| 00:00:01 | -------------------------------------------------------------------------------------- Predicate Information (identified by operation id): ---------------------------------------------------    2 - access("OWNER"='DINGJUN123' AND "OBJECT_ID">=100 AND "OBJECT_ID" IS               NOT NULL) Statistics ----------------------------------------------------------           1  recursive calls           0  db block gets         527  consistent gets          21  physical reads           0  redo size      268134  bytes sent via SQL*Net to client        2275  bytes received via SQL*Net from client         171  SQL*Net roundtrips to/from client           0  sorts (memory)           0  sorts (disk)        2540  rows processed

還是使用idx1_t,沒有問題。看下面的強制使用idx2_t的。

[email protected]> SELECT/*+index(t idx2_t)*/ * FROM t   2  WHERE owner='DINGJUN123'   3  AND object_id>=100; 2540 rows selected. Elapsed: 00:00:00.33 Execution Plan ---------------------------------------------------------- Plan hash value: 3787301248 -------------------------------------------------------------------------------------- | Id  | Operation                   | Name   | Rows  | Bytes | Cost (%CPU)| Time     | -------------------------------------------------------------------------------------- |   0 | SELECT STATEMENT            |        |  2539 |   238K|  3762   (1)| 00:00:46 | |   1 |  TABLE ACCESS BY INDEX ROWID| T      |  2539 |   238K|  3762   (1)| 00:00:46 | |*  2 |   INDEX RANGE SCAN          | IDX2_T |  2539 |       |  3274   (1)| 00:00:40 | -------------------------------------------------------------------------------------- Predicate Information (identified by operation id): ---------------------------------------------------    2 - access("OBJECT_ID">=100 AND "OWNER"='DINGJUN123' AND "OBJECT_ID" IS               NOT NULL)        filter("OWNER"='DINGJUN123') Statistics ----------------------------------------------------------           1  recursive calls           0  db block gets        3763  consistent gets           0  physical reads           0  redo size      268134  bytes sent via SQL*Net to client        2275  bytes received via SQL*Net from client         171  SQL*Net roundtrips to/from client           0  sorts (memory)           0  sorts (disk)        2540  rows processed [email protected]> SELECT COUNT(*) FROM t WHERE object_id >= 100;   COUNT(*) ----------    1032649

 SELECT COUNT(*) FROM t WHERE object_id >= 100;
返回1032649行，但是WHERE owner='DINGJUN123' AND object_id>=100 只返回2540行，要filter掉百萬行，輪詢索引，造成了極大的浪費。


3.如果都是比較，都是<,>之類的表示式
  這種情況，前導列，根據謂詞，選擇條件能夠定位最接近處理結果的基數，並能夠減少索引後filter的工作，因為必然有一列是要走access之後的filter,最好是filter能夠過濾較少資料，不要做過多過濾。
  
例如：

[email protected]> SELECT * FROM t   2  WHERE owner>='DINGJUN123'   3  AND object_id>=107872; 37 rows selected. Elapsed: 00:00:00.00 Execution Plan ---------------------------------------------------------- Plan hash value: 3787301248 -------------------------------------------------------------------------------------- | Id  | Operation                   | Name   | Rows  | Bytes | Cost (%CPU)| Time     | -------------------------------------------------------------------------------------- |   0 | SELECT STATEMENT            |        |   205 | 19680 |    43   (0)| 00:00:01 | |   1 |  TABLE ACCESS BY INDEX ROWID| T      |   205 | 19680 |    43   (0)| 00:00:01 | |*  2 |   INDEX RANGE SCAN          | IDX2_T |   205 |       |     3   (0)| 00:00:01 | -------------------------------------------------------------------------------------- Predicate Information (identified by operation id): ---------------------------------------------------    2 - access("OBJECT_ID">=107872 AND "OWNER">='DINGJUN123' AND "OBJECT_ID"               IS NOT NULL)        filter("OWNER">='DINGJUN123') Statistics ----------------------------------------------------------           1  recursive calls           0  db block gets          14  consistent gets           0  physical reads           0  redo size        6468  bytes sent via SQL*Net to client         438  bytes received via SQL*Net from client           4  SQL*Net roundtrips to/from client           0  sorts (memory)           0  sorts (disk)          37  rows processed [email protected]> SELECT COUNT(*) FROM t WHERE object_id>=107872;   COUNT(*) ----------         37 1 row selected.

  關閉index SKIP SCAN,因為owner種類很少，oracle選擇skip SCAN
 alter session set "_optimizer_skip_scan_enabled" = false;
 

[email protected]> SELECT/*+index(t idx1_t)*/ * FROM t   2  WHERE owner>='DINGJUN123'   3  AND object_id>=107872; 37 rows selected. Elapsed: 00:00:00.23 Execution Plan ---------------------------------------------------------- Plan hash value: 2071967826 -------------------------------------------------------------------------------------- | Id  | Operation                   | Name   | Rows  | Bytes | Cost (%CPU)| Time     | -------------------------------------------------------------------------------------- |   0 | SELECT STATEMENT            |        |   205 | 19680 |  3740   (1)| 00:00:45 | |   1 |  TABLE ACCESS BY INDEX ROWID| T      |   205 | 19680 |  3740   (1)| 00:00:45 | |*  2 |   INDEX RANGE SCAN          | IDX1_T |   205 |       |  3700   (1)| 00:00:45 | -------------------------------------------------------------------------------------- Predicate Information (identified by operation id): ---------------------------------------------------    2 - access("OWNER">='DINGJUN123' AND "OBJECT_ID">=107872 AND "OWNER" IS               NOT NULL)        filter("OBJECT_ID">=107872) Statistics ----------------------------------------------------------           1  recursive calls           0  db block gets  3688  consistent gets           0  physical reads           0  redo size        6468  bytes sent via SQL*Net to client         438  bytes received via SQL*Net from client           4  SQL*Net roundtrips to/from client           0  sorts (memory)           0  sorts (disk)          37  rows processed

因為owner>='DINGJUN123'返回大量行，但是事實結果很少，只有幾十行，過濾object_id>=107872,需要做大量工作，邏輯讀和COST增大千倍+，效能低下。


後記：
     當然如何選擇前導列的順序很複雜，得全盤考慮對應的謂詞，SELECT的列等要素，還要考慮ORDER BY ,GROUP BY等列，比如3列組合索引，如何考慮順序。
     後續會補充更多的組合索引如何建立的要點。