oracle分頁查詢

Oracle的分頁查詢語句基本上可以按照本文給出的格式來進行套用。

分頁查詢格式：

SELECT * FROM
(
SELECT A.*, ROWNUM RN
FROM (SELECT * FROM TABLE_NAME) A
WHERE ROWNUM <= 40
)
WHERE RN >= 21

其中最內層的查詢SELECT * FROM TABLE_NAME表示不進行翻頁的原始查詢語句。ROWNUM <= 40和RN >= 21控制分頁查詢的每頁的範圍。

上面給出的這個分頁查詢語句，在大多數情況擁有較高的效率。分頁的目的就是控制輸出結果集大小，將結果儘快的返回。在上面的分頁查詢語句中，這種考慮主要體現在WHERE ROWNUM <= 40這句上。

選擇第21到40條記錄存在兩種方法，一種是上面例子中展示的在查詢的第二層通過ROWNUM <= 40來控制最大值，在查詢的最外層控制最小值。而另一種方式是去掉查詢第二層的WHERE ROWNUM <= 40語句，在查詢的最外層控制分頁的最小值和最大值。這是，查詢語句如下：

SELECT * FROM
(
SELECT A.*, ROWNUM RN
FROM (SELECT * FROM TABLE_NAME) A
)
WHERE RN BETWEEN 21 AND 40

對比這兩種寫法，絕大多數的情況下，第一個查詢的效率比第二個高得多。

這是由於CBO優化模式下，Oracle可以將外層的查詢條件推到內層查詢中，以提高內層查詢的執行效率。對於第一個查詢語句，第二層的查詢條件WHERE ROWNUM <= 40就可以被Oracle推入到內層查詢中，這樣Oracle查詢的結果一旦超過了ROWNUM限制條件，就終止查詢將結果返回了。

而第二個查詢語句，由於查詢條件BETWEEN 21 AND 40是存在於查詢的第三層，而Oracle無法將第三層的查詢條件推到最內層（即使推到最內層也沒有意義，因爲最內層查詢不知道RN代表什麼）。因此，對於第二個查詢語句，Oracle最內層返回給中間層的是所有滿足條件的數據，而中間層返回給最外層的也是所有數據。數據的過濾在最外層完成，顯然這個效率要比第一個查詢低得多。

上面分析的查詢不僅僅是針對單表的簡單查詢，對於最內層查詢是複雜的多表聯合查詢或最內層查詢包含排序的情況一樣有效。

這裏就不對包含排序的查詢進行說明了，下一篇文章會通過例子來詳細說明。下面簡單討論一下多表聯合的情況。對於最常見的等值表連接查詢，CBO一般可能會採用兩種連接方式NESTED LOOP和HASH JOIN（MERGE JOIN效率比HASH JOIN效率低，一般CBO不會考慮）。在這裏，由於使用了分頁，因此指定了一個返回的最大記錄數，NESTED LOOP在返回記錄數超過最大值時可以馬上停止並將結果返回給中間層，而HASH JOIN必須處理完所有結果集（MERGE JOIN也是）。那麼在大部分的情況下，對於分頁查詢選擇NESTED LOOP作爲查詢的連接方法具有較高的效率（分頁查詢的時候絕大部分的情況是查詢前幾頁的數據，越靠後面的頁數訪問機率越小）。

因此，如果不介意在系統中使用HINT的話，可以將分頁的查詢語句改寫爲：

SELECT /*+ FIRST_ROWS */ * FROM
(
SELECT A.*, ROWNUM RN
FROM (SELECT * FROM TABLE_NAME) A
WHERE ROWNUM <= 40
)
WHERE RN >= 21

這篇文章用幾個例子來說明分頁查詢的效率。首先構造一個比較大的表作爲測試表：

SQL> CREATE TABLE T AS SELECT * FROM DBA_OBJECTS, DBA_SEQUENCES;

表已創建。

SQL> SELECT COUNT(*) FROM T;

COUNT(*)
----------
457992

首先比較兩種分頁方法的區別：

SQL> SET AUTOT ON
SQL> COL OBJECT_NAME FORMAT A30
SQL> EXEC DBMS_STATS.GATHER_TABLE_STATS(USER, 'T')

PL/SQL 過程已成功完成。

SQL> SELECT OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME FROM T
8 )
9 )
10 WHERE RN BETWEEN 11 AND 20;

OBJECT_ID OBJECT_NAME
---------- ------------------------------
5807 ALL_APPLY_PROGRESS
1769 ALL_ARGUMENTS
2085 ALL_ASSOCIATIONS
4997 ALL_AUDIT_POLICIES
4005 ALL_BASE_TABLE_MVIEWS
5753 ALL_CAPTURE
5757 ALL_CAPTURE_PARAMETERS
5761 ALL_CAPTURE_PREPARED_DATABASE
5765 ALL_CAPTURE_PREPARED_SCHEMAS
5769 ALL_CAPTURE_PREPARED_TABLES

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=864 Card=457992 Bytes=42135264)
1 0 VIEW (Cost=864 Card=457992 Bytes=42135264)
2 1 COUNT
3 2 TABLE ACCESS (FULL) OF 'T' (Cost=864 Card=457992 Bytes=9617832)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
8979 consistent gets
7422 physical reads
0 redo size
758 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10 rows processed

SQL> SELECT OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME FROM T
8 )
9 WHERE ROWNUM <= 20
10 )
11 WHERE RN >= 11;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=864 Card=20 Bytes=1840)
1 0 VIEW (Cost=864 Card=20 Bytes=1840)
2 1 COUNT (STOPKEY)
3 2 TABLE ACCESS (FULL) OF 'T' (Cost=864 Card=457992 Bytes=9617832)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
5 consistent gets
0 physical reads
0 redo size
758 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10 rows processed

二者執行效率相差很大，一個需要8000多邏輯讀，而另一個只需要5個邏輯讀。觀察二者的執行計劃可以發現，兩個執行計劃唯一的區別就是第二個查詢在COUNT這步使用了STOPKEY，也就是說，Oracle將ROWNUM <= 20推入到查詢內層，當符合查詢的條件的記錄達到STOPKEY的值，則Oracle結束查詢。

因此，可以預見，採用第二種方式，在翻頁的開始部分查詢速度很快，越到後面，效率越低，當翻到最後一頁，效率應該和第一種方式接近。

SQL> SELECT OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME FROM T
8 )
9 WHERE ROWNUM <= 457990
10 )
11 WHERE RN >= 457980;

OBJECT_ID OBJECT_NAME
---------- ------------------------------
7128 XCF_I_HANDLE_STATUS
7126 XCF_P
7127 XCF_U1
7142 XDF
7145 XDF_I_DF_KEY
7146 XDF_I_HANDLE_STATUS
7143 XDF_P
7144 XDF_U1
TEST.YANGTINGKUN
TEST4.YANGTINGKUN
YANGTK.YANGTINGKUN

已選擇11行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=864 Card=457990 Bytes=42135080)
1 0 VIEW (Cost=864 Card=457990 Bytes=42135080)
2 1 COUNT (STOPKEY)
3 2 TABLE ACCESS (FULL) OF 'T' (Cost=864 Card=457992 Bytes=9617832)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
8979 consistent gets
7423 physical reads
0 redo size
680 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
11 rows processed

繼續看查詢的第二種情況，包含表連接的情況：

SQL> CREATE TABLE T AS SELECT * FROM DBA_USERS;

表已創建。

SQL> CREATE TABLE T1 AS SELECT * FROM DBA_SOURCE;

表已創建。

SQL> ALTER TABLE T ADD CONSTRAINT PK_T PRIMARY KEY (USERNAME);

表已更改。

SQL> ALTER TABLE T1 ADD CONSTRAINT FK_T1_OWNER FOREIGN KEY (OWNER)
2 REFERENCES T(USERNAME);

表已更改。

SQL> CREATE INDEX IND_T1_OWNER ON T1(NAME);

索引已創建。

SQL> EXEC DBMS_STATS.GATHER_TABLE_STATS(USER, 'T')

PL/SQL 過程已成功完成。

SQL> EXEC DBMS_STATS.GATHER_TABLE_STATS(USER, 'T1')

PL/SQL 過程已成功完成。

創建了T表和T1表，默認情況下，HASH JOIN的效率要比NESTED LOOP高很多：

SQL> SET AUTOT TRACE
SQL> SELECT * FROM T, T1 WHERE T.USERNAME = T1.OWNER;

已選擇96985行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=844 Card=96985 Bytes=46164860)
1 0 HASH JOIN (Cost=844 Card=96985 Bytes=46164860)
2 1 TABLE ACCESS (FULL) OF 'T' (Cost=2 Card=12 Bytes=1044)
3 1 TABLE ACCESS (FULL) OF 'T1' (Cost=826 Card=96985 Bytes=37727165)

Statistics
----------------------------------------------------------
39 recursive calls
0 db block gets
14475 consistent gets
7279 physical reads
0 redo size
37565579 bytes sent via SQL*Net to client
71618 bytes received via SQL*Net from client
6467 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
96985 rows processed

SQL> SELECT /*+ FIRST_ROWS */ * FROM T, T1 WHERE T.USERNAME = T1.OWNER;

已選擇96985行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: FIRST_ROWS (Cost=97811 Card=96985 Bytes=46164860)
1 0 NESTED LOOPS (Cost=97811 Card=96985 Bytes=46164860)
2 1 TABLE ACCESS (FULL) OF 'T1' (Cost=826 Card=96985 Bytes=37727165)
3 1 TABLE ACCESS (BY INDEX ROWID) OF 'T' (Cost=1 Card=1 Bytes=87)
4 3 INDEX (UNIQUE SCAN) OF 'PK_T' (UNIQUE)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
117917 consistent gets
7268 physical reads
0 redo size
37565579 bytes sent via SQL*Net to client
71618 bytes received via SQL*Net from client
6467 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
96985 rows processed

但是如果分頁查詢的內層是這種連接查詢的話，使用NESTED LOOP可以更快的得到前N條記錄。

下面看一下這種情況下的分頁查詢情況：

SQL> SELECT USER_ID, USERNAME, NAME
2 FROM
3 (
4 SELECT ROWNUM RN, USER_ID, USERNAME, NAME
5 FROM
6 (
7 SELECT T.USER_ID, T.USERNAME, T1.NAME
8 FROM T, T1
9 WHERE T.USERNAME = T1.OWNER
10 )
11 WHERE ROWNUM <= 20
12 )
13 WHERE RN >= 11;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=830 Card=20 Bytes=1200)
1 0 VIEW (Cost=830 Card=20 Bytes=1200)
2 1 COUNT (STOPKEY)
3 2 HASH JOIN (Cost=830 Card=96985 Bytes=2909550)
4 3 TABLE ACCESS (FULL) OF 'T' (Cost=2 Card=12 Bytes=132)
5 3 TABLE ACCESS (FULL) OF 'T1' (Cost=826 Card=96985 Bytes=1842715)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
8 consistent gets
7 physical reads
0 redo size
574 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10 rows processed

SQL> SELECT /*+ FIRST_ROWS */ USER_ID, USERNAME, NAME
2 FROM
3 (
4 SELECT ROWNUM RN, USER_ID, USERNAME, NAME
5 FROM
6 (
7 SELECT T.USER_ID, T.USERNAME, T1.NAME
8 FROM T, T1
9 WHERE T.USERNAME = T1.OWNER
10 )
11 WHERE ROWNUM <= 20
12 )
13 WHERE RN >= 11;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: FIRST_ROWS (Cost=97811 Card=20 Bytes=1200)
1 0 VIEW (Cost=97811 Card=20 Bytes=1200)
2 1 COUNT (STOPKEY)
3 2 NESTED LOOPS (Cost=97811 Card=96985 Bytes=2909550)
4 3 TABLE ACCESS (FULL) OF 'T1' (Cost=826 Card=96985 Bytes=1842715)
5 3 TABLE ACCESS (BY INDEX ROWID) OF 'T' (Cost=1 Card=1 Bytes=11)
6 5 INDEX (UNIQUE SCAN) OF 'PK_T' (UNIQUE)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
28 consistent gets
0 physical reads
0 redo size
574 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10 rows processed

看上去似乎HASH JOIN效率更高，難道上面說錯了。

其實這個現象是由於這個例子的特殊性造成的。T表是根據DBA_USERS創建，這張表很小。HASH JOIN中第一步也就是第一張表的全表掃描是無法應用STOPKEY的，這就是上面提到的NESTED LOOP比HASH JOIN優勢的地方。但是，這個例子中，恰好第一張表很小，對這張表的全掃描的代價極低，因此，顯得HASH JOIN效率更高。但是，這不具備共性，如果兩張表的大小相近，或者Oracle錯誤的選擇了先掃描大表，則使用HASH JOIN的效率就會低得多。

SQL> SELECT USER_ID, USERNAME, NAME
2 FROM
3 (
4 SELECT ROWNUM RN, USER_ID, USERNAME, NAME
5 FROM
6 (
7 SELECT /*+ ORDERED */ T.USER_ID, T.USERNAME, T1.NAME
8 FROM T1, T
9 WHERE T.USERNAME = T1.OWNER
10 )
11 WHERE ROWNUM <= 20
12 )
13 WHERE RN >= 11;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=951 Card=20 Bytes=1200)
1 0 VIEW (Cost=951 Card=20 Bytes=1200)
2 1 COUNT (STOPKEY)
3 2 HASH JOIN (Cost=951 Card=96985 Bytes=2909550)
4 3 TABLE ACCESS (FULL) OF 'T1' (Cost=826 Card=96985 Bytes=1842715)
5 3 TABLE ACCESS (FULL) OF 'T' (Cost=2 Card=12 Bytes=132)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
8585 consistent gets
7310 physical reads
0 redo size
601 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10 rows processed

通過HINT提示，讓Oracle先掃描大表，這回結果就很明顯了。NESTED LOOP的效果要比HASH JOIN好得多。

下面，繼續比較一下兩個分頁操作的寫法，爲了使結果更具有代表性，這裏都採用了FIRST_ROWS提示，讓Oracle採用NESTED LOOP的方式來進行表連接：

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: FIRST_ROWS (Cost=97811 Card=96985 Bytes=5819100)
1 0 VIEW (Cost=97811 Card=96985 Bytes=5819100)
2 1 COUNT
3 2 NESTED LOOPS (Cost=97811 Card=96985 Bytes=2909550)
4 3 TABLE ACCESS (FULL) OF 'T1' (Cost=826 Card=96985 Bytes=1842715)
5 3 TABLE ACCESS (BY INDEX ROWID) OF 'T' (Cost=1 Card=1 Bytes=11)
6 5 INDEX (UNIQUE SCAN) OF 'PK_T' (UNIQUE)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
105571 consistent gets
7299 physical reads
0 redo size
574 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10 rows processed

兩種寫法的效率差別極大。關鍵仍然是是否能將STOPKEY應用到最內層查詢中。

對於表連接來說，在寫分頁查詢的時候，可以考慮增加FIRST_ROWS提示，它有助於更快的將查詢結果返回。

其實，不光是表連接，對於所有的分頁查詢都可以加上FIRST_ROWS提示。不過需要注意的時，分頁查詢的目標是儘快的返回前N條記錄，因此，無論是ROWNUM還是FIRST_ROWS機制都是提高前幾頁的查詢速度，對於分頁查詢的最後幾頁，採用這些機制不但無法提高查詢速度，反而會明顯降低查詢效率，對於這一點使用者應該做到心中有數。

最後的例子說明內部循環包含排序的情況：

SQL> CREATE TABLE T AS SELECT * FROM DBA_OBJECTS;

表已創建。

SQL> CREATE INDEX IND_T_OBJECT_NAME ON T (OBJECT_NAME);

索引已創建。

SQL> ALTER TABLE T MODIFY OBJECT_NAME NOT NULL;

表已更改。

SQL> EXEC DBMS_STATS.GATHER_TABLE_STATS(USER, 'T')

PL/SQL 過程已成功完成。

下面進行測試包含排序操作的分頁查詢。可以簡單的將查詢分爲兩種不同情況，第一種排序列就是索引列，這種可以利用索引讀取，第二種排序列沒有索引。

第一種情況又可以細分爲：完全索引掃描和通過索引掃描定位到表記錄兩種情況。

無論是那種情況，都可以通過索引的全掃描來避免排序的產生。看下面的例子：

SQL> SET AUTOT TRACE
SQL> SELECT OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_NAME FROM T ORDER BY OBJECT_NAME
8 )
9 WHERE ROWNUM <= 20
10 )
11 WHERE RN >= 11;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=26 Card=20 Bytes=1580)
1 0 VIEW (Cost=26 Card=20 Bytes=1580)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=26 Card=6361 Bytes=419826)
4 3 INDEX (FULL SCAN) OF 'IND_T_OBJECT_NAME' (NON-UNIQUE) (Cost=26 Card=6361 Bytes=108137)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
3 consistent gets
0 physical reads
0 redo size
576 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10 rows processed

這種情況下，通過索引可以完全得到查詢的結果，因此可以避免表掃描的產生，而且，由於索引已經是排序過的，因此通過索引的全掃描，連排序操作都省略了。

SQL> SELECT OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME FROM T ORDER BY OBJECT_NAME
8 )
9 WHERE ROWNUM <= 20
10 )
11 WHERE RN >= 11;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=43 Card=20 Bytes=1840)
1 0 VIEW (Cost=43 Card=20 Bytes=1840)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=43 Card=6361 Bytes=502519)
4 3 SORT (ORDER BY STOPKEY) (Cost=43 Card=6361 Bytes=133581)
5 4 TABLE ACCESS (FULL) OF 'T' (Cost=9 Card=6361 Bytes=133581)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
81 consistent gets
0 physical reads
0 redo size
673 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
10 rows processed

由於不能僅僅通過索引掃描得到查詢結果，這裏Oracle選擇了表掃描。這是由於初始化參數設置決定的。因此，建議在分頁的時候使用FIRST_ROWS提示。

SQL> SELECT /*+ FIRST_ROWS */ OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME FROM T ORDER BY OBJECT_NAME
8 )
9 WHERE ROWNUM <= 20
10 )
11 WHERE RN >= 11;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: FIRST_ROWS (Cost=826 Card=20 Bytes=1840)
1 0 VIEW (Cost=826 Card=20 Bytes=1840)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=826 Card=6361 Bytes=502519)
4 3 TABLE ACCESS (BY INDEX ROWID) OF 'T' (Cost=826 Card=6361 Bytes=133581)
5 4 INDEX (FULL SCAN) OF 'IND_T_OBJECT_NAME' (NON-UNIQUE) (Cost=26 Card=6361)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
22 consistent gets
0 physical reads
0 redo size
673 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10 rows processed

使用了FIRST_ROWS提示後，Oracle不需要掃描全表，而且避免了排序操作。

下面討論最後一種情況，排序列不是索引列。這個時候排序不可避免，但是利用給出分頁格式，Oracle不會對所有數據進行排序，而是隻排序前N條記錄。

SQL> SELECT OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME FROM T ORDER BY TIMESTAMP
8 )
9 )
10 WHERE RN BETWEEN 11 AND 20;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=64 Card=6361 Bytes=585212)
1 0 VIEW (Cost=64 Card=6361 Bytes=585212)
2 1 COUNT
3 2 VIEW (Cost=64 Card=6361 Bytes=502519)
4 3 SORT (ORDER BY) (Cost=64 Card=6361 Bytes=260801)
5 4 TABLE ACCESS (FULL) OF 'T' (Cost=9 Card=6361 Bytes=260801)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
81 consistent gets
0 physical reads
0 redo size
690 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
10 rows processed

SQL> SELECT OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME FROM T ORDER BY TIMESTAMP
8 )
9 WHERE ROWNUM <= 20
10 )
11 WHERE RN >= 11;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=64 Card=20 Bytes=1840)
1 0 VIEW (Cost=64 Card=20 Bytes=1840)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=64 Card=6361 Bytes=502519)
4 3 SORT (ORDER BY STOPKEY) (Cost=64 Card=6361 Bytes=260801)
5 4 TABLE ACCESS (FULL) OF 'T' (Cost=9 Card=6361 Bytes=260801)

觀察兩種不同寫法的ORDER BY步驟，一個是帶STOPKEY的ORDER BY，另一個不帶。在大數據量需要排序的情況下，帶STOPKEY的效率要比不帶STOPKEY排序的效率高得多。

SQL> INSERT INTO T SELECT T.* FROM T, USER_OBJECTS;

已創建407104行。

SQL> COMMIT;

提交完成。

SQL> SELECT OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME FROM
5 (
6 SELECT OBJECT_ID, OBJECT_NAME FROM T ORDER BY TIMESTAMP
7 )
8 WHERE ROWNUM <= 20
9 )
10 WHERE RN >= 11;

已選擇10行。

已用時間: 00: 00: 03.78

Statistics
----------------------------------------------------------
268 recursive calls
0 db block gets
6215 consistent gets
6013 physical reads
0 redo size
740 bytes sent via SQL*Net to client
385 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
6 sorts (memory)
0 sorts (disk)
10 rows processed

SQL> SELECT OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME FROM
5 (
6 SELECT OBJECT_ID, OBJECT_NAME FROM T ORDER BY TIMESTAMP
7 )
8 )
9 WHERE RN BETWEEN 11 AND 20;

已選擇10行。

已用時間: 00: 00: 11.86

Statistics
----------------------------------------------------------
26 recursive calls
12 db block gets
6175 consistent gets
9219 physical reads
0 redo size
737 bytes sent via SQL*Net to client
385 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
1 sorts (disk)
10 rows processed

觀察兩個查詢語句的執行時間，以及統計信息中的排序信息。對於第一個查詢語句，Oracle利用了ORDER BY STOPKEY方式進行排序，排序操作只排序需要的TOP N的數據，因此排序操作放到了內存中，而對於第二個查詢語句來說，進行的數據的全排序，排序數據量大，排序操作不得不在磁盤上完成，因此耗時比較多。

通過上面的例子可以看出給出的標準分頁查詢格式，對於包含排序的操作仍然可以在很大程度上提高分頁查詢性能。

前面的各種例子已經說明了分頁查詢語句的標準寫法所帶來的性能提升。

這裏簡單總結一下，並簡單的說明分頁查詢語句在何時無法帶來性能提升。

分頁查詢語句之所以可以很快的返回結果，是因爲它的目標是最快的返回第一條結果。如果每頁有20條記錄，目前翻到第5頁，那麼只需要返回前100條記錄都可以滿足查詢的要求了，也許還有幾萬條記錄也符合查詢的條件，但是由於分頁的限制，在當前的查詢中可以忽略這些數據，而只需儘快的返回前100條數據。這也是爲什麼在標準分頁查詢語句中經常會使用FIRST_ROWS提示的原因。

對於行操作，可以在得到結果的同時將結果直接返回給上一層調用。但是對於結果集操作，Oracle必須得到結果集中所有的數據，因此分頁查詢中所帶的ROWNUM信息不起左右。如果最內層的子查詢中包含了下面這些操作中的一個以上，則分頁查詢語句無法體現出任何的性能優勢：UNION、UNION ALL、MINUS、INTERSECT、GROUP BY、DISTINCT、UNIQUE以及聚集函數如MAX、MIN和分析函數等。

除了這些操作以外，分頁查詢還有一個很明顯的特點，就是處理的頁數越小，效率就越高，越到後面，查詢速度越慢。

分頁查詢用來提高返回速度的方法都是針對數據量較小的前N條記錄而言。無論是索引掃描，NESTED LOOP連接，還是ORDER BY STOPKEY，這些方法帶來性能提升的前提都是數據量比較小，一旦分頁到了最後幾頁，會發現這些方法不但沒有辦法帶來性能的提升，而且性能比普通查詢還要低得多。這一點，在使用分頁查詢的時候，一定要心裏有數。

最後看幾個例子：

首先看看UNION ALL、GROUP BY以及分析函數使外層的ROWNUM限制對內層查詢無效。

SQL> SET AUTOT TRACE
SQL> SELECT /*+ FIRST_ROWS */ OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME FROM T ORDER BY OBJECT_NAME
8 )
9 WHERE ROWNUM <= 20
10 )
11 WHERE RN >= 11;

已選擇10行。

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
23 consistent gets
0 physical reads
0 redo size
597 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10 rows processed

這是分頁查詢ROWNUM起作用的情況，下面看看如果內層查詢包括了集操作時的情況：

SQL> SELECT /*+ FIRST_ROWS */ OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME FROM T
8 UNION ALL
9 SELECT OBJECT_ID, OBJECT_NAME FROM T
10 ORDER BY OBJECT_NAME
11 )
12 WHERE ROWNUM <= 20
13 )
14 WHERE RN >= 11;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: FIRST_ROWS (Cost=85 Card=20 Bytes=1840)
1 0 VIEW (Cost=85 Card=20 Bytes=1840)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=85 Card=12722 Bytes=1005038)
4 3 SORT (ORDER BY STOPKEY) (Cost=18 Card=12722 Bytes=267162)
5 4 UNION-ALL
6 5 TABLE ACCESS (FULL) OF 'T' (Cost=9 Card=6361 Bytes=133581)
7 5 TABLE ACCESS (FULL) OF 'T' (Cost=9 Card=6361 Bytes=133581)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
322 consistent gets
0 physical reads
0 redo size
546 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
10 rows processed

SQL> SELECT /*+ FIRST_ROWS */ OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT /*+ INDEX(T) */ OBJECT_ID, OBJECT_NAME FROM T
8 UNION ALL
9 SELECT /*+ INDEX(T) */ OBJECT_ID, OBJECT_NAME FROM T
10 ORDER BY OBJECT_NAME
11 )
12 WHERE ROWNUM <= 20
13 )
14 WHERE RN >= 11;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: FIRST_ROWS (Cost=1719 Card=20 Bytes=1840)
1 0 VIEW (Cost=1719 Card=20 Bytes=1840)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=1719 Card=12722 Bytes=1005038)
4 3 SORT (ORDER BY STOPKEY) (Cost=1652 Card=12722 Bytes=267162)
5 4 UNION-ALL
6 5 TABLE ACCESS (BY INDEX ROWID) OF 'T' (Cost=826 Card=6361 Bytes=133581)
7 6 INDEX (FULL SCAN) OF 'IND_T_OBJECT_NAME' (NON-UNIQUE) (Cost=26 Card=6361)
8 5 TABLE ACCESS (BY INDEX ROWID) OF 'T' (Cost=826 Card=6361 Bytes=133581)
9 8 INDEX (FULL SCAN) OF 'IND_T_OBJECT_NAME' (NON-UNIQUE) (Cost=26 Card=6361)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
24004 consistent gets
0 physical reads
0 redo size
546 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
10 rows processed

如果說全表掃描的情況下，ROWNUM不起作用效果也不是很明顯的話，通過使用HINT，讓Oracle使用索引掃描，這時ROWNUM不起作用的效果就相當驚人了。

SQL> SELECT /*+ FIRST_ROWS */ OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME FROM T
8 GROUP BY OBJECT_ID, OBJECT_NAME
9 ORDER BY OBJECT_NAME
10 )
11 WHERE ROWNUM <= 20
12 )
13 WHERE RN >= 11;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: FIRST_ROWS (Cost=43 Card=20 Bytes=1840)
1 0 VIEW (Cost=43 Card=20 Bytes=1840)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=43 Card=6361 Bytes=502519)
4 3 SORT (GROUP BY STOPKEY) (Cost=43 Card=6361 Bytes=133581)
5 4 TABLE ACCESS (FULL) OF 'T' (Cost=9 Card=6361 Bytes=133581)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
161 consistent gets
0 physical reads
0 redo size
673 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
10 rows processed

SQL> SELECT /*+ FIRST_ROWS */ OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME, ROW_NUMBER() OVER(ORDER BY OBJECT_NAME)
8 FROM T
9 ORDER BY OBJECT_NAME
10 )
11 WHERE ROWNUM <= 20
12 )
13 WHERE RN >= 11;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: FIRST_ROWS (Cost=826 Card=20 Bytes=1840)
1 0 VIEW (Cost=826 Card=20 Bytes=1840)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=826 Card=6361 Bytes=502519)
4 3 WINDOW (BUFFER) (Cost=826 Card=6361 Bytes=133581)
5 4 TABLE ACCESS (BY INDEX ROWID) OF 'T' (Cost=826 Card=6361 Bytes=133581)
6 5 INDEX (FULL SCAN) OF 'IND_T_OBJECT_NAME' (NON-UNIQUE) (Cost=26 Card=6361)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
12002 consistent gets
0 physical reads
0 redo size
597 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
10 rows processed

上面的例子說明了分頁查詢的優化作用對於哪些種查詢不起作用，下面看看，分頁查詢在翻頁到最後的時候的性能。

SQL> SET AUTOT OFF
SQL> SELECT COUNT(*) FROM T;

COUNT(*)
----------
12722

SQL> SET AUTOT TRACE
SQL> SELECT /*+ FIRST_ROWS */ OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME
8 FROM T
9 ORDER BY OBJECT_NAME
10 )
11 WHERE ROWNUM <= 20
12 )
13 WHERE RN >= 11;

已選擇10行。

SQL> SELECT OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME
8 FROM T
9 ORDER BY OBJECT_NAME
10 )
11 )
12 WHERE RN BETWEEN 11 AND 20;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=43 Card=6361 Bytes=585212)
1 0 VIEW (Cost=43 Card=6361 Bytes=585212)
2 1 COUNT
3 2 VIEW (Cost=43 Card=6361 Bytes=502519)
4 3 SORT (ORDER BY) (Cost=43 Card=6361 Bytes=133581)
5 4 TABLE ACCESS (FULL) OF 'T' (Cost=9 Card=6361 Bytes=133581)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
161 consistent gets
0 physical reads
0 redo size
597 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
10 rows processed

首先看一下，分頁查詢的性能比不使用ROWNUM的情況要強很多，但是，如果將分頁的範圍設置到12710和12720之間，這時候再來對比一下兩種查詢的效率。

SQL> SELECT /*+ FIRST_ROWS */ OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME
8 FROM T
9 ORDER BY OBJECT_NAME
10 )
11 WHERE ROWNUM <= 12720
12 )
13 WHERE RN >= 12711;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: FIRST_ROWS (Cost=826 Card=6361 Bytes=585212)
1 0 VIEW (Cost=826 Card=6361 Bytes=585212)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=826 Card=6361 Bytes=502519)
4 3 TABLE ACCESS (BY INDEX ROWID) OF 'T' (Cost=826 Card=6361 Bytes=133581)
5 4 INDEX (FULL SCAN) OF 'IND_T_OBJECT_NAME' (NON-UNIQUE) (Cost=26 Card=6361)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
12001 consistent gets
0 physical reads
0 redo size
612 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10 rows processed

已選擇10行。

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
161 consistent gets
0 physical reads
0 redo size
612 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
10 rows processed

不難發現，對於第二個查詢，效率和第一次執行完全一樣，但是分頁查詢的效率則大大的下降，而且此時的效率遠遠低於沒有使用ROWNUM的查詢。

這篇文章通過例子說明分頁查詢使用的NESTED LOOP操作，在分頁查詢翻到最後幾頁時的性能問題：

SQL> CREATE TABLE T AS SELECT * FROM DBA_USERS;

表已創建。

SQL> CREATE TABLE T1 AS SELECT * FROM DBA_SOURCE;

表已創建。

SQL> ALTER TABLE T ADD CONSTRAINT PK_T PRIMARY KEY (USERNAME);

表已更改。

SQL> ALTER TABLE T1 ADD CONSTRAINT FK_T1_OWNER FOREIGN KEY (OWNER)
2 REFERENCES T(USERNAME);

表已更改。

SQL> CREATE INDEX IND_T1_OWNER ON T1(NAME);

索引已創建。

SQL> EXEC DBMS_STATS.GATHER_TABLE_STATS(USER, 'T')

PL/SQL 過程已成功完成。

SQL> EXEC DBMS_STATS.GATHER_TABLE_STATS(USER, 'T1')

PL/SQL 過程已成功完成。

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=830 Card=96985 Bytes=5819100)
1 0 VIEW (Cost=830 Card=96985 Bytes=5819100)
2 1 COUNT
3 2 HASH JOIN (Cost=830 Card=96985 Bytes=2909550)
4 3 TABLE ACCESS (FULL) OF 'T' (Cost=2 Card=12 Bytes=132)
5 3 TABLE ACCESS (FULL) OF 'T1' (Cost=826 Card=96985 Bytes=1842715)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
8586 consistent gets
8052 physical reads
0 redo size
574 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10 rows processed

在分頁查詢的前幾頁，NESTED LOOP操作比HASH JOIN操作效率高得多。

SQL> SET AUTOT OFF
SQL> SELECT COUNT(*) FROM T, T1 WHERE USERNAME = OWNER;

COUNT(*)
----------
96985

SQL> SET AUTOT TRACE

已選擇10行。

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
8586 consistent gets
8068 physical reads
0 redo size
571 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10 rows processed

對於最後幾頁，採用HASH JOIN的方式，執行效率幾乎沒有任何改變，而採用NESTED LOOP方式，則效率嚴重下降，而且遠遠低於HASH JOIN的方式。

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: FIRST_ROWS (Cost=97811 Card=96980 Bytes=5818800)
1 0 VIEW (Cost=97811 Card=96980 Bytes=5818800)
2 1 COUNT (STOPKEY)
3 2 NESTED LOOPS (Cost=97811 Card=96985 Bytes=2909550)
4 3 TABLE ACCESS (FULL) OF 'T1' (Cost=826 Card=96985 Bytes=1842715)
5 3 TABLE ACCESS (BY INDEX ROWID) OF 'T' (Cost=1 Card=1 Bytes=11)
6 5 INDEX (UNIQUE SCAN) OF 'PK_T' (UNIQUE)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
105566 consistent gets
8068 physical reads
0 redo size
571 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10 rows processed

分頁查詢一般情況下，很少會翻到最後一篇，如果只是偶爾碰到這種情況，對系統性能不會有很大的影響，但是如果經常碰到這種情況，在設計分頁查詢時應該給予足夠的考慮。

在文章的最後看一下ORDER BY STOPKEY和ORDER BY在翻頁查詢的最後幾頁的性能差異：

SQL> CREATE TABLE T AS SELECT A.* FROM DBA_OBJECTS A, DBA_USERS B, TAB;

表已創建。

SQL> SELECT COUNT(*) FROM T;

COUNT(*)
----------
458064

SQL> EXEC DBMS_STATS.GATHER_TABLE_STATS(USER, 'T')

PL/SQL 過程已成功完成。

SQL> SET AUTOT TRACE
SQL> SET TIMING ON
SQL> SELECT OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME FROM T ORDER BY TIMESTAMP
8 )
9 WHERE ROWNUM <= 20
10 )
11 WHERE RN >= 11;

已選擇10行。

已用時間: 00: 00: 00.03

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=13888 Card=20 Bytes=1840)
1 0 VIEW (Cost=13888 Card=20 Bytes=1840)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=13888 Card=458064 Bytes=36187056)
4 3 SORT (ORDER BY STOPKEY) (Cost=13888 Card=458064 Bytes=18780624)
5 4 TABLE ACCESS (FULL) OF 'T' (Cost=537 Card=458064 Bytes=18780624)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
5579 consistent gets
0 physical reads
0 redo size
694 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
10 rows processed

SQL> SELECT OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME FROM T ORDER BY TIMESTAMP
8 )
9 )
10 WHERE RN BETWEEN 11 AND 20;

已選擇10行。

已用時間: 00: 00: 09.05

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=13888 Card=458064 Bytes=42141888)
1 0 VIEW (Cost=13888 Card=458064 Bytes=42141888)
2 1 COUNT
3 2 VIEW (Cost=13888 Card=458064 Bytes=36187056)
4 3 SORT (ORDER BY) (Cost=13888 Card=458064 Bytes=18780624)
5 4 TABLE ACCESS (FULL) OF 'T' (Cost=537 Card=458064 Bytes=18780624)

Statistics
----------------------------------------------------------
0 recursive calls
41 db block gets
5579 consistent gets
7935 physical reads
0 redo size
689 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
1 sorts (disk)
10 rows processed

對於翻頁查詢的前幾頁，採用ORDER BY STOPKEY的方式比ORDER BY性能上有很大的優勢，那麼對於分頁查詢的最後幾頁，ORDER BY STOPKEY是否和其他分頁查詢技術一樣，性能比普通方式還要低很多：

SQL> SELECT OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME FROM T ORDER BY TIMESTAMP
8 )
9 WHERE ROWNUM <= 458060
10 )
11 WHERE RN >= 458051;

已選擇10行。

已用時間: 00: 00: 09.07

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=13888 Card=458060 Bytes=42141520)
1 0 VIEW (Cost=13888 Card=458060 Bytes=42141520)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=13888 Card=458064 Bytes=36187056)
4 3 SORT (ORDER BY STOPKEY) (Cost=13888 Card=458064 Bytes=18780624)
5 4 TABLE ACCESS (FULL) OF 'T' (Cost=537 Card=458064 Bytes=18780624)

Statistics
----------------------------------------------------------
0 recursive calls
41 db block gets
5579 consistent gets
7933 physical reads
0 redo size
667 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
1 sorts (disk)
10 rows processed

SQL> SELECT OBJECT_ID, OBJECT_NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OBJECT_ID, OBJECT_NAME
5 FROM
6 (
7 SELECT OBJECT_ID, OBJECT_NAME FROM T ORDER BY TIMESTAMP
8 )
9 )
10 WHERE RN BETWEEN 458051 AND 458060;

已選擇10行。

已用時間: 00: 00: 10.01

Statistics
----------------------------------------------------------
0 recursive calls
41 db block gets
5579 consistent gets
7935 physical reads
0 redo size
649 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
1 sorts (disk)
10 rows processed

出乎意料的是，雖然ORDER BY STOPKEY的方式在分頁查詢的最後幾頁性能也有明顯的下降，但是在和普通的ORDER BY相比，無論從邏輯讀、物理讀還是從執行時間上看，二者都屬於一個數量級上的。

看來ORDER BY STOPKEY排序方式，在STOPKEY接近排序總量的時候也不會有明顯的性能下降。

前面提到了對於集操作和聚集操作，使用標準的分頁函數沒有太大的意義，下面通過幾篇文章專門討論集操作和聚集操作的情況。這裏首先討論集操作的情況。

當查詢需要分頁時，大多數情況都會包含排序操作，因爲如果缺少排序操作，很難保證分頁返回的數據是連續不重複的。

因此這裏只考慮包含排序的情況：

SQL> CREATE TABLE T AS SELECT * FROM DBA_TABLES;

表已創建。

SQL> CREATE TABLE T1 AS SELECT * FROM DBA_INDEXES;

表已創建。

SQL> CREATE INDEX IND_T_OWNER ON T(OWNER);

索引已創建。

SQL> CREATE INDEX IND_T1_OWNER ON T1(OWNER);

索引已創建。

SQL> EXEC DBMS_STATS.GATHER_TABLE_STATS(USER, 'T')

PL/SQL 過程已成功完成。

SQL> EXEC DBMS_STATS.GATHER_TABLE_STATS(USER, 'T1')

PL/SQL 過程已成功完成。

SQL> SET AUTOT ON
SQL> SELECT /*+ FIRST_ROWS */ OWNER, NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OWNER, NAME
5 FROM
6 (
7 SELECT OWNER, TABLE_NAME NAME
8 FROM T
9 UNION ALL
10 SELECT OWNER, INDEX_NAME NAME
11 FROM T1
12 ORDER BY OWNER
13 )
14 WHERE ROWNUM <= 20
15 )
16 WHERE RN > 10;

OWNER NAME
------------------------------ ------------------------------
CTXSYS DR$SECTION_GROUP
CTXSYS DR$THS_BT
CTXSYS DR$THS_FPHRASE
CTXSYS DR$THS_PHRASE
CTXSYS DR$THS
CTXSYS DR$SQE
CTXSYS SYS_IOT_OVER_26472
CTXSYS DR$INDEX_OBJECT
CTXSYS DR$POLICY_TAB
CTXSYS DR$INDEX_PARTITION

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: FIRST_ROWS (Cost=17 Card=20 Bytes=940)
1 0 VIEW (Cost=17 Card=20 Bytes=940)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=17 Card=2877 Bytes=97818)
4 3 SORT (ORDER BY STOPKEY) (Cost=7 Card=2877 Bytes=76522)
5 4 UNION-ALL
6 5 TABLE ACCESS (FULL) OF 'T' (Cost=3 Card=1157 Bytes=30082)
7 5 TABLE ACCESS (FULL) OF 'T1' (Cost=4 Card=1720 Bytes=46440)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
44 consistent gets
0 physical reads
0 redo size
639 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
10 rows processed

SQL> SELECT /*+ FIRST_ROWS */ OWNER, NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OWNER, NAME
5 FROM
6 (
7 SELECT OWNER, TABLE_NAME NAME
8 FROM T
9 UNION ALL
10 SELECT OWNER, INDEX_NAME NAME
11 FROM T1
12 ORDER BY OWNER
13 )
14 )
15 WHERE RN > 10 AND RN <=20 ;

OWNER NAME
------------------------------ ------------------------------
CTXSYS DR$THS_BT
CTXSYS DR$THS_FPHRASE
CTXSYS DR$THS_PHRASE
CTXSYS DR$THS
CTXSYS DR$SQE
CTXSYS SYS_IOT_OVER_26472
CTXSYS DR$INDEX_OBJECT
CTXSYS DR$POLICY_TAB
CTXSYS DR$PART_STATS
CTXSYS DR$STATS

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: FIRST_ROWS (Cost=17 Card=2877 Bytes=135219)
1 0 VIEW (Cost=17 Card=2877 Bytes=135219)
2 1 COUNT
3 2 VIEW (Cost=17 Card=2877 Bytes=97818)
4 3 SORT (ORDER BY) (Cost=7 Card=2877 Bytes=76522)
5 4 UNION-ALL
6 5 TABLE ACCESS (FULL) OF 'T' (Cost=3 Card=1157 Bytes=30082)
7 5 TABLE ACCESS (FULL) OF 'T1' (Cost=4 Card=1720 Bytes=46440)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
44 consistent gets
0 physical reads
0 redo size
626 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
10 rows processed

從執行SQL的統計中看到，由於集操作的存在，導致了Oracle沒有辦法將ROWNUM信息推入到查詢內部，導致標準分頁方式的效率和其他分頁方式效率差別不大。

當存在排序操作，且集操作爲UNION ALL操作時，可以改寫SQL語句爲：

SQL> SELECT /*+ FIRST_ROWS */ OWNER, NAME
2 FROM
3 (
4 SELECT ROWNUM RN, OWNER, NAME
5 FROM
6 (
7 SELECT OWNER, NAME
8 FROM
9 (
10 SELECT OWNER, TABLE_NAME NAME
11 FROM T
12 ORDER BY OWNER
13 )
14 WHERE ROWNUM <= 20
15 UNION ALL
16 SELECT *
17 FROM
18 (
19 SELECT OWNER, TABLE_NAME NAME
20 FROM T1
21 ORDER BY OWNER
22 )
23 WHERE ROWNUM <= 20
24 ORDER BY OWNER
25 )
26 WHERE ROWNUM <= 20
27 )
28 WHERE RN > 10;

OWNER NAME
------------------------------ ------------------------------
CTXSYS DR$POLICY_TAB
CTXSYS DR$INDEX_OBJECT
CTXSYS SYS_IOT_OVER_26472
CTXSYS DR$SQE
CTXSYS DR$THS
CTXSYS DR$THS_PHRASE
CTXSYS DR$THS_FPHRASE
CTXSYS DR$THS_BT
CTXSYS DR$SECTION_GROUP
CTXSYS DR$SECTION

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: FIRST_ROWS (Cost=1654 Card=20 Bytes=940)
1 0 VIEW (Cost=1654 Card=20 Bytes=940)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=1654 Card=40 Bytes=1360)
4 3 SORT (ORDER BY STOPKEY) (Cost=1652 Card=40 Bytes=1360)
5 4 UNION-ALL
6 5 COUNT (STOPKEY)
7 6 VIEW (Cost=826 Card=1157 Bytes=39338)
8 7 TABLE ACCESS (BY INDEX ROWID) OF 'T' (Cost=826 Card=1157 Bytes=30082)
9 8 INDEX (FULL SCAN) OF 'IND_T_OWNER' (NON-UNIQUE) (Cost=26 Card=1157)
10 5 COUNT (STOPKEY)
11 10 VIEW (Cost=826 Card=1720 Bytes=58480)
12 11 TABLE ACCESS (BY INDEX ROWID) OF 'T1' (Cost=826 Card=1720 Bytes=39560)
13 12 INDEX (FULL SCAN) OF 'IND_T1_OWNER' (NON-UNIQUE) (Cost=26 Card=1720)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
7 consistent gets
2 physical reads
0 redo size
631 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
10 rows processed

進行了上面的等價改寫，查詢的邏輯讀大大的減少。其實，這裏使用的方法就是人爲的將ROWNUM推入到UNION ALL操作的每個子查詢中，使用這種方法保證查詢結果正確的同時，提高了查詢的效率。

不過上面給出的改寫方法只對包含排序的UNION ALL操作有效。而其他集操作不能使用這種方法，比如UNION操作使用這種方法可能導致查詢結果的數量小於查詢開始限定的數量。而對於另外兩種

本文簡單討論一下包含GROUP BY、DISTINCT、UNIQUE等操作的分頁查詢語句。

由於包含上述的集操作，Oracle必須處理完所有的數據纔會將結果全部的提交給用戶。即使包含了ROWNUM，Oracle也只是在處理完所有的數據之後，將需要的數據進行返回。

不過並不是說前面提到的標準寫法在這裏沒有意義，而是最耗時的部分已經處理完畢。通過ROWNUM來限制意義不大。

雖然標準分頁的寫法對於GROUP BY之類的操作意義不大，但是如果在執行分頁之前需要計算總數的話，那麼可以對分頁的查詢語句稍做修改，將計算總數和查詢第一頁的SQL結合起來，避免一次計算總數的操作。

SQL> ALTER SESSION SET NLS_DATE_FORMAT = 'YYYY-MM-DD HH24:MI:SS';

會話已更改。

SQL> CREATE TABLE T AS SELECT ROWNUM ID, A.* FROM DBA_OBJECTS A;

表已創建。

SQL> ALTER TABLE T ADD PRIMARY KEY (ID);

表已更改。

SQL> EXEC DBMS_STATS.GATHER_TABLE_STATS(USER, 'T')

PL/SQL 過程已成功完成。

SQL> SET AUTOT ON
SQL> SELECT *
2 FROM
3 (
4 SELECT ROWNUM RN, A.*
5 FROM
6 (
7 SELECT OBJECT_TYPE, CREATED, COUNT(*)
8 FROM T
9 GROUP BY OBJECT_TYPE, CREATED
10 ORDER BY CREATED
11 ) A
12 WHERE ROWNUM < 20
13 )
14 WHERE RN >= 10;

RN OBJECT_TYPE CREATED COUNT(*)
---------- ------------------ ------------------- ----------
10 TABLE 2003-11-13 01:41:01 16
11 CLUSTER 2003-11-13 01:41:02 3
12 INDEX 2003-11-13 01:41:02 31
13 LOB 2003-11-13 01:41:02 2
14 SEQUENCE 2003-11-13 01:41:02 4
15 TABLE 2003-11-13 01:41:02 20
16 INDEX 2003-11-13 01:41:03 16
17 LOB 2003-11-13 01:41:03 6
18 SEQUENCE 2003-11-13 01:41:03 2
19 SYNONYM 2003-11-13 01:41:03 1

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=97 Card=19 Bytes=874)
1 0 VIEW (Cost=97 Card=19 Bytes=874)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=97 Card=32185 Bytes=1062105)
4 3 SORT (GROUP BY STOPKEY) (Cost=97 Card=32185 Bytes=547145)
5 4 TABLE ACCESS (FULL) OF 'T' (Cost=24 Card=32185 Bytes=547145)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
232 consistent gets
0 physical reads
0 redo size
759 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
10 rows processed

SQL> SELECT *
2 FROM
3 (
4 SELECT ROWNUM RN, A.*
5 FROM
6 (
7 SELECT OBJECT_TYPE, CREATED, COUNT(*)
8 FROM T
9 GROUP BY OBJECT_TYPE, CREATED
10 ORDER BY CREATED
11 ) A
12 )
13 WHERE RN >= 10 AND RN < 20;

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=97 Card=32185 Bytes=1480510)
1 0 VIEW (Cost=97 Card=32185 Bytes=1480510)
2 1 COUNT
3 2 VIEW (Cost=97 Card=32185 Bytes=1062105)
4 3 SORT (GROUP BY) (Cost=97 Card=32185 Bytes=547145)
5 4 TABLE ACCESS (FULL) OF 'T' (Cost=24 Card=32185 Bytes=547145)

由於查詢操作中GROUP BY操作是耗時最大的SQL，因此標準分頁方式在這裏所帶來的性能提升十分有限。但是，如果在執行分頁查詢前需要執行COUNT(*)的話，那麼可以考慮將COUNT(*)的結果和分頁結果一起返回，從而減少了一次計算總數所需的時間。

SQL> SELECT COUNT(*)
2 FROM
3 (
4 SELECT OBJECT_TYPE, CREATED, COUNT(*)
5 FROM T
6 GROUP BY OBJECT_TYPE, CREATED
7 ORDER BY CREATED
8 );

COUNT(*)
----------
3570

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=97 Card=1)
1 0 SORT (AGGREGATE)
2 1 VIEW (Cost=97 Card=32185)
3 2 SORT (GROUP BY) (Cost=97 Card=32185 Bytes=547145)
4 3 TABLE ACCESS (FULL) OF 'T' (Cost=24 Card=32185 Bytes=547145)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
232 consistent gets
0 physical reads
0 redo size
377 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
1 rows processed

SQL> SELECT *
2 FROM
3 (
4 SELECT ROWNUM RN, A.*
5 FROM
6 (
7 SELECT OBJECT_TYPE, CREATED, COUNT(*)
8 FROM T
9 GROUP BY OBJECT_TYPE, CREATED
10 ORDER BY CREATED
11 ) A
12 WHERE ROWNUM < 20
13 )
14 WHERE RN >= 10;

已選擇10行。

與上面的兩個查詢相比，下面的兩種方法都可以通過一個SQL語句來實現：

SQL> SELECT *
2 FROM
3 (
4 SELECT ROWNUM RN, A.*
5 FROM
6 (
7 SELECT COUNT(*) OVER() CNT, OBJECT_TYPE, CREATED, COUNT(*) CN
8 FROM T
9 GROUP BY OBJECT_TYPE, CREATED
10 ORDER BY CREATED
11 ) A
12 WHERE ROWNUM < 20
13 )
14 WHERE RN >= 10;

RN CNT OBJECT_TYPE CREATED CN
---------- ---------- ------------------ ------------------- ----------
10 3570 TABLE 2003-11-13 01:41:01 16
11 3570 CLUSTER 2003-11-13 01:41:02 3
12 3570 INDEX 2003-11-13 01:41:02 31
13 3570 LOB 2003-11-13 01:41:02 2
14 3570 SEQUENCE 2003-11-13 01:41:02 4
15 3570 TABLE 2003-11-13 01:41:02 20
16 3570 INDEX 2003-11-13 01:41:03 16
17 3570 LOB 2003-11-13 01:41:03 6
18 3570 SEQUENCE 2003-11-13 01:41:03 2
19 3570 SYNONYM 2003-11-13 01:41:03 1

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=97 Card=19 Bytes=1121)
1 0 VIEW (Cost=97 Card=19 Bytes=1121)
2 1 COUNT (STOPKEY)
3 2 VIEW (Cost=97 Card=32185 Bytes=1480510)
4 3 WINDOW (BUFFER) (Cost=97 Card=32185 Bytes=547145)
5 4 SORT (GROUP BY) (Cost=97 Card=32185 Bytes=547145)
6 5 TABLE ACCESS (FULL) OF 'T' (Cost=24 Card=32185 Bytes=547145)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
232 consistent gets
0 physical reads
0 redo size
808 bytes sent via SQL*Net to client
503 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
2 sorts (memory)
0 sorts (disk)
10 rows processed

SQL> SELECT *
2 FROM
3 (
4 SELECT COUNT(*) OVER() CNT, ROWNUM RN, A.*
5 FROM
6 (
7 SELECT OBJECT_TYPE, CREATED, COUNT(*) CN
8 FROM T
9 GROUP BY OBJECT_TYPE, CREATED
10 ORDER BY CREATED
11 ) A
12 )
13 WHERE RN >= 10 AND RN < 20;

CNT RN OBJECT_TYPE CREATED CN
---------- ---------- ------------------ ------------------- ----------
3570 10 TABLE 2003-11-13 01:41:01 16
3570 11 CLUSTER 2003-11-13 01:41:02 3
3570 12 INDEX 2003-11-13 01:41:02 31
3570 13 LOB 2003-11-13 01:41:02 2
3570 14 SEQUENCE 2003-11-13 01:41:02 4
3570 15 TABLE 2003-11-13 01:41:02 20
3570 16 INDEX 2003-11-13 01:41:03 16
3570 17 LOB 2003-11-13 01:41:03 6
3570 18 SEQUENCE 2003-11-13 01:41:03 2
3570 19 SYNONYM 2003-11-13 01:41:03 1

已選擇10行。

Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=97 Card=32185 Bytes=1898915)
1 0 VIEW (Cost=97 Card=32185 Bytes=1898915)
2 1 WINDOW (BUFFER) (Cost=97 Card=32185 Bytes=1062105)
3 2 COUNT
4 3 VIEW (Cost=97 Card=32185 Bytes=1062105)
5 4 SORT (GROUP BY) (Cost=97 Card=32185 Bytes=547145)
6 5 TABLE ACCESS (FULL) OF 'T' (Cost=24 Card=32185 Bytes=547145)

第一種方法採用了標準分頁方式，效率相對更高一些，但是第一種方法需要對原始SQL進行修改，而第二種方式不需要修改原始SQL，直接在原始SQL外面添加一些代碼就可以實現。

集操作，這種方法可能會導致錯誤的查詢結果。

Oracle10g的新功能GROUP BY STOPKEY，使得Oracle10g解決了上一篇文章中提到的GROUP BY操作無法分頁的問題。

在10g以前，Oracle的GROUP BY操作必須完全執行完，才能將結果返回給用戶。但是Oracle10g增加了GROUP BY STOPKEY執行路徑，使得用戶在執行GROUP BY操作時，可以根據STOPKEY隨時中止正在運行的操作。

這使得標準分頁函數對於GROUP BY操作重新發揮了作用。

SQL> CREATE TABLE T AS SELECT ROWNUM ID, A.* FROM DBA_OBJECTS A;

表已創建。

SQL> CREATE INDEX IND_T_CREATED ON T (CREATED);

索引已創建。

SQL> ALTER TABLE T MODIFY CREATED NOT NULL;

表已更改。

SQL> ALTER SESSION SET NLS_DATE_FORMAT = 'YYYY-MM-DD HH24:MI:SS';

會話已更改。

SQL> EXEC DBMS_STATS.GATHER_TABLE_STATS(USER, 'T')

PL/SQL 過程已成功完成。

SQL> SET AUTOT ON
SQL> SET TIMING ON
SQL> SELECT *
2 FROM
3 (
4 SELECT ROWNUM RN, A.*
5 FROM
6 (
7 SELECT CREATED, COUNT(*)
8 FROM T
9 GROUP BY CREATED
10 ) A
11 WHERE ROWNUM < 20
12 )
13 WHERE RN >= 10;

RN CREATED COUNT(*)
---------- ------------------- ----------
10 2005-12-19 17:07:57 50
11 2005-12-19 17:07:58 36
12 2005-12-19 17:08:24 10
13 2005-12-19 17:08:25 49
14 2005-12-19 17:08:26 66
15 2005-12-19 17:08:27 62
16 2005-12-19 17:08:28 81
17 2005-12-19 17:08:29 82
18 2005-12-19 17:08:33 1
19 2005-12-19 17:08:35 3

已選擇10行。

已用時間: 00: 00: 00.04

執行計劃
----------------------------------------------------------
Plan hash value: 3639065582

------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)|
-------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 19 | 665 | 1 (0)|
|* 1 | VIEW | | 19 | 665 | 1 (0)|
|* 2 | COUNT STOPKEY | | | | |
| 3 | VIEW | | 973 | 21406 | 1 (0)|
|* 4 | SORT GROUP BY STOPKEY| | 973 | 7784 | 1 (0)|
| 5 | INDEX FULL SCAN | IND_T_CREATED | 984 | 7872 | 1 (0)|
------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

1 - filter("RN">=10)
2 - filter(ROWNUM<20)
4 - filter(ROWNUM<20)

統計信息
----------------------------------------------------------
1 recursive calls
0 db block gets
67 consistent gets
0 physical reads
0 redo size
730 bytes sent via SQL*Net to client
385 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
10 rows processed

SQL> SELECT *
2 FROM
3 (
4 SELECT ROWNUM RN, A.*
5 FROM
6 (
7 SELECT CREATED, COUNT(*)
8 FROM T
9 GROUP BY CREATED
10 ) A
11 )
12 WHERE RN >= 10
13 AND RN < 20;

RN CREATED COUNT(*)
---------- ------------------- ----------
10 2005-12-19 17:09:27 34
11 2005-12-19 17:09:31 29
12 2005-12-19 17:09:40 29
13 2005-12-19 17:09:58 11
14 2005-12-19 17:10:06 6
15 2005-12-19 17:10:12 48
16 2005-12-19 17:10:20 24
17 2005-12-19 17:10:37 8
18 2005-12-19 17:10:40 2
19 2005-12-19 17:10:49 2

已選擇10行。

已用時間: 00: 00: 00.06

執行計劃
----------------------------------------------------------
Plan hash value: 4036621539

------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)|
-----------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 973 | 34055 | 14 (36)|
|* 1 | VIEW | | 973 | 34055 | 14 (36)|
| 2 | COUNT | | | | |
| 3 | VIEW | | 973 | 21406 | 14 (36)|
| 4 | HASH GROUP BY | | 973 | 7784 | 14 (36)|
| 5 | INDEX FAST FULL SCAN| IND_T_CREATED | 50359 | 393K| 10 (10)|
------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

1 - filter("RN">=10 AND "RN"<20)

統計信息
----------------------------------------------------------
1 recursive calls
0 db block gets
73 consistent gets
0 physical reads
0 redo size
724 bytes sent via SQL*Net to client
385 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
10 rows processed

通過上面的執行計劃可以看到，無論使用標準分頁函數，還是使用其他方式，Oracle採用的都是10g特有的執行計劃GROUP BY STOPKEY和HASH GROUP BY，與9i的GROUP BY相比較，二者的執行效率都很高，不過利用了標準分頁模式，Oracle可以將STOPKEY推入到查詢最內層，使得Oracle在掃描的開始階段就將結果集限制住。從查詢的邏輯讀和執行時間上也可以看出來，GROUP BY STOPKEY的執行效率更高一些。

從這一點上看，Oracle10g的新功能確實使得查詢效率得到提高。

oracle分頁查詢

通過HPA+CronHPA組合應對業務複雜彈性伸縮場景

RSA加密算法個人理解

Calling web service using AJAX

java使用Unicode碼對中文進行排序

Oracle中慎用帶有order by的分頁

table行間隔交換顏色顯示方法

https://yachay.unat.edu.pe/blog/index.php?comment_area=format_blog&comment_component=blog&comment_co

linux以太網驅動總結