【案例5】你是否仍迷信rowid分页？
一直以来，看到的说法大都是rowid要强于rownum分页：在索引设计的没问题的前提下，两种方式在前几页性能上接近，越向后翻rowid的效果越明显。
注意红字部分，红字部分一定成立吗？实验说明一切。
我们先来看看11g的情况，建立实验环境：
11gR2 >create table test(id number,status VARCHAR2(7),type VARCHAR2(19),created date);
Table created.
11gR2 >insert into test select OBJECT_ID,STATUS,OBJECT_TYPE,CREATED from dba_objects;
12926 rows created.
11gR2 >alter table test modify created not null;
Table altered.
11gR2 >create index test_ind1 on test(CREATED);
Index created.
11gR2 >ANALYZE TABLE TEST compute statistics;
Table analyzed.
测试11g的rownum分页：
11gR2 >select *
from (
select rownum rn,t.*
from
(selectid,status,type,created from test order by created) t
where rownum<1000)
where rn >900;
99 rows selected.
--------------------------------------------------------------------------------------------
| Id| Operation | Name | Rows| Bytes | Cost (%CPU)| Time |
--------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 999 | 50949 | 9 (0)| 00:00:01 |
|*1 |VIEW | | 999 | 50949 | 9 (0)| 00:00:01 |
|*2 | COUNT STOPKEY | | | | | |
| 3 | VIEW | | 999 | 37962 | 9 (0)| 00:00:01 |
| 4 | TABLE ACCESS BY INDEX ROWID| TEST | 12926 | 277K| 9 (0)| 00:00:01 |
| 5 | INDEX FULL SCAN | TEST_IND1 | 999 | | 4 (0)| 00:00:01 |
--------------------------------------------------------------------------------------------
Statistics
----------------------------------------------------------
0recursive calls
0db block gets
22consistent gets
0physical reads
0redo size
3703bytes sent via SQL*Net to client
590bytes received via SQL*Net from client
8SQL*Net roundtrips to/from client
0sorts (memory)
0sorts (disk)
99rows processed
下面测试rowid分页
11gR2 >select /*+ ordered use_nl(p s) */ *
from (
select rownum rn,rd
from (selectrowid rd from test order by created)
t where rownum<1000) p,
test s
where rn>900 and p.rd=s.rowid;2 3 4 5 6 7
99 rows selected.
-----------------------------------------------------------------------------------------
| Id| Operation | Name | Rows| Bytes | Cost (%CPU)| Time |
-----------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 999 | 52947 |1003 (0)| 00:00:13 |
| 1 |NESTED LOOPS | | 999 | 52947 |1003 (0)| 00:00:13 |
|*2 | VIEW | | 999 | 24975 | 4 (0)| 00:00:01 |
|*3 | COUNT STOPKEY | | | | | |
| 4 | VIEW | | 999 | 11988 | 4 (0)| 00:00:01 |
| 5 | INDEX FULL SCAN | TEST_IND1 | 12926 | 239K| 4 (0)| 00:00:01 |
| 6 | TABLE ACCESS BY USER ROWID| TEST | 1 | 28 | 1 (0)| 00:00:01 |
-----------------------------------------------------------------------------------------
Statistics
----------------------------------------------------------
0recursive calls
0db block gets
19consistent gets
0physical reads
0redo size
5450bytes sent via SQL*Net to client
590bytes received via SQL*Net from client
8SQL*Net roundtrips to/from client
0sorts (memory)
0sorts (disk)
99rows processed
我们看可以看到，当读900-1000行时，rowid分页（19逻辑读）确实是优于rownum分页（22逻辑读）。因为rowid方式减少了前900行的回表。
我们来看看10g的情况。
在10g建立同样的TEST表和索引，然后：
测试10g的rownum分页：
10gR2 >select *
from (
select rownum rn,t.*
from
(selectid,status,type,created from test order by created) t
where rownum<1000)
where rn >900;
99 rows selected.
--------------------------------------------------------------------------------------------
| Id| Operation | Name | Rows| Bytes | Cost (%CPU)| Time |
--------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 999 | 50949 | 17 (0)| 00:00:01 |
|*1 |VIEW | | 999 | 50949 | 17 (0)| 00:00:01 |
|*2 | COUNT STOPKEY | | | | | |
| 3 | VIEW | | 999 | 37962 | 17 (0)| 00:00:01 |
| 4 | TABLE ACCESS BY INDEX ROWID| TEST | 45620 |1113K| 17 (0)| 00:00:01 |
| 5 | INDEX FULL SCAN | TEST_IND1 | 999 | | 3 (0)| 00:00:01 |
--------------------------------------------------------------------------------------------
Statistics
----------------------------------------------------------
0recursive calls
0db block gets
19consistent gets
0physical reads
0redo size
3842bytes sent via SQL*Net to client
558bytes received via SQL*Net from client
8SQL*Net roundtrips to/from client
0sorts (memory)
0sorts (disk)
99rows processed
19个逻辑读，和11g是很接近的，是个正常值，也符合我们的预期。
继续测试10g的rowid分页：
swp1 >select /*+ ordered use_nl(p s) */ *
from (
select rownum rn,rd
from (selectrowid rd from test order by created)
t where rownum<1000) p,
test s
where rn>900 and p.rd=s.rowid;
2 3 4 5 6 7
99 rows selected.
-----------------------------------------------------------------------------------------
| Id| Operation | Name | Rows| Bytes | Cost (%CPU)| Time |
-----------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 999 | 55944 |1061 (1)| 00:00:15 |
| 1 |NESTED LOOPS | | 999 | 55944 |1061 (1)| 00:00:15 |
|*2 | VIEW | | 999 | 24975 | 62 (2)| 00:00:01 |
|*3 | COUNT STOPKEY | | | | | |
| 4 | VIEW | | 45620 | 534K| 62 (2)| 00:00:01 |
| 5 | INDEX FULL SCAN | TEST_IND1 | 45620 | 846K| 62 (2)| 00:00:01 |
| 6 | TABLE ACCESS BY USER ROWID| TEST | 1 | 31 | 1 (0)| 00:00:01 |
-----------------------------------------------------------------------------------------
Statistics
----------------------------------------------------------
0recursive calls
0db block gets
109consistent gets
1physical reads
0redo size
5585bytes sent via SQL*Net to client
558bytes received via SQL*Net from client
8SQL*Net roundtrips to/from client
0sorts (memory)
0sorts (disk)
99rows processed
怎么逻辑读变成了109？ 执行计划是和11g一样的。为什么会有这么大的差别？
rowid分页的cost一下变成了rownum方式的好几倍。
为什么？
下面的实验揭露了原因，也为我们揭露了11g一个不为人知的新特性：
11gR2 >select rowid,id from test where rowid in ('AAADSHAABAAAH3hAAA','AAADSHAABAAAH3hAAB','AAADSHAABAAAH3hAAC','AAADSHAABAAAH3hAAD','AAADSHAABAAAH3hAAE','AAADSHAABAAAH3hAAF','AAADSHAABAAAH3hAAG','AAADSHAABAAAH3hAAH','AAADSHAABAAAH3hAAI','AAADSHAABAAAH3hAAJ');
ROWID ID
------------------ ----------
AAADSHAABAAAH3hAAA 20
AAADSHAABAAAH3hAAB 46
AAADSHAABAAAH3hAAC 28
AAADSHAABAAAH3hAAD 15
AAADSHAABAAAH3hAAE 29
AAADSHAABAAAH3hAAF 3
AAADSHAABAAAH3hAAG 25
AAADSHAABAAAH3hAAH 41
AAADSHAABAAAH3hAAI 54
AAADSHAABAAAH3hAAJ 40
10 rows selected.
------------------------------------------------------------------------------------
| Id| Operation | Name | Rows| Bytes | Cost (%CPU)| Time |
------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 16 | 1 (0)| 00:00:01 |
| 1 |INLIST ITERATOR | | | | | |
| 2 | TABLE ACCESS BY USER ROWID| TEST | 1 | 16 | 1 (0)| 00:00:01 |
------------------------------------------------------------------------------------
Statistics
----------------------------------------------------------
0recursive calls
0db block gets
2consistent gets
0physical reads
0redo size
875bytes sent via SQL*Net to client
524bytes received via SQL*Net from client
2SQL*Net roundtrips to/from client
0sorts (memory)
0sorts (disk)
10rows processed
10gR2 >select rowid,id from test where rowid in ('AAAzo5AABAAAKnWAAA','AAAzo5AABAAAKnWAAB','AAAzo5AABAAAKnWAAC','AAAzo5AABAAAKnWAAD','AAAzo5AABAAAKnWAAE','AAAzo5AABAAAKnWAAF','AAAzo5AABAAAKnWAAG','AAAzo5AABAAAKnWAAH','AAAzo5AABAAAKnWAAI','AAAzo5AABAAAKnWAAJ');
ROWID ID
------------------ ----------
AAAzo5AABAAAKnWAAA 30
AAAzo5AABAAAKnWAAB 8
AAAzo5AABAAAKnWAAC 14
AAAzo5AABAAAKnWAAD 34
AAAzo5AABAAAKnWAAE 45
AAAzo5AABAAAKnWAAF 39
AAAzo5AABAAAKnWAAG 47
AAAzo5AABAAAKnWAAH 51
AAAzo5AABAAAKnWAAI 11
AAAzo5AABAAAKnWAAJ 48
10 rows selected.
------------------------------------------------------------------------------------
| Id| Operation | Name | Rows| Bytes | Cost (%CPU)| Time |
------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 16 | 1 (0)| 00:00:01 |
| 1 |INLIST ITERATOR | | | | | |
| 2 | TABLE ACCESS BY USER ROWID| TEST | 1 | 16 | 1 (0)| 00:00:01 |
------------------------------------------------------------------------------------
Statistics
----------------------------------------------------------
1recursive calls
0db block gets
10consistent gets
0physical reads
0redo size
861bytes sent via SQL*Net to client
492bytes received via SQL*Net from client
2SQL*Net roundtrips to/from client
0sorts (memory)
0sorts (disk)
10rows processed
我们看到，同样用rowid读取同一个block的10行，在11g中仅仅耗费了2个读，而在10g中耗费10个读。
更进一步的测试也确认了10g中，每一个rowid都会产生一个逻辑读，即使这些rowid来自一个block，也不能重用。
而11g引入了对rowid读的新特性， 那就是，如果rowid是来自一个block，那么是可以重用的。
总结：
在11g中，得益于这个11g不为人知的新特性，总体来说，rowid分页要优于rownum。
可是在10g中，相当一部分情况下rowid方式性能是远远不如rownum方式的。只有在后翻的页数非常大，并且页的大小很小，譬如10000-10010的情况下，rowid的分页才会优于rownum方式。
老师的实验讲的很细，致敬！:)
页:
[1]