利用虚拟索引（Virtual Index）优化数据库的案例分析

当我们在对生产库做优化的时候，主要就是对SQL语句的优化，包括语句的等价改写等，但其中很大一部分情况，又与索引有关。如果能合理利用合适的索引，可以使原本走全表扫描产生的逻辑读大大降低，提高数据库的性能。由于Oracle数据库中的索引本身就要占用磁盘空间，维护索引需要一定的开销，如何才能知道创建某个索引，会给数据带来性能的提升，而又不至于判断失误，创建了一个不恰当的索引，最后又不得不删除呢？这种情况下，我们可以利用Oralce提供的虚拟索引，即nosegment索引，它并不占用磁盘资源，只是在数据字典中增加一个定义。它为DBA在创建索引对提升数据库性能的方面提供了一定的参考。下面来看具体测试和分析：

SQL> startup
ORACLE instance started.

Total System Global Area 835104768 bytes
Fixed Size 2232960 bytes
Variable Size 675286400 bytes
Database Buffers 155189248 bytes
Redo Buffers 2396160 bytes
Database mounted.
Database opened.

--本测试在11.2.0.3.0环境，与10g略有不同
SQL> select * from v$version;

BANNER
--------------------------------------------------------------------------------
Oracle Database 11g Enterprise Edition Release 11.2.0.3.0 - 64bit Production
PL/SQL Release 11.2.0.3.0 - Production
CORE 11.2.0.3.0 Production
TNS for Linux: Version 11.2.0.3.0 - Production
NLSRTL Version 11.2.0.3.0 - Production

--创建测试表fakeind并插入数据

SQL> drop table fakeind_test;
drop table fakeind_test
*
ERROR at line 1:
ORA-00942: table or view does not exist

SQL> create table fakeind_test as select * from dba_objects;

Table created.

SQL> insert into fakeind_test select * from fakeind_test;

75540 rows created.

SQL> /

151080 rows created.

SQL> /

302160 rows created.

SQL> select count(*) from fakeind_test;

COUNT(*)
----------
604320

--开始测试，执行查询
SQL> set line 130 pages 130

SQL> select object_id,object_name from fakeind_test where object_id in(select distinct object_id from fakeind_test where object_id>44500 and object_id<45000);

3992 rows selected.

Execution Plan
----------------------------------------------------------
Plan hash value: 1190425891

-------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 3816 | 160K| 4667 (1)| 00:00:57 |
|* 1 | HASH JOIN RIGHT SEMI| | 3816 | 160K| 4667 (1)| 00:00:57 |
| 2 | VIEW | VW_NSO_1 | 3819 | 49647 | 2333 (1)| 00:00:28 |
|* 3 | TABLE ACCESS FULL | FAKEIND_TEST | 3819 | 19095 | 2333 (1)| 00:00:28 |
| 4 | TABLE ACCESS FULL | FAKEIND_TEST | 604K| 17M| 2331 (1)| 00:00:28 |
-------------------------------------------------------------------------------------

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

1 - access("OBJECT_ID"="OBJECT_ID")
3 - filter("OBJECT_ID">44500 AND "OBJECT_ID"<45000)

Statistics
----------------------------------------------------------
23 recursive calls
0 db block gets
17436 consistent gets
0 physical reads
0 redo size
144488 bytes sent via SQL*Net to client
3445 bytes received via SQL*Net from client
268 SQL*Net roundtrips to/from client
3 sorts (memory)
0 sorts (disk)
3992 rows processed

可以看到，用CTAS创建的测试表fakeind上目前并没有索引，因此在生成的执行计划中，该条SQL语句只能走全表扫描

--创建虚拟索引(在普通创建索引命令后加一个nosegmnet即可)
SQL> create index ind_fake_id on fakeind_test(object_id) nosegment;

Index created.

--设置隐含参数使虚拟索引生效
SQL> alter session set "_use_nosegment_indexes"=true; --注意必须要写双引号，单引号不行

Session altered.

SQL> set autot off

--查看表是否被分析过
SQL> select table_name,last_analyzed from dba_tables where table_name='FAKEIND_TEST';

TABLE_NAME LAST_ANALYZED
------------------------------ ------------------
FAKEIND_TEST

--收集测试表的统计信息

SQL> exec dbms_stats.gather_table_stats(ownname=>'SYS',tabname=>'FAKEIND_TEST',degree=>4,estimate_percent=>100,cascade=>true);

PL/SQL procedure successfully completed.

--再次确认表的分析情况
SQL> select table_name,last_analyzed from dba_tables where table_name='FAKEIND_TEST';

TABLE_NAME LAST_ANALYZED
------------------------------ ------------------
FAKEIND_TEST 17-SEP-14

--再次查询测试表

SQL> set autot trace

SQL> select object_id,object_name from fakeind_test where object_id in(select distinct object_id from fakeind_test where object_id>44500 and object_id<45000);

3992 rows selected.

Execution Plan
----------------------------------------------------------
Plan hash value: 2531911586

-----------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-----------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 3904 | 308K| 12 (17)| 00:00:01 |
| 1 | VIEW | VM_NWVW_2 | 3904 | 308K| 12 (17)| 00:00:01 |
| 2 | HASH UNIQUE | | 3904 | 179K| 12 (17)| 00:00:01 |
|* 3 | HASH JOIN | | 3904 | 179K| 11 (10)| 00:00:01 |
|* 4 | INDEX RANGE SCAN | IND_FAKE_ID | 3819 | 19095 | 2 (0)| 00:00:01 |
| 5 | TABLE ACCESS BY INDEX ROWID| FAKEIND_TEST | 3819 | 156K| 8 (0)| 00:00:01 |
|* 6 | INDEX RANGE SCAN | IND_FAKE_ID | 3819 | | 2 (0)| 00:00:01 |
-----------------------------------------------------------------------------------------------

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

3 - access("OBJECT_ID"="OBJECT_ID")
4 - access("OBJECT_ID">44500 AND "OBJECT_ID"<45000)
6 - access("OBJECT_ID">44500 AND "OBJECT_ID"<45000)

Statistics
----------------------------------------------------------
1 recursive calls
0 db block gets
17418 consistent gets
0 physical reads
0 redo size
144488 bytes sent via SQL*Net to client
3445 bytes received via SQL*Net from client
268 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
3992 rows processed

SQL> set autot off

此时利用虚拟索引获得的执行计划中，COST从之前的4000多降低到12，执行时间也从57s到1s，由此可以判断，当加上真实索引后，性能会大大提高。

--创建真实索引

SQL> create index ind_real_id on fakeind_test(object_id);

Index created.

SQL> set autot trace
SQL> select object_id,object_name from fakeind_test where object_id in(select distinct object_id from fakeind_test where object_id>45500 and object_id<50000);

35992 rows selected.

Execution Plan
----------------------------------------------------------
Plan hash value: 2531911586

-------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes |TempSpc| Cost (%CPU)| Time |
-------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 41816 | 3307K| | 548 (1)| 00:00:07 |
| 1 | VIEW | VM_NWVW_2 | 41816 | 3307K| | 548 (1)| 00:00:07 |
| 2 | HASH UNIQUE | | 41816 | 1919K| 2472K| 548 (1)| 00:00:07 |
|* 3 | HASH JOIN | | 41816 | 1919K| | 53 (2)| 00:00:01 |
|* 4 | INDEX RANGE SCAN | IND_FAKE_ID | 34375 | 167K| | 3 (0)| 00:00:01 |
| 5 | TABLE ACCESS BY INDEX ROWID| FAKEIND_TEST | 34375 | 1409K| | 49 (0)| 00:00:01 |
|* 6 | INDEX RANGE SCAN | IND_FAKE_ID | 34375 | | | 3 (0)| 00:00:01 |
-------------------------------------------------------------------------------------------------------

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

3 - access("OBJECT_ID"="OBJECT_ID")
4 - access("OBJECT_ID">45500 AND "OBJECT_ID"<50000)
6 - access("OBJECT_ID">45500 AND "OBJECT_ID"<50000)

Statistics
----------------------------------------------------------
1 recursive calls
0 db block gets
11017 consistent gets
82 physical reads
0 redo size
1293055 bytes sent via SQL*Net to client
26908 bytes received via SQL*Net from client
2401 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
35992 rows processed

虽然创建了真实索引，但数据库却仍旧在用虚拟索引，此时COST和TIME反而还上去了一点，那么需要先禁用虚拟索引

SQL> alter session set "_use_segment_indexes"=false;

--禁用虚拟索引后继续查看刚才的SQL

SQL> select object_id,object_name from fakeind_test where object_id in(select distinct object_id from fakeind_test where object_id>45500 and object_id<50000);

35992 rows selected.

Execution Plan
----------------------------------------------------------
Plan hash value: 750753197

-------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 34375 | 1443K| 2414 (1)| 00:00:29 |
|* 1 | HASH JOIN RIGHT SEMI| | 34375 | 1443K| 2414 (1)| 00:00:29 |
| 2 | VIEW | VW_NSO_1 | 34375 | 436K| 79 (0)| 00:00:01 |
|* 3 | INDEX RANGE SCAN | IND_REAL_ID | 34375 | 167K| 79 (0)| 00:00:01 |
| 4 | TABLE ACCESS FULL | FAKEIND_TEST | 604K| 17M| 2331 (1)| 00:00:28 |
-------------------------------------------------------------------------------------

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

1 - access("OBJECT_ID"="OBJECT_ID")
3 - access("OBJECT_ID">45500 AND "OBJECT_ID"<50000)

Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
11017 consistent gets
0 physical reads
0 redo size
1293055 bytes sent via SQL*Net to client
26908 bytes received via SQL*Net from client
2401 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
35992 rows processed

虽然使用真实索引之后，性能提升并不如使用虚拟索引时那样多，但至少比最初没有索引的情况下，要快了将近28秒，COST也减少了将近一半，当真实索引建立完毕以后，可以对虚拟索引进行删除，以免白白占用一个对象名，删除语法和删除普通索引一致。

虚拟索引有几个要注意的地方：

--虚拟索引并不存在于dba_indexes视图

SQL> select index_name from dba_indexes where index_name='IND_FAKE_ID';

no rows selected

--无法创建与虚拟索引同名的真实索引
SQL> create index ind_fake_id on fakeind_test(object_name);
create index ind_fake_id on fakeind_test(object_name)
*
ERROR at line 1:
ORA-00955: name is already used by an existing object

--无法使用alter命令来修改或重建索引
SQL> alter index ind_fake_id rename to ind_fake_name;
alter index ind_fake_id rename to ind_fake_name
*
ERROR at line 1:
ORA-08114: can not alter a fake index

SQL> alter index ind_fake_id rebuild;
alter index ind_fake_id rebuild
*
ERROR at line 1:
ORA-08114: can not alter a fake index

--查看虚拟索引的方法
SQL> set autot off
SQL> SELECT index_owner, index_name
2 FROM dba_ind_columns
3 WHERE index_name NOT LIKE 'BIN$%'
4 MINUS
5 SELECT owner, index_name
6 FROM dba_indexes;

INDEX_OWNER INDEX_NAME
------------------------------ ------------------------------
SYS IND_FAKE_ID

--drop table并不会直接删除虚拟索引，必须purge recyclebin或者显示drop index

SQL> conn zlm/oracle --由于SYS用户默认没有recyclebin，这里用zlm用户做测试

Connected.

SQL> create table t1 as select * from dba_objects;

Table created.

SQL> create index ind_t1_id on t1(object_id) nosegment;

Index created.

SQL> drop table t1;

Table dropped.

SQL> SELECT index_owner, index_name
2 FROM dba_ind_columns
3 WHERE index_name NOT LIKE 'BIN$%'
4 MINUS
5 SELECT owner, index_name
6 FROM dba_indexes;

INDEX_OWNER INDEX_NAME
------------------------------ ------------------------------
ZLM IND_T1_ID --尽管已经删除了T1表，但是虚拟索引仍然还在

SQL> purge recyclebin;

Recyclebin purged.

SQL> SELECT index_owner, index_name
2 FROM dba_ind_columns
3 WHERE index_name NOT LIKE 'BIN$%'
4 MINUS
5 SELECT owner, index_name
6 FROM dba_indexes;

no rows selected

--清空回收站后，虚拟索引才被彻底删除了，或者直接显示删除虚拟索引
SQL> drop index ind_t1_id;

总结：

这个测试其实并没有做的很完善，SQL语句选取得不好，正常情况下，原有语句所涉及的表至少且肯定会有一个主键索引，没有索引的表在OLTP生产库中是不太现实的。本文主要是起到一个抛砖引玉的作用，当我们面对一个数据库优化的场景，需要测试创建某个特定条件的索引是否会给系统带来性能提升，就可以借助虚拟索引来测试，因为在生产库中通常不允许随意创建索引，而维护索引对数据库而言也是一笔不小的开销，况且如果索引创建不当，可能使数据库性能更糟糕。