首先,什么是SQL的执行计划,包括为什么一个SQL会有多个不同执行计划等类似问题,我这里就不做讨论了,各位可以网上百度
在这里,我主要讨论一下,当一个常用的SQL,执行计划忽然发生改变,我们如何最快速度的将其执行计划稳定为平日使用的合理执行计划。
11G现在比较多,我们先讨论11G
11G,有个新特性,或者说一个新的功能包(package) ,SPM(Sql Plan Management)
有两个初始化参数与其相关
NAME TYPE VALUE------------------------------------ ----------- ------------------------------optimizer_capture_sql_plan_baselines boolean FALSEoptimizer_use_sql_plan_baselines boolean TRUE
我们用来稳定执行计划的,就是SPM中的SQL baseline。
第二个参数很好理解
第一个参数,则是是否开启捕获baseline ,什么意思呢
当一个新SQL 执行,并产生执行计划时,如果 第一个参数为true时,oracle会自动为其产生一个baseline,该baseline对应这个执行计划(
当该SQL因为统计信息,或什么其他东西 而导致执行计划发生改变时,那么会执行一次,并产生一个新的baseline,但是下一次再执行时,依旧会使用第一个baseline(起到稳定执行计划的作用),只有当DBA确认 第二个执行计划确实比第一个好时,DBA可以通过DBMS_SPM包进行调整,使用第二个执行计划产生的basaline,那么以后该sql的执行计划就会使用第二个执行计划。
默认,我们捕获baseline是关闭的,所以一个SQL在执行时,不会有baseline产生。
我们通过从library cache中load一个SQL合理的执行计划 为该SQL的baseline,从而保证其计划稳定
SQL> create table haha(a varchar2(30),b number);Table created.SQL> insert into haha select object_id,object_id from dba_objects;75407 rows created.SQL> create index haha_idx on haha(a);Index created.
SQL>exec dbms_stats.gather_table_stats('SYS','HAHA',NO_INVALIDATE =>FALSE);PL/SQL procedure successfully completed.SQL> select * from haha where a='10000';A B------------------------------ ----------10000 10000SQL> select sql_id,child_number,plan_hash_value,sql_text from v$sql where sql_text like '%haha where%';SQL_ID CHILD_NUMBER PLAN_HASH_VALUE------------- ------------ ---------------SQL_TEXT--------------------------------------------------------------------------------auwza0aq10mk0 0 1624320650select * from haha where a='10000'我们收集了统计信息,并且执行语句select * from haha where a=10000;
通过查v$sql视图获取它的sql_id,child_number,我们现在查看执行计划
QL> select * from table(dbms_xplan.display_cursor('auwza0aq10mk0',0));PLAN_TABLE_OUTPUT------------------------------------------------------------------------------------------------------------------------------------------------------SQL_ID auwza0aq10mk0, child number 0-------------------------------------select * from haha where a='10000'Plan hash value: 1624320650----------------------------------------------------------------------------------------| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |----------------------------------------------------------------------------------------| 0 | SELECT STATEMENT | | | | 2 (100)| || 1 | TABLE ACCESS BY INDEX ROWID| HAHA | 1 | 11 | 2 (0)| 00:00:01 ||* 2 | INDEX RANGE SCAN | HAHA_IDX | 1 | | 1 (0)| 00:00:01 |----------------------------------------------------------------------------------------Predicate Information (identified by operation id):--------------------------------------------------- 2 - access("A"='10000')19 rows selected.这个执行计划使用了索引,执行计划是良好的。
现在我们通过修改统计信息,优化器模式,改变执行计划(产生新的执行计划并使用它,这个执行计划是不良好的,也即是错误的执行计划)。
SQL> alter session set optimizer_mode=first_rows_1;PL/SQL procedure successfully completed.SQL> select * from haha where a='10000';A B------------------------------ ----------10000 10000SQL> exec dbms_stats.set_index_stats(ownname=>user,indname=>'HAHA_IDX',NUMROWS=>1,numdist=>1,no_invalidate=>false);PL/SQL procedure successfully completed.SQL> select * from haha where a='10000';A B------------------------------ ----------10000 10000SQL> select sql_id,child_number,plan_hash_value,sql_text from v$sql where sql_text like '%haha where%';SQL_ID CHILD_NUMBER PLAN_HASH_VALUE------------- ------------ ---------------SQL_TEXT------------------------------------------------------------------------------------------------------------------------------------------------------auwza0aq10mk0 0 1624320650select * from haha where a='10000'auwza0aq10mk0 1 3694936490select * from haha where a='10000'
我们可以看到,这里已经产生了两个不同的执行计划,注意看PLAN_HASH_VALUE列。
我们来查询一下新产生的PLAN
SQL> select * from table(dbms_xplan.display_cursor('auwza0aq10mk0',1));PLAN_TABLE_OUTPUT------------------------------------------------------------------------------------------------------------------------------------------------------SQL_ID auwza0aq10mk0, child number 1-------------------------------------select * from haha where a='10000'Plan hash value: 3694936490--------------------------------------------------------------------------| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |--------------------------------------------------------------------------| 0 | SELECT STATEMENT | | | | 32 (100)| ||* 1 | TABLE ACCESS FULL| HAHA | 1 | 11 | 32 (4)| 00:00:01 |--------------------------------------------------------------------------Predicate Information (identified by operation id):--------------------------------------------------- 1 - filter("A"='10000')18 rows selected.
这个新的PLAN,为全表扫描,是不正确的执行计划
现在我在当前环境下执行该sql,会一直采用这个错误的plan,在业务系统中是很致命的。
SQL> select * from haha where a='10000';A B------------------------------ ----------10000 10000SQL> select * from table(dbms_xplan.display_cursor());PLAN_TABLE_OUTPUT------------------------------------------------------------------------------------------------------------------------------------------------------SQL_ID auwza0aq10mk0, child number 1-------------------------------------select * from haha where a='10000'Plan hash value: 3694936490--------------------------------------------------------------------------| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |--------------------------------------------------------------------------| 0 | SELECT STATEMENT | | | | 32 (100)| ||* 1 | TABLE ACCESS FULL| HAHA | 1 | 11 | 32 (4)| 00:00:01 |--------------------------------------------------------------------------Predicate Information (identified by operation id):--------------------------------------------------- 1 - filter("A"='10000')18 rows selected.
那么我想将正确的执行计划稳定上,该如何做?
通过使用11G的新包,DBMS_SPM就可以实现
SQL BASELINE有两个状态,一个是ENABLED,一个是ACCEPTED,两个都为true,oracle才会使用,如果有同一个SQL的多个SQL_BASELINE存在,且状态都为true,那么会根据成本大小进行判断。(注:当自动捕获开启时,一个SQL,每产生一个不同的执行计划,都会产生一个baseline,enabled都为true,但只有第一个获取的baseline,accepted才为true,其他都为false)
SQL>variable cnt number; SQL> EXECUTE :cnt :=DBMS_SPM.LOAD_PLANS_FROM_CURSOR_CACHE(sql_id=>'auwza0aq10mk0',plan_hash_value=>1624320650);PL/SQL procedure successfully completed.SQL> SELECT sql_handle, sql_text, plan_name, enabled,accepted FROM dba_sql_plan_baselines;SQL_HANDLE SQL_TEXT PLAN_NAME ENA ACC------------------------------ -------------------------------------------------------------------------------- ------------------------------ --- ---SQL_32507237f9e5cfe6 select * from haha where a='10000' SQL_PLAN_34n3k6zwybmz6dd4bd292 YES YES现在,我们再从当前环境下执行该SQL
SQL> select * from haha where a='10000';A B------------------------------ ----------10000 10000SQL> select * from table(dbms_xplan.display_cursor());PLAN_TABLE_OUTPUT------------------------------------------------------------------------------------------------------------------------------------------------------SQL_ID auwza0aq10mk0, child number 3-------------------------------------select * from haha where a='10000'Plan hash value: 1624320650----------------------------------------------------------------------------------------| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |----------------------------------------------------------------------------------------| 0 | SELECT STATEMENT | | | | 14685 (100)| || 1 | TABLE ACCESS BY INDEX ROWID| HAHA | 1 | 11 | 14685 (1)| 00:02:57 ||* 2 | INDEX RANGE SCAN | HAHA_IDX | 1 | | 178 (0)| 00:00:03 |----------------------------------------------------------------------------------------Predicate Information (identified by operation id):--------------------------------------------------- 2 - access("A"='10000')Note----- - SQL plan baseline SQL_PLAN_34n3k6zwybmz6dd4bd292 used for this statement
执行计划已经变为了正常的索引扫描,注意Note,说明有一个SQL_BASELINE在使用。
如果我们不想使用该SQL_BASELINE了,该如何呢?
可以有两个方法实现
第一、将该BASELINE的属性,Enabled,Accepted 改为false(Accepted在11r2里面改不了,只能改enabled)
第二、将该BASELINE删除(SQL_SPM.DROP_SQL_PLAN_BASELINE)
我们来一一演示
第一、修改属性
SQL> select SQL_HANDLE,SQL_TEXT,PLAN_NAME,ENABLED,ACCEPTED from dba_sql_plan_baselines;SQL_HANDLE SQL_TEXT PLAN_NAME ENA ACC------------------------------ ----------------------------------- ------------------------------ --- ---SQL_32507237f9e5cfe6 select * from haha where a='10000' SQL_PLAN_34n3k6zwybmz6dd4bd292 YES YESSQL>variable cnt number; SQL>EXECUTE :cnt :=DBMS_SPM.ALTER_SQL_PLAN_BASELINE(sql_handle =>'SQL_32507237f9e5cfe6' ,plan_name =>'SQL_PLAN_34n3k6zwybmz6dd4bd292',attribute_name=>'enabled',attribute_value =>'NO');PL/SQL procedure successfully completed.SQL> select SQL_HANDLE,SQL_TEXT,PLAN_NAME,ENABLED,ACCEPTED from dba_sql_plan_baselines;SQL_HANDLE SQL_TEXT PLAN_NAME ENA ACC------------------------------ ----------------------------------- ------------------------------ --- ---SQL_32507237f9e5cfe6 select * from haha where a='10000' SQL_PLAN_34n3k6zwybmz6dd4bd292 NO YES
看,我们执行完DBMS_SPM.ALTER_SQL_PLAN_BASELINE之后,通过查询dba_sql_plan_baselines可以看到 Enabled属性已经变成了NO
现在我们执行select * from haha where a='10000';
SQL> select * from haha where a='10000';A B------------------------------ ----------10000 10000SQL> select * from table(dbms_xplan.display_cursor());PLAN_TABLE_OUTPUT------------------------------------------------------------------------------------------------------------------------------------------------------SQL_ID auwza0aq10mk0, child number 0-------------------------------------select * from haha where a='10000'Plan hash value: 3694936490--------------------------------------------------------------------------| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |--------------------------------------------------------------------------| 0 | SELECT STATEMENT | | | | 32 (100)| ||* 1 | TABLE ACCESS FULL| HAHA | 1 | 11 | 32 (4)| 00:00:01 |PLAN_TABLE_OUTPUT--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------Predicate Information (identified by operation id):--------------------------------------------------- 1 - filter("A"='10000')18 rows selected.
看,这里的执行计划走成了全表扫描,而且note也没有了。(baseline不起作用了)
第二、删除sql plan baseline
我们除了修改属性以外,还有一种方法,那就是删除sql plan baseline
现在我们将属性改回(改回不是必须的,不过为了模拟场景而已),并删除该baseline
SQL>variable cnt number; SQL>EXECUTE :cnt :=DBMS_SPM.ALTER_SQL_PLAN_BASELINE(sql_handle =>'SQL_32507237f9e5cfe6' ,plan_name =>'SQL_PLAN_34n3k6zwybmz6dd4bd292',attribute_name=>'enabled',attribute_value =>'YES');
好,现在我们查看,并删除(SQL_SPM.DROP_SQL_PLAN_BASELINE 该函数很简单,只有两个元素(输入值))
SQL> select SQL_HANDLE,SQL_TEXT,PLAN_NAME,ENABLED,ACCEPTED from dba_sql_plan_baselines;SQL_HANDLE SQL_TEXT PLAN_NAME ENA ACC------------------------------ ----------------------------------- ------------------------------ --- ---SQL_32507237f9e5cfe6 select * from haha where a='10000' SQL_PLAN_34n3k6zwybmz6dd4bd292 YES YESSQL>variable cnt number; SQL>EXECUTE :cnt :=DBMS_SPM.DROP_SQL_PLAN_BASELINE (sql_handle =>'SQL_32507237f9e5cfe6' ,plan_name =>'SQL_PLAN_34n3k6zwybmz6dd4bd292');PL/SQL procedure successfully completed.SQL> select SQL_HANDLE,SQL_TEXT,PLAN_NAME,ENABLED,ACCEPTED from dba_sql_plan_baselines;no rows selectedSQL> select * from haha where a='10000';A B------------------------------ ----------10000 10000SQL> select * from table(dbms_xplan.display_cursor());PLAN_TABLE_OUTPUT------------------------------------------------------------------------------------------------------------------------------------------------------SQL_ID auwza0aq10mk0, child number 0-------------------------------------select * from haha where a='10000'Plan hash value: 3694936490--------------------------------------------------------------------------| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |--------------------------------------------------------------------------| 0 | SELECT STATEMENT | | | | 32 (100)| ||* 1 | TABLE ACCESS FULL| HAHA | 1 | 11 | 32 (4)| 00:00:01 |PLAN_TABLE_OUTPUT--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------Predicate Information (identified by operation id):--------------------------------------------------- 1 - filter("A"='10000')18 rows selected.
我们看到执行计划中note也消失了,且走为全表扫描。
to be continue--