执行计划变化的处理

oracle 10g

4、操做步骤

方法一：

    先尝试用 oracle自带的调优包获得建议，若是建议里能够采用profile获得原来的执行计划，则接受建议，执行相关命令；若是建议里没有原来的执行计划，转到方法二。

注意：将下面 SQL中的“8sjqpgjwztzm8”替换为问题SQL的sql_id便可运行。

步骤1、建立任务

set serverout on

DECLARE

 l_sql_tune_task_id VARCHAR2(100);

BEGIN

 l_sql_tune_task_id := DBMS_SQLTUNE.create_tuning_task(sql_id       => '8sjqpgjwztzm8',

                                                        scope       => DBMS_SQLTUNE.scope_comprehensive,

                                                        time_limit => 600,

                                                        task_name   => '8sjqpgjwztzm8_tuning_task',

                                                        description => 'Tuning task for statement 8sjqpgjwztzm8.');

 DBMS_OUTPUT.put_line('l_sql_tune_task_id: ' || l_sql_tune_task_id);

END;

步骤2、确认任务已经建立

--有返回结果表示正常

select task_name

 from dba_advisor_log

 where owner = 'SYS'

   and task_name = '8sjqpgjwztzm8_tuning_task';

运行任务

Execute dbms_sqltune.Execute_tuning_task (task_name => '8sjqpgjwztzm8_tuning_task');

步骤3、确认任务状态

select status

 from dba_advisor_log

 where task_name = '8sjqpgjwztzm8_tuning_task';

步骤4、打印报表

--查看调优建议，这一步能够获得调优建议，若是有须要的执行计划，则接受 profile

set long 999999

set longchunksize 1000

set lin 140 pages 1000

select dbms_sqltune.report_tuning_task('8sjqpgjwztzm8_tuning_task')

 from dual;

 

这里举个例子，若是确认执行计划可用，运行红框命令便可。

步骤5、删除任务

--调优完成后此任务再也不须要

exec dbms_sqltune.DROP_TUNING_TASK(task_name => '8sjqpgjwztzm8_tuning_task');

 

方法二：

因统计信息变化致使执行计划效率变差的状况，出现的频率仍是很高的，但愿轮班的同事，至少是值班班长有能力去处理这类的问题。

 

基本的思路以下，你们须要不断在实战中锻炼本身，作几回就熟了，不要怯战。

步骤1、查看历史信息

--经过看 awr的历史信息，很容易就知道执行计划的效率更高

select s.instance_number,

       to_char(sn.end_interval_time, 'YYYYMMDD HH24:MI:SS'),

       s.plan_hash_value,

       s.executions_delta,

       round(s.elapsed_time_delta / s.executions_delta),

       round(s.BUFFER_GETS_delta / s.executions_delta),

       round(s.CPU_TIME_delta / s.executions_delta)

 from dba_hist_snapshot sn, sys.WRH$_SQLSTAT s

 where s.snap_id = sn.snap_id

   and s.sql_id = '1gu8t96d0bdmu'

   and s.instance_number = sn.instance_number

   and s.executions_delta > 0

 order by sn.end_interval_time desc

步骤2、对比新旧执行计划区别

--对比效率高和低的执行计划的区别

SELECT * FROM table(DBMS_XPLAN.DISPLAY_AWR('&sql_id',null,null,'ALL'));

 

步骤3、作hint 获得好的执行计划

通常状况下，经过 index， leading，use_hash or use_nl  就能搞定，估计80%~90%的sql用这几个hint就能搞定了

实在搞不定的，再去找主管 dba看下

对于 10g以上库，有个理论上方法，将DG 设置回归点后激活，将sql相关的对象的统计信息回退到 好执行计划的时间点

对 explain plan for sqlstat 作10053 的trace，好的执行计划的hint就能直接在trace文件中找到

反正无论怎么说，只要肯花点时间，总能作出来

 

步骤4、outline 互换

第一步：先为加Hint后的好的执行计划生成outline：

CREATE OR REPLACE OUTLINE zhuofhb for category special ON

SELECT T1.DISPATCHNO,

       T1.AUTO_TASK_NO,

       T1.TASK_NO,

       T1.REPORT_ID,

       T1.TASK_DEPT,

       T1.HANDLER_ID,

       T1.CAR_MARK,

       T1.ARRIVE_TIME,

       T1.ARRIVE_TIMES,

       T1.TASK_STATUS,

       T1.APPLY_USER,

       T1.APPLY_TACHE_ID,

       T1.TASK_ATTRIBUTE,

       T1.REMARK,

       T1.CREATED_BY,

       T1.CREATED_DATE,

       T1.UPDATE_BY,

       T1.UPDATE_DATE,

       T1.LOCAL_FLAG

 FROM T_AUTO_DISPATCH_TASK T1

 WHERE (T1.TASK_STATUS = '2' OR T1.AUTO_TASK_NO IS NULL)

   AND EXISTS (SELECT *

          FROM (SELECT DISTINCT T.DISPATCHNO

                  FROM (SELECT T2.DISPATCHNO

                          FROM T_AUTO_DISPATCH_COMMON T2

                         WHERE T2.SYNCSTATUS = '0'

                           AND T2.REPORTFLAG = 'Y'

                           AND T2.CASE_STATUS = '1'

                           AND EXISTS

                         (SELECT 1

                                  FROM C_ACC_SYNC_INFO T

                                 WHERE T.SYNC_STATUS
||= '2'

                                   AND T.CASENO = T2.CASENO)

                         ORDER BY T2.CASE_DIS_TIME) T

                 WHERE ROWNUM < 20) XX

         WHERE XX.DISPATCHNO = T1.DISPATCHNO)

 ORDER BY T1.ARRIVE_TIME

;   

第二步：再为旧的坏的执行计划生成outline：

CREATE OR REPLACE OUTLINE zhuofha for category special ON

SELECT T1.DISPATCHNO,

       T1.AUTO_TASK_NO,

       T1.TASK_NO,

       T1.REPORT_ID,

       T1.TASK_DEPT,

       T1.HANDLER_ID,

       T1.CAR_MARK,

       T1.ARRIVE_TIME,

       T1.ARRIVE_TIMES,

       T1.TASK_STATUS,

       T1.APPLY_USER,

       T1.APPLY_TACHE_ID,

       T1.TASK_ATTRIBUTE,

       T1.REMARK,

       T1.CREATED_BY,

       T1.CREATED_DATE,

       T1.UPDATE_BY,

       T1.UPDATE_DATE,

       T1.LOCAL_FLAG

 FROM T_AUTO_DISPATCH_TASK T1

 WHERE (T1.TASK_STATUS = '2' OR T1.AUTO_TASK_NO IS NULL)

   AND EXISTS (SELECT *

          FROM (SELECT DISTINCT T.DISPATCHNO

                  FROM (SELECT T2.DISPATCHNO

                          FROM T_AUTO_DISPATCH_COMMON T2

                         WHERE T2.SYNCSTATUS = '0'

                           AND T2.REPORTFLAG = 'Y'

                           AND T2.CASE_STATUS = '1'

                           AND EXISTS

                         (SELECT 1

                                  FROM C_ACC_SYNC_INFO T

                                 WHERE T.SYNC_STATUS = '2'

                                   AND T.CASENO = T2.CASENO)

                         ORDER BY T2.CASE_DIS_TIME) T

                 WHERE ROWNUM < 20) XX

         WHERE XX.DISPATCHNO = T1.DISPATCHNO)

 ORDER BY T1.ARRIVE_TIME

;   

第三步：将两个outline的名字互换，让outline生效

UPDATE OUTLN.OL$HINTS

SET OL_NAME=DECODE(OL_NAME,'ZHUOFHB','ZHUOFHA','ZHUOFHA','ZHUOFHB')

WHERE OL_NAME IN ('ZHUOFHB','ZHUOFHA');

 

exec dbms_outln.update_signatures;

alter system set use_stored_outlines=special;

第四步：删除不用的outline

--这个必定要作，不然之后他们下发优化后的 sql语句后，会用到这个outline中的坏的执行计划

drop outline zhuofhb;

第五步：验证

验证执行计划，对 pol_main不走全表扫描就对了  

select hash_value,sql_text,outline_category,child_number from v$sql where hash_value=3671808433