RE: Performance issue post database analyze, need to understand whats wrong in analyze

As a good friend of mine told me: train time is think time so I thought I'd add my comments to this.  

1. Why did the plan change - see Niall's and Greg's answers for an explanation. Additionally I like to check the plans for a statement (proper licensing assumed) using dbms_xplan.display_awr for that sql id as well. It helped greatly in understanding the problem after developers told me "I have no idea what the good plan is. All I know is that it's slow now"
2. Preventing unpredictable execution plans with partitioned tables has been discussed a number of times by much more knowledgeable people on this list. In my experience (that was on 10.2.0.4 Linux/3 node RAC) copying partition statistics forward seemed to be a great way of getting predictable performance, especially if you partition by range on a time column

Hope this helps,  

Martin  

Martin Bach

Martin Bach Consulting Ltd

http://www.linkedin.com/in/martincarstenbach

http://martincarstenbach.wordpress.com  

From: oracle-l-bounce_at_freelists.org [mailto:oracle-l-bounce_at_freelists.org] On Behalf Of Shastry(DBA)
Sent: 07 June 2011 06:30
To: joshuasingham_at_gmail.com
Cc: oracle-l
Subject: Re: Performance issue post database analyze, need to understand whats wrong in analyze  

Hi Joshua,

Thanks for the response. I am looking for the approach to find the issue where post analyze why the plan got changed? How to dig out on this and any measures to prevent the same issue in coming analyze job?

Regards,
Anantha

On Tue, Jun 7, 2011 at 10:01 AM, joshuasingham <joshuasingham_at_gmail.com> wrote:

Hi Shastry,  

Instead of using a hint you can use a solution for plan stability like stored oulines, sql profile (see kerry osborne blog for more details) and if you using 11g you can use SQL plan management .  

Regards  

joshua

On Tue, Jun 7, 2011 at 11:44 AM, Shastry(DBA) <shastry17_at_gmail.com> wrote:

Hi Team,  

I need your help in knowing whats the issue with my database analyze? Last Sunday analyze job has been run and post that the explain plan got changed which caused severe performance issue with high GC BUFFER BUSY waits. Please help me in understanding the issue.  

Here is the problematic query :  

SELECT DECODE (COUNT (1), 0, 'N', 'Y')
FROM CPR_LDAP_SUMMARY
WHERE LOWER (USERID) = :B1 AND UPPER (IPCACCESS) IN ('12', 'YES')   And plan details from stage & Prod    

Plan
SELECT STATEMENT ALL_ROWSCost: 4 Bytes: 15 Cardinality: 1 3 SORT AGGREGATE Bytes: 15 Cardinality: 1 2 TABLE ACCESS BY GLOBAL INDEX ROWID TABLE DMADM.CPR_LDAP_SUMMARY Cost: 4 Bytes: 15 Cardinality: 1 Partition #: 2 Partitions accessed #1 1 INDEX RANGE SCAN INDEX DM.CPR_LS_USERID_IDX Cost: 3 Cardinality: 1

Prod plan is

Plan
SELECT STATEMENT ALL_ROWSCost: 2 Bytes: 566 Cardinality: 1 3 SORT AGGREGATE Bytes: 566 Cardinality: 1 2 PARTITION RANGE SINGLE Cost: 2 Bytes: 566 Cardinality: 1 Partition #: 2 Partitions accessed #1
1 TABLE ACCESS FULL TABLE DM.CPR_LDAP_SUMMARY Cost: 2 Bytes: 566 Cardinality: 1 Partition #: 3 Partitions accessed #1  

The issue was with plan change since last gather stats and performance restored after re-gathering stats.  

SQL_ID=dx87pq17pthq1
Bad Plan Hash=1572312302 (FTS)
Good Plan Hash= 2166829021 (function based index usage)

The plan change from index usage to FTS since 05-jun-11 18hrs and gather stats on this table was on 05-jun-11 10:00 hrs.

After re-gathering stats on DM.CPR_LDAP_SUMMARY the static plan used index again and performance improved.

If the issue re-appears with next gather stats, perf team suggested to consider using following hint:
SELECT /*+ index(CPR_LDAP_SUMMARY CPR_LS_USERID_IDX) */ COUNT (1) ...  

Thanks,

Shastry    

--

http://www.freelists.org/webpage/oracle-l Received on Tue Jun 07 2011 - 01:33:05 CDT