RE: Delete Performance Issue

From: Ken Naim <kennaim_at_gmail.com>
Date: Thu, 21 Dec 2006 17:04:21 -0500
Message-ID: <005f01c7254b$fb0df7c0$68b016ac@KenHome>

Most likely it is doing a full scan/hash join per distinct value or something along those lines.  

---

From: oracle-l-bounce_at_freelists.org [mailto:oracle-l-bounce_at_freelists.org] On Behalf Of Don Doo
Sent: Thursday, December 21, 2006 4:39 PM To: Stephane Faroult
Cc: oracle-l_at_freelists.org
Subject: Re: Delete Performance Issue  

Hi Stephane,  

The targetperiod and runseq are not part of primary key. The target period represent weeks
and we have about 100 distinct values for that ( 2 years of data) and runseq has about 750
distinct values. The table has 65 million rows. Just for testing I created an index on these
two columns and the query was not using the index. I think that is the right thing.
The query has the same execution plan when it deletes 200,000 rows and 2.5 million rows.
It is using full scans for three tables and one NL and two hash joins. What could make a
hash join delete so slow when the number of records increases. I would really appreciate if
somebody could shed some light into that.  

Regards,  

Don  

On 12/21/06, Stephane Faroult <sfaroult_at_roughsea.com> wrote: Don,

   You don't say what your primary key actually is. Might it be (targetperiod, runseq)? This would not sound too bad to me. In that case, perhaps that

delete from c_trace
where (targetperiod, runseq) in
(select 200612, RunSeq
from C_Run pr,

              C_Summary ss,
              C_Stage st
     where  pr.RunSeq = ss.RunSeq 
               and    ss.stageType = st.stageType
               and    st.name = 'load'
             and    pr.period = 361
      and    ss.Active   = 'yes')

would help Oracle to see the light. I think that in such a case I'd pay much attention to things such as the clustering factor (for which, of course, you cannot do much. But it may give you an idea about what you can hope for) and the order of the columns in the PK index, that may or may not favor an effective index scan.

HTH Stephane Faroult

Don Doo wrote:
>
> Hi,
>
> We are facing a serious performance
> issue. This is a delete statement and
> it takes 4 hours to delete 2.3 million
> rows from a 65 million row table.
> The c_trace table has only one index
> (the primary key) The table is not
> partitioned (we don't have the budget
> to pay for partition option) Other tables
> have less than 5000 rows.
> The sub-query returns 3 to 6 rows
> depending on the values
>
> Query
>
> delete from c_trace
> where targetperiod= 200612
> and RUNSEQ in (select
> RunSeq from C_Run pr,
> C_Summary ss,
> C_Stage st
> where pr.RunSeq = ss.RunSeq
> and ss.stageType = st.stageType
> and st.name <http://st.name> = 'load'
> and pr.period = 361
> and ss.Active = 'yes')
>
>
> The V$session_longops shows
>
> 1 select OPNAME||' '||MESSAGE||' '||ELAPSED_SECONDS from
> v$session_longops
> 2* where sql_hash_value=2467621466
> SQL> /
> OPNAME||''||MESSAGE||''||ELA
>

---

---

>
> Hash Join Hash Join: : 6592 out of 6592 Blocks done 13688
> Hash Join Hash Join: : 6272 out of 6272 Blocks done 12753
> Hash Join Hash Join: : 6272 out of 6272 Blocks done 13594
> Hash Join Hash Join: : 7488 out of 7488 Blocks done 14050
>
> Looks like it takes 14050 seconds to complete the hash join which
> matches the time taken to complete the delete.
>
> select HASH_VALUE,CPU_TIME,elapsed_time/(1000000*60),
> fetches,disk_reads,
> BUFFER_GETS,ROWS_PROCESsed
> from v$sql where hash_value = 2467621466
>
> HASH_VALUE CPU_TIME FETCHES DISK_READS BUFFER_GETS ROWS_PROCESSED
> ---------- ---------- ------------------------- ---------- ----------
> ----------- --------------
> 2467621466 193010000 0 1265770
> 13820713 2325397
>
> Oracle version 9.2.0.4 < http://9.2.0.4 <http://9.2.0.4> >
> Hash_area_size 8 MB
> Sort_area_size 4 MB
> Statistics are current..
> We are using ASSM for these tables.
> I would really appreciate any ideas to improve this statement. We
> cannot do a event trace
> here until mid of January next year because we are behind the SLA and
> don't want to make it slower.
>
> Regards,
>
> Don
>
 

--
http://www.freelists.org/webpage/oracle-l

Received on Thu Dec 21 2006 - 16:04:21 CST