Wednesday, January 30, 2008

Rolling invalidations

There have been discussions which I have seen related to the feature of auto invalidation in dbms_stats. A couple of references are

http://forums.oracle.com/forums/thread.jspa?threadID=592771&tstart=30
and
http://www.orafaq.com/maillist/oracle-l/2006/10/10/0429.htm

I have tested the relevant parameter “_optimizer_invalidation_period” on 10.2.0.3 and believe that this is working as expected

Let us take the below testcase where the parameter (it is dynamic) is set to a value of 120

SQL> show parameter optimizer_inva

NAME TYPE VALUE
------------------------------------ ----------- ------------------------------
_optimizer_invalidation_period integer 120

We have the following sql statement

11:00:00 SQL> select * from source where rownum<2;

OBJ# LINE SOURCE
---------- ---------- ---------------------------------------------------------------------------
194107 171 -- *Action: Start a new job, or attach to an existing job that has a


1 row selected.

Elapsed: 00:00:00.12
11:00:00 SQL> select * from source where rownum<2;

OBJ# LINE SOURCE
---------- ---------- ---------------------------------------------------------------------------
194107 171 -- *Action: Start a new job, or attach to an existing job that has a


1 row selected.

Elapsed: 00:00:00.00
11:00:00 SQL> select a.child_number,LAST_LOAD_TIME, to_char(LAST_ACTIVE_TIME,'dd-mon-yyyy hh24:mi:ss') ,b.invalidations from
v$SQL_SHARED_CURSOR a, v$sql b where a.sql_id='954g5yyw5tn1s' and a.child_address=b.child_address ;

CHILD_NUMBER LAST_LOAD_TIME TO_CHAR(LAST_ACTIVE_ INVALIDATIONS
------------ ------------------- -------------------- -------------
0 2008-01-29/11:00:00 29-jan-2008 11:00:00 0

1 row selected.

Elapsed: 00:00:00.14
11:00:00 SQL>
11:00:00 SQL> select executions, invalidations,child_number from v$sql where sql_id='954g5yyw5tn1s';

EXECUTIONS INVALIDATIONS CHILD_NUMBER
---------- ------------- ------------
2 0 0

1 row selected.

Now we gather stats on the table with the auto_invalidate parameter passed to the API.

11:00:00 SQL> exec dbms_stats.gather_table_stats('REGOFA','SOURCE',no_invalidate => DBMS_STATS.AUTO_INVALIDATE);

PL/SQL procedure successfully completed.

Elapsed: 00:00:01.50

Then we keep executing the sql statement of interest to check when the new cursor will be generated.

Elapsed: 00:00:01.50
11:00:13 SQL> select a.child_number,LAST_LOAD_TIME, to_char(LAST_ACTIVE_TIME,'dd-mon-yyyy hh24:mi:ss') ,b.invalidations from
v$SQL_SHARED_CURSOR a, v$sql b where a.sql_id='954g5yyw5tn1s' and a.child_address=b.child_address ;

CHILD_NUMBER LAST_LOAD_TIME TO_CHAR(LAST_ACTIVE_ INVALIDATIONS
------------ ------------------- -------------------- -------------
0 2008-01-29/11:00:00 29-jan-2008 11:00:09 0

1 row selected.

Elapsed: 00:00:00.05
11:00:13 SQL> select executions, invalidations,child_number from v$sql where sql_id='954g5yyw5tn1s';

EXECUTIONS INVALIDATIONS CHILD_NUMBER
---------- ------------- ------------
3 0 0

1 row selected.

Elapsed: 00:00:00.00
11:00:13 SQL> select * from v$sql_shared_cursor where sql_id='954g5yyw5tn1s';

SQL_ID ADDRESS CHILD_ADDRESS CHILD_NUMBER U S O O S L S E B P
------------- ---------------- ---------------- ------------ - - - - - - - - - -
I S T A B D L T R I I R L I O S M U T N F A I T D L D B P C S R P T M B M R O P
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
M F L
- - -
954g5yyw5tn1s 00000007D3BBCBD8 00000007D5644028 0 N N N N N N N N N N
N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N
N N N


1 row selected.
…….
11:00:37 SQL> select * from source where rownum<2;

OBJ# LINE SOURCE
---------- ---------- ---------------------------------------------------------------------------
194107 171 -- *Action: Start a new job, or attach to an existing job that has a


1 row selected.

Elapsed: 00:00:00.01
11:00:39 SQL> select a.child_number,LAST_LOAD_TIME, to_char(LAST_ACTIVE_TIME,'dd-mon-yyyy hh24:mi:ss') ,b.invalidations from
v$SQL_SHARED_CURSOR a, v$sql b where a.sql_id='954g5yyw5tn1s' and a.child_address=b.child_address ;

CHILD_NUMBER LAST_LOAD_TIME TO_CHAR(LAST_ACTIVE_ INVALIDATIONS
------------ ------------------- -------------------- -------------
0 2008-01-29/11:00:00 29-jan-2008 11:00:25 0
1 2008-01-29/11:00:37 29-jan-2008 11:00:37 0

2 rows selected.

Elapsed: 00:00:00.04
11:00:39 SQL> select executions, invalidations,child_number from v$sql where sql_id='954g5yyw5tn1s';

EXECUTIONS INVALIDATIONS CHILD_NUMBER
---------- ------------- ------------
7 0 0
1 0 1

2 rows selected.

Elapsed: 00:00:00.00
11:00:39 SQL> select * from v$sql_shared_cursor where sql_id='954g5yyw5tn1s';

SQL_ID ADDRESS CHILD_ADDRESS CHILD_NUMBER U S O O S L S E B P
------------- ---------------- ---------------- ------------ - - - - - - - - - -
I S T A B D L T R I I R L I O S M U T N F A I T D L D B P C S R P T M B M R O P
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
M F L
- - -
954g5yyw5tn1s 00000007D3BBCBD8 00000007D5644028 0 N N N N N N N N N N
N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N
N N N

954g5yyw5tn1s 00000007D3BBCBD8 00000007D3753DC0 1 N N N N N N N N N N
N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N Y N N
N N N

So somewhere between 11:00:00 and 11:00:39 (within the 2 minute window) a new child cursor has been generated with roll_invalid_mismatch set to ‘Y”

I have tested for the following values of _optimizer_invalidation_period and I see consistent results

120
210
600
1800
18000

Hence this would be an ideal way to avoid a hard parse storm

6 comments:

Timur Ahmadeev said...

Hi, Fairlie.
Thank you for a number of interesting postings.
I think one thing is missing in all your posts: pretty output formatting :)
Could you please fix it?

Fairlie Rego said...

Thanks for your kind words.. Will try to fix it. would love to move to wordpress but time is the problem...

Dion_Cho said...

Greate test and analysis. Thanks

Konstantinos said...

Great test and analysis. Thank's

Anonymous said...

Thanks for the info.

I have one question - once the new child cursor is created, what happens to the old child cursor? Currently I have a situation where the old child cursor does a full table scan and the new one uses an index. From time to time, oracle randomly seems to pick the old plan and cause a spike in response time. Once a new child is created, when does oracle invalidate or flush out the old cursor?

Thanks in advance!

Anonymous said...

Greetings

Nice one, Just like the previous poster, I too have the same issue of old cursor being executed, for now I am flushing the shared_pool. Planning to test Stored outlines next week.

Plese keep up the good work

Regards & Thanks
BN