Archive

Archive for the ‘Trouble’ Category

Cardinality estimates on list subpartitioned tables with binds on partition, literals on subpartition

19/08/2012 Leave a comment

Acutually, I decided blogging is just too much pain especially the formating. Additionally considering the good quality of most of the other blogs, I would not want distribute half knowledge. However as some blog services do not allow direct formated long comments or limit comment to 4K, I will have to do this here (but even this does not format as I want).

While the bugs mentioned in http://oracle-randolf.blogspot.ch/2009/03/partition-oddities-part-1-severe-bug-in.html are according to oracle support fixed, still not all works well in 11.2.0.2. At least for range partitioned, sub list partitioned in combination with partial use of bind variables and add partition operations.

A partitioned table with identical subpartitions seems to work ok, however add one subpartition to the maxvalue partition (or any other or remove one) and you are back to square one, except for the first subpartition.

Unless I made an error or misunderstood something, a plan with cardinality 1 results:

	SQL_ID  a1d6qv0615qjg, child number 0
	-------------------------------------
	SELECT /*+ full(t) 1 off h*/ COUNT(*) FROM TESTSZ WHERE I=:B1 AND J=1
	 
	Plan hash value: 835289084
	 
	--------------------------------------------------------------------------------------------------
	| Id  | Operation               | Name   | Rows  | Bytes | Cost (%CPU)| Time     | Pstart| Pstop |
	--------------------------------------------------------------------------------------------------
	|   0 | SELECT STATEMENT        |        |       |       |    14 (100)|          |       |       |
	|   1 |  SORT AGGREGATE         |        |     1 |    10 |            |          |       |       |
	|   2 |   PARTITION RANGE SINGLE|        |     1 |    10 |    14   (0)| 00:00:01 |   KEY |   KEY |
	|   3 |    PARTITION LIST SINGLE|        |     1 |    10 |    14   (0)| 00:00:01 |       |       |
	|*  4 |     TABLE ACCESS FULL   | TESTSZ |     1 |    10 |    14   (0)| 00:00:01 |   KEY |   KEY |
	--------------------------------------------------------------------------------------------------
	 
	Predicate Information (identified by operation id):
	---------------------------------------------------
	 
	   4 - filter(("J"=1 AND "I"=:B1))
	 

following first, abbreviated output of plan of different sql statements, then the script:

WITH added subpartition, note the cardinality estimate of 1 for j=1, j=2 (statement 3 and 4)
PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------------------------------
SELECT /*+ full(t) f off h*/ COUNT(*) FROM TESTSZ WHERE
|* 4 | TABLE ACCESS FULL | TESTSZ | 90 | 990 | 14 (0)| 00:00:01 | 2 | 2 |
SELECT /*+ full(t) 0 off h*/ COUNT(*) FROM TESTSZ WHERE I=:B1 AND J=0
|* 4 | TABLE ACCESS FULL | TESTSZ | 30 | 300 | 14 (0)| 00:00:01 | KEY | KEY |
SELECT /*+ full(t) 1 off h*/ COUNT(*) FROM TESTSZ WHERE I=:B1 AND J=1
|* 4 | TABLE ACCESS FULL | TESTSZ | 1 | 10 | 14 (0)| 00:00:01 | KEY | KEY |
SELECT /*+ full(t) 2 off h*/ COUNT(*) FROM TESTSZ WHERE I=:B1 AND J=2
|* 4 | TABLE ACCESS FULL | TESTSZ | 1 | 10 | 14 (0)| 00:00:01 | KEY | KEY |

SELECT /*+ full(t) 3 off h*/ COUNT(*) FROM TESTSZ WHERE I=:B2 AND J=:B1
|* 4 | TABLE ACCESS FULL | TESTSZ | 100 | 1100 | 14 (0)| 00:00:01 | KEY | KEY |

WITHOUT added subpartition resulting in reasonable estimates

PLAN_TABLE_OUTPUT
---------------------------------------------------------------------------------------------------------
SELECT /*+ full(t) f off h*/ COUNT(*) FROM TESTSZ WHERE
|* 4 | TABLE ACCESS FULL | TESTSZ | 90 | 990 | 14 (0)| 00:00:01 | 2 | 2 |
SELECT /*+ full(t) 0 off h*/ COUNT(*) FROM TESTSZ WHERE I=:B1 AND J=0
|* 4 | TABLE ACCESS FULL | TESTSZ | 30 | 300 | 14 (0)| 00:00:01 | KEY | KEY |
SELECT /*+ full(t) 1 off h*/ COUNT(*) FROM TESTSZ WHERE I=:B1 AND J=1
|* 4 | TABLE ACCESS FULL | TESTSZ | 90 | 990 | 14 (0)| 00:00:01 | KEY | KEY |
SELECT /*+ full(t) 2 off h*/ COUNT(*) FROM TESTSZ WHERE I=:B1 AND J=2
|* 4 | TABLE ACCESS FULL | TESTSZ | 180 | 1980 | 14 (0)| 00:00:01 | KEY | KEY |

SELECT /*+ full(t) 3 off h*/ COUNT(*) FROM TESTSZ WHERE I=:B2 AND J=:B1
|* 4 | TABLE ACCESS FULL | TESTSZ | 100 | 1100 | 14 (0)| 00:00:01 | KEY | KEY |

and the script to generate above output:

set linesize 200
set pagesize 50
set echo on
set feedback off
column plan_table_output format a120

drop table testsz purge;
create table testsz
(i date, j number, v varchar2(100))
partition by range(i)
subpartition by list (j)
subpartition template
(subpartition p0 values (0),
 subpartition p1 values (1),
 subpartition p2 values (2)
)
(
partition p_100 values less than (to_date('3101201201','DDMMYYYYSS')),
partition p_200 values less than (to_date('3103201201','DDMMYYYYSS')),
partition p_max values less than (maxvalue)
);

alter table testsz modify partition p_max add subpartition values(3);


insert into testsz
select to_date('31012012','DDMMYYYY'),
        case when mod(rownum,100)  <10 then 0 when mod(rownum,100) <40 then 1 else 2 end,
       lpad('x',100)
  from dual connect by level <301;

insert into testsz
select to_date('31032012','DDMMYYYY'),
        case when mod(rownum,100)  <10 then 0 when mod(rownum,100) <40 then 1 else 2 end,
       lpad('x',100)
  from dual connect by level <301;
  
commit;

begin
 dbms_stats.gather_table_stats(ownname => user,tabname => 'TESTSZ',
                               method_opt => 'for all columns size 1',no_invalidate => false);
end;
/

-- info
select owner, table_name, partitioning_type, subpartitioning_type, partition_count, def_subpartition_count
 from dba_part_Tables where table_name='TESTSZ' and owner like 'TEST%';
select table_name, partition_name, num_rows,partition_position from user_tab_partitions where table_name='TESTSZ'
order by 1;
select table_name, subpartition_name, num_rows from user_tab_subpartitions where table_name='TESTSZ'
order by 2;

--
declare
 i1 date:=to_date('31012012','DDMMYYYY');
 h number:=1;
 j number;
begin
  select  /*+ full(t) f off h*/ count(*) into j from testsz where  I=to_date('31012012','DDMMYYYY')and j=1;
  select  /*+ full(t) 0 off h*/ count(*) into j from testsz  where I=i1 and j=0;
  select  /*+ full(t) 1 off h*/ count(*) into j from testsz  where I=i1 and j=1;
  select  /*+ full(t) 2 off h*/ count(*) into j from testsz  where I=i1 and j=2;
  select  /*+ full(t) 3 off h*/ count(*) into j from testsz  where I=i1 and j=h;
  --
  select  /*+ full(t) 4a off h*/ count(*) into j from testsz  where I=to_date('31012012','DDMMYYYY');
  select  /*+ full(t) 4b off h*/ count(*) into j from testsz  where I=to_date('31012012','DDMMYYYY') and j=h;
end;
/  

select t.sql_id, o.* from v$sql t ,
      table(dbms_xplan.display_cursor(t.sql_id, t.CHILD_NUMBER,'ADVANCED')) o
 where sql_Text like 'SELECT %full%TESTSZ%%'
 and (o.plan_table_output like '%TESTSZ %')
 and last_active_time > sysdate -1/24/60
 order by t.sql_text, o.plan_table_output desc
;

pause ...
Categories: Trouble, Uncategorized

Influence of new Oracle version – well actually not

22/04/2011 Leave a comment

Recently I have been working on application testing  related to a migration from 10.2.0.3 to 11.1.0.7.  New hardware as well. More and faster CPUs and significantly faster IO created some expectations.

So (as usual) some significantly slower batch jobs  were a bad, but somewhat expected surprise. Here is one example:

A batch job took 37 minutes instead of 15 minutes. I took one of the worst performing statements, taking 270 instead of 60 seconds, to investigate. Rather quickly the usual suspects, such as a changed plan, could be ruled out. As a matter of fact, after a while I could reduce the statement to a simple  reproducible test.

Preparation:


create table testsz
(i number, wa varchar2(5), vt varchar2(1), v varchar2(400));

create table testsz2
(i number, wa varchar2(5), vt varchar2(1), v varchar2(400));
create index testsz2_i on testsz2(wa,vt);

insert into testsz
with gen as (select rownum rn  from dual connect by level<=1e5)
select mod(rn, 1000),
       case  mod(rn,3) when 0 then 'A' when 1 then 'B' when 2 then 'C' else 'D' end||mod(rn, 2000),
       case when mod(rn,1) =0 then 'Y' else 'Q' end,
       lpad('x',300,'x'
 from gen;
commit;

Test:

set timing on

truncate table testsz2;

insert into testsz2
  select *
   from testsz;

The exact times vary depending on system and hardware, but the difference was significant on all systems tested.

Results on 10g vs 11g:   5 seconds vs 20 seconds.

This bugged me quite some time. I did all kinds of really stupid and other non-relevant stuff to investigate, to no avail. There must be something wrong with Oracle, right?

Finally I detected a increase in FB enqueue contention on 11g.  So there must be something wrong with the new Oracle release, right?

Looking for bug reports like mad, nothing found. (Not that the increase actually was significant, and if it had not been on ASSM tablespaces it would have been different anyway). Still, this triggered me to remember, that I asked a few weeks ago, about a changed init setting. But did not bother to follow up. Quickly checked, luckily the parameter  is session modifiable:

Modified test, on 11g only:

set timing on

prompt block checking false
alter session set db_block_checking=false;
truncate table testsz2;

insert into testsz2
select * from testsz;

prompt block checking true
alter session set db_block_checking=true;
truncate table testsz2;

insert into testsz2
select * from testsz;

3 seconds vs 20 seconds. So this has nothing to do with the new oracle release, but a changed initialization parameter. I have not tested much further about the causes, but the difference is also there in 10g.  At least on the systems I have access to.

Obviously the db_block_checking setting has an overhead, but frankly I am surprised by the magnitude, considering the quotes in the oracle documentation (from 11.1):

Block checking typically causes 1% to 10% overhead, depending on workload and the parameter value. The more updates or inserts in a workload, the more expensive it is to turn on block checking. You should set DB_BLOCK_CHECKING to FULL if the performance overhead is acceptable.

That difference is certainly not acceptable to me. Anyone else seen such a big difference?

Categories: Performance, Trouble

TM enqueues and deadlock with one user session

16/04/2011 2 comments

For quite some time now, I have been thinking about writing some blog items, but never did. A recent blog entry on http://jonathanlewis.wordpress.com/2011/04/08/deadlock prompted me to give it a try.

So the first post is just a repetition of the comment I made:

TM-enqueues for foreign key locking seems to be a well known fact. What I personally would like to know, what other reasons can cause TM-Locks.
A few times already I had issues with them, not involving foreign keys.

Another reason are exclusive lock and the following. Did anyone know, that a single user connected session can cause a deadlock?

on 10.2.0.3 and 11.1.0.7 verified:


DROP TABLE TESTSZ;
CREATE TABLE TESTSZ
( I NUMBER,
  K VARCHAR2(50));
 INSERT INTO TESTSZ
 SELECT MOD(ROWNUM,1000), LPAD(‘X’,40,’Y') FROM DUAL CONNECT BY ROWNUM < 1000000;
 COMMIT;

LOCK TABLE TESTSZ IN EXCLUSIVE MODE;
 INSERT INTO TESTSZ X
 WITH X1 AS  (SELECT /*+ no_PARALLEL(T) MATERIALIZE*/ 2000, K FROM TESTSZ T
 WHERE I=1)
 SELECT * FROM X1;
 rollback;

LOCK TABLE TESTSZ IN EXCLUSIVE MODE;
 set timing on
 INSERT INTO TESTSZ X
 WITH X1 AS (SELECT /*+ PARALLEL(T,4) MATERIALIZE*/ 2000, K FROM TESTSZ T
 WHERE I=1)
 SELECT * FROM X1;
 rollback;

Result of the last insert statement:

INSERT INTO TESTSZ X
*
FEHLER in Zeile 1:
ORA-12801: error signaled in parallel query server P000
ORA-00060: deadlock detected while waiting for resource

Before anyone complains there are more than one session involved,
yes, but the importance is only one “user” connected session.

Moving back to TM locks, in the case above with the factored subquery
in combination with parallel query, the deadlock occurs while the sessions
are trying to get TM-locks.

SID PROGRAM LAST_CALL_ET BLOCKING_SESSION EVENT SECONDS_IN_WAIT
116 sqlplus.exe               48        PX Deq: Parse Reply     3  WAITING
117 oracle@cyber-ora10 (P005) 48        PX Deq: Execution Msg   3  WAITING
128 oracle@cyber-ora10 (P004) 48        PX Deq: Execution Msg   3  WAITING
115 oracle@cyber-ora10 (P003) 48        PX Deq: Execution Msg   3  WAITING
129 oracle@cyber-ora10 (P002) 48  116   enq: TM – contention   48 WAITING
120 oracle@cyber-ora10 (P001) 48  116   enq: TM – contention   48 WAITING
136 oracle@cyber-ora10 (P000) 48  116   enq: TM – contention   48 WAITING

To finish, Oracle support agreed, that this is not supposed to happen. The workaround is to use the inline hint.

And to credit, this originally happened to a co-worker sitting two desks next to me.

Categories: Locking, Trouble