Wrong Result Bug using group-by placement optimization

Wednesday, February 5, 2014

Last week I’ve mentioned on Twitter that we ran into wrong result bug. We found workaround quickly but I’ve decided to spend some time to reproduce error and write blog post to warn you about this optimizer behavior.
Special thanks to my colleague who spotted odd results which led us to this finding.

My test (virtual) environment is:
OS: Oracle Enterprise Linux 5.8
DB: Oracle EE 11.1.0.7.12

In test I will use three tables:

CONT
Name    Null Type          
------- ---- ------------- 
CUST_ID      NUMBER(38)    
CODE         VARCHAR2(100) 

CUST
Name    Null     Type       
------- -------- ---------- 
CUST_ID NOT NULL NUMBER(38) 

DRAG
Name    Null Type      
------- ---- --------- 
DRAG_ID      NUMBER(6) 


To gather fresh statistics for the tables:
begin
  dbms_stats.gather_table_stats(ownname=>user,tabname=>'CONT',estimate_percent=>100, cascade=>TRUE);
  dbms_stats.gather_table_stats(ownname=>user,tabname=>'CUST',estimate_percent=>100, cascade=>TRUE);
  dbms_stats.gather_table_stats(ownname=>user,tabname=>'DRAG',estimate_percent=>100, cascade=>TRUE);
end;
/

More details about tables:
select table_name, num_rows, blocks, partitioned, last_analyzed
from dba_tables
where table_name in ('CONT','CUST','DRAG');


TABLE_NAME     NUM_ROWS     BLOCKS PARTITIONED LAST_ANALYZED     
------------ ---------- ---------- ----------- -------------------
CONT            1181949       2892 NO          04.02.2014 14:49:24 
DRAG                314          5 NO          04.02.2014 14:49:25 
CUST             576233        902 NO          04.02.2014 14:49:25 

Information about indexes:
select index_name, table_name, uniqueness, distinct_keys, clustering_factor
from dba_indexes
where table_name in ('CONT','CUST','DRAG');


INDEX_NAME     TABLE_NAME   UNIQUENESS DISTINCT_KEYS CLUSTERING_FACTOR
-------------- ------------ ---------- ------------- -----------------
I_CUST_ID      CONT         NONUNIQUE         468738            753983 
PK_CUST_ID     CUST         UNIQUE            576233               878 

We have three small and simple tables with just two indexes. CUST table has primary key on “cust_id” column.


After this little introduction it is time for some tests.

I will flush buffer cache and shared pool before every query execution.
SQL> alter system flush shared_pool;
System altered.

SQL> alter system flush buffer_cache;
System altered.

First query execution and execution plan:
select /*+ gather_plan_statistics */ 
       count(co.code) as cnt
from drag t, 
     cust cus, 
     cont co
where 1=1
and t.drag_id = cus.cust_id
and cus.cust_id = co.cust_id(+) 
group by t.drag_id;

            CNT
---------------
              2
              2
              2
              2
              1
              2
              2
              2
              2
...
303 rows
SQL> select * from table(dbms_xplan.display_cursor(null,null,'ALLSTATS LAST'));

PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------------------------------------------
SQL_ID  gpnrgy2vawafq, child number 0
-------------------------------------
select /*+ gather_plan_statistics */        count(co.code) as cnt from
drag t,      cust cus,      cont co where 1=1 and t.drag_id =
cus.cust_id and cus.cust_id = co.cust_id(+) group by t.drag_id

Plan hash value: 3989628059

---------------------------------------------------------------------------------------------------------------------------
| Id  | Operation            | Name | Starts | E-Rows | A-Rows |   A-Time   | Buffers | Reads  |  OMem |  1Mem | Used-Mem |
---------------------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT     |      |      1 |        |    303 |00:00:00.62 |    3734 |   3724 |       |       |          |
|   1 |  HASH GROUP BY       |      |      1 |    303 |    303 |00:00:00.62 |    3734 |   3724 |  1096K|  1096K| 1264K (0)|
|*  2 |   HASH JOIN OUTER    |      |      1 |    792 |   1084 |00:00:00.16 |    3734 |   3724 |  1206K|  1206K| 1244K (0)|
|*  3 |    HASH JOIN         |      |      1 |    314 |    314 |00:00:00.04 |     890 |    885 |  1452K|  1452K| 1470K (0)|
|   4 |     TABLE ACCESS FULL| DRAG |      1 |    314 |    314 |00:00:00.01 |       7 |      6 |       |       |          |
|   5 |     TABLE ACCESS FULL| CUST |      1 |    576K|    576K|00:00:00.02 |     883 |    879 |       |       |          |
|   6 |    TABLE ACCESS FULL | CONT |      1 |   1181K|   1181K|00:00:00.01 |    2844 |   2839 |       |       |          |
---------------------------------------------------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

   2 - access("CUS"."CUST_ID"="CO"."CUST_ID")
   3 - access("T"."DRAG_ID"="CUS"."CUST_ID")

Check result of the query - this is correct query result.

Now to simulate what we experienced in production.
SQL> alter system flush shared_pool;
System altered.
SQL> alter system flush buffer_cache;
System altered.   

With hint I want to force PK_CUST_ID index usage because this was preferred plan in production.
select /*+ gather_plan_statistics index(cus PK_CUST_ID) */ 
       count(co.code) as cnt
from drag t, 
     cust cus, 
     cont co
where 1=1
and t.drag_id = cus.cust_id
and cus.cust_id = co.cust_id(+) 
group by t.drag_id;

            CNT
---------------
              0
              0
              0
              0
              0
              0
              0
              0
              0
              0
... 303 rows
SQL> select * from table(dbms_xplan.display_cursor(null,null,'ALLSTATS LAST'));

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------------------------------------------------------
SQL_ID  9vf9uf7mhdmdz, child number 0
-------------------------------------
select /*+ gather_plan_statistics index(cus PK_CUST_ID) */
count(co.code) as cnt from drag t,      cust cus,      cont co where
1=1 and t.drag_id = cus.cust_id and cus.cust_id = co.cust_id(+) group
by t.drag_id

Plan hash value: 3263881209

-----------------------------------------------------------------------------------------------------------------------------------
| Id  | Operation              | Name       | Starts | E-Rows | A-Rows |   A-Time   | Buffers | Reads  |  OMem |  1Mem | Used-Mem |
-----------------------------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT       |            |      1 |        |    303 |00:00:00.70 |    3459 |   3094 |       |       |          |
|   1 |  HASH GROUP BY         |            |      1 |    303 |    303 |00:00:00.70 |    3459 |   3094 |   934K|   934K| 1267K (0)|
|*  2 |   HASH JOIN OUTER      |            |      1 |    764 |   1046 |00:00:00.22 |    3459 |   3094 |  1134K|  1134K| 1198K (0)|
|   3 |    NESTED LOOPS        |            |      1 |    303 |    303 |00:00:02.02 |     615 |    255 |       |       |          |
|   4 |     VIEW               | VW_GBC_9   |      1 |    303 |    303 |00:00:00.01 |       7 |      6 |       |       |          |
|   5 |      HASH GROUP BY     |            |      1 |    303 |    303 |00:00:00.01 |       7 |      6 |  1012K|  1012K| 1249K (0)|
|   6 |       TABLE ACCESS FULL| DRAG       |      1 |    314 |    314 |00:00:00.01 |       7 |      6 |       |       |          |
|*  7 |     INDEX UNIQUE SCAN  | PK_CUST_ID |    303 |      1 |    303 |00:00:00.22 |     608 |    249 |       |       |          |
|   8 |    TABLE ACCESS FULL   | CONT       |      1 |   1181K|   1181K|00:00:00.01 |    2844 |   2839 |       |       |          |
-----------------------------------------------------------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

   2 - access("CUS"."CUST_ID"="CO"."CUST_ID")
   7 - access("ITEM_1"="CUS"."CUST_ID")

Check result of the query!
Count is displaying all 0 values because it received only NULLs to count.
Other functions like max and min are also affected by this error.

Check steps 4,5 and 6 in execution plan.

Instead of quick full scan on DRAG table Oracle transformed query and created inline view using smart group-by optimization.

In 10053 trace I could easily find what Oracle was doing.
SELECT
  /*+ INDEX ("CUS" "PK_CUST_ID") */
  SUM("VW_GBC_9"."ITEM_2") "CNT"
FROM
  (SELECT "T"."DRAG_ID" "ITEM_1",
    COUNT("CO"."CODE") "ITEM_2",
    "T"."DRAG_ID" "ITEM_3"
  FROM "ADMIN"."DRAG" "T"
  WHERE 1=1
  GROUP BY "T"."DRAG_ID",
    "T"."DRAG_ID"
  ) "VW_GBC_9",
  "ADMIN"."CUST" "CUS",
  "ADMIN"."CONT" "CO"
WHERE "VW_GBC_9"."ITEM_1"="CUS"."CUST_ID"
AND "CUS"."CUST_ID"      ="CO"."CUST_ID"(+)
GROUP BY "VW_GBC_9"."ITEM_3";


Quick workaround to fix this bug:
- Set "_optimizer_group_by_placement"=FALSE.

You could check in 10053 trace value of this parameter.
In my case: _optimizer_group_by_placement = true

SQL> alter session set "_optimizer_group_by_placement"=FALSE;

Session altered.

select /*+ gather_plan_statistics index(cus PK_CUST_ID) */ 
       count(co.code) as cnt
from drag t, 
     cust cus, 
     cont co
where 1=1
and t.drag_id = cus.cust_id
and cus.cust_id = co.cust_id(+) 
group by t.drag_id;

       CNT
----------
         2
         2
         2
         2
         1
         2
         2
         2
         2
...
303 rows
SQL> select * from table(dbms_xplan.display_cursor(null,null,'ALLSTATS LAST'));

PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------------------------------------------
SQL_ID  a91bzhvupzquh, child number 0
-------------------------------------
select /*+ gather_plan_statistics index(cus PK_CUST_ID)*/
count(co.code) as cnt from drag t,      cust cus,      cont co where
1=1 and t.drag_id = cus.cust_id and cus.cust_id = co.cust_id(+) group
by t.drag_id

Plan hash value: 2460166079

------------------------------------------------------------------------------------------------------------------------
| Id  | Operation            | Name       | Starts | E-Rows | A-Rows |   A-Time   | Buffers |  OMem |  1Mem | Used-Mem |
------------------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT     |            |      1 |        |    303 |00:00:00.16 |    3481 |       |       |          |
|   1 |  HASH GROUP BY       |            |      1 |    303 |    303 |00:00:00.16 |    3481 |  1096K|  1096K| 1232K (0)|
|*  2 |   HASH JOIN OUTER    |            |      1 |    792 |   1084 |00:00:00.01 |    3481 |  1206K|  1206K| 1529K (0)|
|   3 |    NESTED LOOPS      |            |      1 |    314 |    314 |00:00:00.01 |     637 |       |       |          |
|   4 |     TABLE ACCESS FULL| DRAG       |      1 |    314 |    314 |00:00:00.01 |       7 |       |       |          |
|*  5 |     INDEX UNIQUE SCAN| PK_CUST_ID |    314 |      1 |    314 |00:00:00.01 |     630 |       |       |          |
|   6 |    TABLE ACCESS FULL | CONT       |      1 |   1181K|   1181K|00:00:00.01 |    2844 |       |       |          |
------------------------------------------------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

   2 - access("CUS"."CUST_ID"="CO"."CUST_ID")
   5 - access("T"."DRAG_ID"="CUS"."CUST_ID")


27 rows selected.  



Oracle Support note associated with "_optimizer_group_by_placement" parameter.
Note.8945586.8 Ext/Pub Bug 8945586 - Wrong results using GROUP BY placement:
Description
Wrong results can occur when using GROUP BY placement where the aggregate column gets pruned from select list.


I’ve even found that “_optimizer_group_by_placement” parameter was mentioned in "Oracle® Fusion Middleware Oracle WebCenter Analytics Installation and Upgrade Guide".

Oracle 11g (11.1.0.6 and above) in default or Oracle Real Application Clusters (RAC) configuration

When running Oracle 11g versions prior to 11.1.0.7.0 the Oracle system parameter _optimizer_group_by_placement must be set to false. This can either be set in the init.ora file of the respective database instances or by by issuing an ALTER SYSTEM command as follows:

SQLPLUS /nolog
CONNECT / AS SYSDBA
ALTER SYSTEM SET "_optimizer_group_by_placement"=false


If you are running affected Oracle version - consider changing this parameter ;-)


Share/Bookmark

0 comments:

Post a Comment