Performance Tuning :: Change In Explain Plan Due To Oracle Upgrade
Dec 14, 2010
I need to warn readers that I am not a DBA but am heavily involved in application development. Whatever I know about database tuning is whatever I've managed to pick up via self-learning, and I must admit that the sum total of my knowledge isn't a lot.
Anyway, our "DBAs" recently did an upgrade to our 10g database, going from version 10.2.0.2.0 to 10.2.0.4.0. Immediately after the upgrade, a particular query has started to under-perform. The query itself was not altered in any way during the upgrade.
We have two explain plans for the query, a before and an after plan. The two plans are similar but not identical. The plans are too massive to post here, so I hope the following synopsis of the differences will do.
The 10.2.0.2.0 plan:
shows a HASH GROUP BY
has a TempSpc column in the explain plan
shows a particular table (EMP_HISTORY) as having ~1700 rows
The 10.2.0.4.0 plan:
shows SORT GROUP BY instead of HASH GROUP BY does not show a TempSpc column in the explain plan shows the EMP_HISTORY table as having only 25 rows
Other than these points, no other discernible differences can be noted. I'm wondering what would cause HASH to change to SORT. I'm told that stats are up-to-date.
what privilege is require for a user to execute explain plan? I get below error while try to execute explain plan.
SQL> explain plan for SELECT /*+ FULL(t) */ COUNT(*) FROM "DREAM"."CONSUMER.TAB" t WHERE ROWNUM <= 1000000; explain plan for SELECT /*+ FULL(t) */ COUNT(*) FROM "DREAM"."CONSUMER.TAB" t WHERE ROWNUM <= 1000000 * ERROR at line 1: ORA-01031: insufficient privileges
I am facing a weird situation wherein the explain plan of same sql in SIT and PROD is different.In fact the explain plan is very costly in Prod.Also the DB version of both SIT and PROD is same.
Below is the sql and corresponding explain plan in Prod and SIT respectively.
Query: SELECT seq,CCN,ProcessorPart,root_item,comp_path,Item,comp_item,comp_item_type, lag(comp_item_type,1,'PART') over(PARTITION BY seq ORDER BY lvl)Nxt_comp_item_type,lvl,bom_qty, ROUND(CASE min(abs(bom_qty)) OVER (PARTITION BY seq ORDER BY lvl) WHEN 0 THEN 0 ELSE 1 END * EXP (SUM (LN (nullif(abs(bom_qty),0))) OVER (PARTITION BY seq ORDER BY lvl))) Ulti_qty, 'AMER'
[code]...
The tables referred in above query is small tables containing arnd 10k records.The above tables are partitioned on Region and not indexed.
I am not able to attribute why there is a huge change in Cost between SIT and Prod.Apparently the Job is going for 3-5 hours which used to get completed within 20mins in SIT.
however I was able to identify a poorly performing query that seemed to be maxing out our CPU. I have been trying to understand the Explain Plan. The plan below is from our test system which has considerably less information in the tables than our PROD system.
I can see there are a bunch of table scans at the end which may indicate missing indexes, but I am unclear on whether this is actually a problem as the %CPU seems to be worse for the JOIN near the top of the plan.
------------------------------------------------------------------------------------------------------------------------------- | Id | Operation | Name | Rows | Bytes |TempSpc| Cost (%CPU)| Time | Inst |IN-OUT| ------------------------------------------------------------------------------------------------------------------------------- | 0 | SELECT STATEMENT | | 1870M| 3018G| | 677M (1)|999:59:59 | | | | 1 | SORT ORDER BY | | 1870M| 3018G| 3567G| 677M (1)|999:59:59 | | |
I have two Oracle instances that are setup identically.When I run a query on one of them, it takes around 3 seconds, on the other it takes around 200 seconds.
I have looked at the explain plans, and it has shown me what I think is the problem. On one instance, it does a join on two tables, then runs the other filter/access predicates. On the other instance it runs the filter/access predicated first, then does the expensice join. The one that does the join first is the one that takes around 200 seconds. How to tell Oracle to make this join after runnning the other predicates?
when i runnung the explain plan syntax , show error : running --- SELECT * FROM TABLE(dbms_xplan.display) ; ERROR: an uncaught error in function display has happened; please contact Oracle support Please provide also a DMP file of the used plan table PLAN_TABLE ORA-00904: "OTHER_TAG": 無效的 ID
since the optimizer (during explain plan) assumes all bind variable to be of varchar type, while checking plan for SQL statement using bind variable of numeric and date type shall we convert (typecast) it as following?
variable n_sal number variable dt_joining date exec n_sal:= 1000 exec dt_joining := '12-dec-2005' select first_name from emp_data where sal=to_number(n_sal) and joining=to_date(dt_joining);
I have created an non unique index lk_fein on lookup_fein( code,map_id,trash). When I check the explain plan it does a full table scan on lookup_fein. if I force it to use index by it does and the cost also decreases.
Session 1 create table tab1 as select * from dba_objects where object_id is not null; alter session set events '10046 trace name context forever, level 12'; declare x number; begin for i in 1..4 loop
[code]....
Session 2
after "starting" the above pl/sql block from Session 1, I keep on querying tab2 from Session 2 And as soon as 2 records are inserted in tab2, I create index from Session 2
select * from tab2; select * from tab2; select * from tab2; N ---------- 1 2 create index i on tab1(object_id);
As I have tested from a single session (just before this test) such index is used for the sql statement
select count(1) into x from tab1 where object_id=2331;
However when I checked the trace file I am not geeting results as expected
I am expecting 4 execution plans - 2 FTS and 2 Index Access scans and for this I am issuing following command
SELECT COUNT(1) FROM TAB1 WHERE OBJECT_ID=2331 call count cpu elapsed disk query current rows ------- ------ -------- ---------- ---------- ---------- ---------- ---------- Parse 1 0.00 0.00 0 1 0 0 Execute 4 0.00 0.00 0 2 0 0
[code]....
1) Why I am unable to see 4 execution plans - 2 with FTS and 2 with Index access when I mentioned 'aggregate=no'?
2) Whether the index i will be used for last 2 iterations after first 2 iterations of FTS?
If answer to above question 2) is 'No'
By which method I can force an ongoing sql statement in loop to take different execution path? Of course I can't hard parse sql in 'that' current session Will flushing Shared pool work in above case?
The prod stats has been implemented in development. The stats has been gathered 2 months back on dev while in production the stats has been gathered 2 weeks back.
My question shouldn't the high volume of data causes changes in plan in both the environment? My thinking is that plan can be different as the high volume of data are changing in prod it may lead to a different plan.
Why the query is behaving differently with the different database.(execution plan)
Whatever the production database is having same database instance replicated to a new schema. I tried both the queries running on both environment.In prod the index has been used but in newdev it is not. This case existing primary key index were not been used.
How can i check the avg time taken by an execution plan. Actually i have a very big query and it changes its execution plan very often, we would like to lock the best execution plan and to find it , i would like to know the Average Execution Time the query takes when it runs using different different execution plans.
I have queries on the execution plan of a sql statement
Following is the example
create table t1 as select s1.nextval id,a.* from dba_objects a; create table t2 as select s2.nextval id,a.* from dba_objects a; insert into t1 select s1.nextval id,a.* from dba_objects a; insert into t1 select s1.nextval id,a.* from dba_objects a; insert into t2 select s2.nextval id,a.* from dba_objects a; insert into t2 select s2.nextval id,a.* from dba_objects a; insert into t2 select s2.nextval id,a.* from dba_objects a; commit;
create index i1 on t1(id); create index i2 on t2(id); create index i11 on t1(object_type);
(1) First index on object_type is accessed to get rowids - t1.object_type='VIEW' (2) Then the filter on owner is applied - t1.owner='SYS' (3) Then the table T1 is accessed to fetch data from the rowids returned by the index I11 and filer application - TABLE ACCESS BY INDEX ROWID
Though I am unable to understand how filter can be applied to the rowids retrieved from index, we can see from the plan below that The rows accessed have reduced from 8550 to 1221 before we access the table...Thus filter "t1.owner='SYS'" is applied in between. Right?
another question is
Case 1 - do we retrieve a rowid from index for a given value, then retrieve required values from table for that rowid Thus row at a time in both ... in loop OR Case 2 - we first fetch all rowids from index and then retrieve values from table one row at a time from the collection of rowids fetched?
Suppose Case 1 is what is happening then can we say, both the steps mentioned by IDS 2,3 in plan below are executed exactly equal number of times and the filter "t1.owner='SYS'" is applied at some later stage? Of course in this case the values in ROWS stand misleading then
select * from t1,t2 where t1.id = t2.id and t1.object_type='VIEW' and t1.owner='SYS';
Execution Plan ---------------------------------------------------------- Plan hash value: 26873579 ------------------------------------------------------------------------------------- | Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time | ------------------------------------------------------------------------------------- | 0 | SELECT STATEMENT | | 1221 | 233K| 915 (1)| 00:00:11 | |* 1 | HASH JOIN | | 1221 | 233K| 915 (1)| 00:00:11 | |* 2 | TABLE ACCESS BY INDEX ROWID| T1 | 1221 | 116K| 381 (1)| 00:00:05 | |* 3 | INDEX RANGE SCAN | I11 | 8550 | | 24 (0)| 00:00:01 | | 4 | TABLE ACCESS FULL | T2 | 161K| 15M| 533 (1)| 00:00:07 | ------------------------------------------------------------------------------------- Predicate Information (identified by operation id): --------------------------------------------------- 1 - access("T1"."ID"="T2"."ID") 2 - filter("T1"."OWNER"='SYS') 3 - access("T1"."OBJECT_TYPE"='VIEW')
- Both of these databases run on different hardware (A is a VM, B is on a physical host)
- The 20 tables in A and B have exactly same number of rows and after preparing the data, the schemas were analysed using the same DBMS_STATS parameters
Despite this, the execution plans appear to be quite different for the same queries between A and B
I imagine there is something outside of the Oracle table rowcounts, table stats, column stats, index stats that's resulting in the different execution plans.
refere to below 2 queries and their execution plans:
First Query INSERT INTO temp_vendor(vendor_record_seq_no,checksum,rownumber,transaction_type,iu_flag) SELECT /*+ USE_NL ( vd1 ,vd2 ,vd3 ) leading ( vd1 ,vd2 ,vd3 , tvd) */ vd1.vendor_record_seq_no, tvr.checksum, tvr.rownumber, tvr.transaction_type, 'U' FROM vendor_data vd1,
[code]...
Second Query SELECT vd1.vendor_record_seq_no, tvr.checksum, tvr.rownumber, tvr.transaction_type, 'U' FROM ( select * from vendor_data vd1 where vd1.study_seq_no = 99903 AND vd1.control_column_seq_no = 435361232
[code]...
Both are to achieve same output but written in different ways. CAn I get same exectuion plan from 1st query as there is for 2nd using hints
We have recently upgraded application (from Oracle Applications 11.5.9 to 12.1.3) and database (from 9.2.0.5.0 to 11.2.0.3.0).Since we are confronting to performances issues, i try to analyse some queries which Explains plans seems strange (in my opinion).Studying one of them i discover the next case (which according to my logic, i can't explain): --
Just bind variable --select *from MTL_MATERIAL_TRANSACTIONS mmtwhere 1 = 1and mmt.INVENTORY_ITEM_ID = :p1and mmt.ORGANIZATION_ID = :p2and mmt.TRANSACTION_REFERENCE = :p3--and mmt.SUBINVENTORY_CODE = :p4 PlanSELECT STATEMENT ALL_ROWS Cost: 5 Bytes: 361 Cardinality: 1 2 TABLE ACCESS BY INDEX ROWID TABLE INV.MTL_MATERIAL_TRANSACTIONS Cost: 5 Bytes: 361 Cardinality: 1 1 INDEX RANGE SCAN INDEX XXSPE.XXSPE_MTL_MATERIAL_TRANSAC_N99 Cost: 3 Cardinality: 2-- Nvl on bind variable --select *from MTL_MATERIAL_TRANSACTIONS mmtwhere 1 = 1and mmt.INVENTORY_ITEM_ID = :p1and
One of our clients is using Rule Based Optimizer on Oracle 10.2.0.3.0
2-3 weeks backs, during performance issue in one of the sql queries, one of our team members executed tuning adviser for it, created SQL profile and the subsequent execution of the SQL did not took much time (less I/O). Now it took hardly a minute to execute
When this happened I checked that the SQL profile forced that particular query to use CBO (say plan_hash_value is PHV1 here). Yesterday the same query again took 15-20 minutes for execution. I checked that even for this execution the query used the same SQL profile but "this time" with different plan_hash_value - say PHV2.
Today again the query executed in less than a minute and used the plan_hash_value as PHV1.
select distinct plan_hash_value,timestamp from dba_hist_sql_plan where sql_id='mysqlid' order by 1,2;
I confirmed from awrsqrpt as well that different plans were used for different plan_hash_values and every time same SQL profile was used
SQL> select name,CATEGORY,SIGNATURE,CREATED,LAST_MODIFIED,TYPE,STATUS,FORCE_MATCHING from dba_sql_profiles;
NAME CATEGORY SIGNATURE CREATED LAST_MODIFIED TYPE STATUS FOR ------------------------------ ------------------------------ ---------- -------------------- -------------------- --------- -------- --- SYS_SQLPROF_015ffffcc3e1c5b000 DEFAULT 1.5512E+19 20-feb-2013 16:30:48 20-feb-2013 16:30:48 MANUAL ENABLED NO
I am unable to understand how execution plan and thus plan_hash_value is changing for the same SQL Profile. I read that SQL Profile (unlike stored outline) keeps up with increasing data volume and may not keep up with changing data distribution.
I checked that values for 4 bind variables out of 81 are different for execution between today and yesterdays' run(queried v$sql_bind_capture based on last_captured)
My questions are 1) does the different plan_hash_values with different execution plans for query using same SQL profile mean the query was hard parsed multiple times and still used the same SQL profile? 2) If that is the case why I never saw child_number = 1 in any of the views for the same sql_id. I tried it repeatedly over last 2 weeks and always found child_number=0 in v$sql (also loaded_versions=1) 3) Does the different values of bind variable are causing this flip-flop of the plans? How can I conclude this?
I have 2 plans with 2 different plan_hash_values. I know which would be better. How can I force the sql to use better plan in the two in this case where I am using Rule Based Optimizer and have SQL profile created If this is not possible then how can I create stored outline from the existing plan (not waiting for subsequent execution to take place).
The types of query I refer to in the title are of this pseudo-code ilk:
select t.column_value from table1 o, xmltable('for $co in $data where $co/path1=$bind1 and $co/path2=$bind2 passing o.field as "data", :b1 as "bind1", :b2 as "bind2") t where o.field = :b3
They're querying a table with a (binary) xmltype with a path/domain index over this column.As those who have had the (mis)fortune to run into these will know, the queries are extensively rewritten under the covers to access to xml via the paths supplied.
getting a baseline to work with queries like this? I was suspicious because whilst I can hint it to pick a certain access path first (leading()), the plan hashes remain the same.
I'm not sure, however, if I'm simply "doing it wrong" or it is just not possible with the level of recursive rewriting going on.NB: I consider myself reasonably competent in applying baselines to "traditional" queries...
1. In previous version of Toad (9.7.2) during executing a statement I was able to click on 'Explain plan current statement'. In this version of Toad (10.6.0. 42) is not a possible to do it.In the Toad options I've filled the sign: "use a separate connection when Toad itself is generating transactions"
2. During executing a statement I still see 'clock' cursor.Is it a possible to disable to see it?
I have an APP that truncates tables and loads data, which in turn makes the stats stale. I ran the query advisor (see attachment) and of course it ecommends running stats or accept a profile.I really don't want to do that as it may cause a load on my DB.
In turn, I would like to consider having my APP team change the query to pass a hint to use the best query plan.syntax to pass the hint to emulate good attached plan? Or is this a bad way to proceed?
select /* INDEX FAST FULL SCAN PK_PLACEMENT_REQUEST_QUEUE */ sum(lastshares) as "ROSEN" from nyeo.fix_exec_reports fer, nyeo.placement_request_queue q, nyeo.nyeo_block_control bc where fer.clordid = q.sequence_number and q.blockid = bc.blockid and upper(bc.deskname) like '%ROSEN%'
Elapsed times include waiting on following events:
Event waited on Times Max. Wait Total Waited ---------------------------------------- Waited ---------- ------------ SQL*Net message to client 1 0.00 0.00 db file sequential read 85704 0.31 460.55 latch free 1 0.00 0.00 SQL*Net message from client 1 14.98 14.98
[code]...
Why the elasped time changed when data and plan hasn't changed at all? Also why the plan has different stats for round 1 and 2 on db1 and db2?
I ran it 2 times each round each database so hard parsing shall not be issue.Also why the number of rows accessed are different in db1,db2 and db3,db4 especially for step1 when count of crt.qtn_cun_id is similar?
In fact when the query was taking long I was the only user on the system Also I used hard coded value (no bind variables at all)
I checked num_rows, distinct keys as well which are quite similar across all 4 databases Also no stats where gather during the query execution
We are currently migrated our database to Oracle 11g i.e 11.2.0.3 (from 10.2.0.3). Now we are facing application slowness on 11g, we found the issue is with missing histogram or some unnecessary histograms in place. Given below are my details.
When we were in 10g, we were having schema level gather stats job with 'METHOD_OPT' as 'AUTO' running each weekend and gathering only STALE stats. And we had faced some performance issue due to optimizer following wrong path of execution due to some missing histograms and some extras too.So we had done skew ness(for each column in table in database) analysis on DB level using 'WIDTH_BUCKET' function manually on each column,and we deleted no of unnecessary histograms and created some on 10g DB. And updated the default schema level gather stats job with 'METHOD_OPT' as 'FOR ALL COLUMN SIZE REPEAT'. And we were getting stable environment.
Now we just migrated and its done by our techops(Oracle) team, i am confused how the histograms has been changed/modified for columns during migration, causing headache now. So now below are my plans and fears associated with them.
1. If i will make the histograms exactly same in both the environments, i.e. will delete histograms which are present in new DB(11g) but were not in old DB(10g).and creating those which were there in old DB(10g) but now not in new DB(11g). My question is , will it be safe to do this or it may impact negatively with 11g optimizers features in place?
2. Or should i try easier option once by making 'METHOD_OPT' for the schema level stats job as 'AUTO' on 11g and try once, as because i believe 11g is more stable and optimizer is equipped with more intelligence in case of stats gathering using AUTO?
I have used the above to get a copy of schema stats and gather new stats for specific tables into a STATS TABLE in my personal schema. What I want to do now is use this stats table to generate plans for queries where I believe stats are off. Is it even possible? To be clear, I do not want to import stats because this replaces the stats currently there. I just want to point the CBO to my stats table for generating plans.
there was a session parameter I could set to tell oracle to use my stats table when generating plans, or an explain plan clause I could use or a DBMS_XPLAN paramter I could provide that would tell these tools to use my stats table when generating a plan, or even some way to tell autotrace. But I have found none of this.
Performance issues with the below mentioned sql.After gone through execution plan we have found out the reason but we couldn't able to change the execution plan the way we want.
If we could able to join
HRMGR.HR_EXPANDED_BOOK table with MISBOMGR.ibm_client_mgr7_empid, MISBOMGR.ibm_client_mgr6_empid at earlier stage means before HRMGR.HR_EMP_STATUS_LOOKUP then my issue will be solved but somehow optimizer is not considering that path. Even i have added push_subq hint which will push sub queries to execute at earlier stage but no use. Why push_subq hint is not working in this scenario and what can be the other alternative to change the driving path.
Query :-
select /*+ push_subq */CEMP.EMP_ID, CEMP.EMP_STATUS_CD, EMP_STATUS_DESC, MGR_6_EMP_ID, MGR_7_EMP_ID FROM [code]........
I observed a strange behaviour of a query after using stored outline on it. I created a stored outline for a query in one database before creating the outline i had set 2 session level parameters optimizer_index_caching and optimizer_cost adjust. i then took an export of the stored outline and then imported into another db. but here the plan of the query seems to be different from the database from which the outline was taken. why is this change? My version of oracle is 10.2.0.5.0
After upgrade 10g to 11g, the below sql is not working. I have issue with connect by, if we use it with subquery it will hang.
select item_code from bom_list_pos where ln_id in (select ln_id from bom_list_nodes start with ln_id IN (select ln_id from bom_used_work_pack where rownum =1) connect by prior ln_id = parent_ln_id)
I ran 10g, able to get it less than minute, but 11g hang. below is explain plan.
