Plan Shaping and Cardinality Miss Estimate on row_number over partition by

Recalling talks with Tim Hall I wrote in my last post. He mentionend Jonathan Lewis telling about plan shaping. Every now and then there comes a need to tell the optimizer where the query execution should be starting. Write your query from part in order and use ordered hint or most often leading hint will be enough. Hitting the need for such plan shaping comes when the optimizer sees inline views in a query returning only one row and actually there are more in execution time. If there are several missleading one liners and a cartesian join between those a quite simple query can consume significant query time. 12.1.0.2 database has a anoying bug involving top-n queries with analytic partition by part. Luckily a patch for the 21971099 bug is available for some environments. Here is a simplified test getting cardinality one. Imagine having a couple of such in your bigger query.

create table s as (
select round(level/2) n 
     , level n2 
 from dual connect by level < 10e4
)
;

select /*+gather_plan_statistics*/ n
     , n2
  from (
   select n
        , n2
        , row_number()over(partition by n order by n2 desc) rn
     from s)
 where rn <= 1
;

select * 
  from table(dbms_xplan.display_cursor(format=>'iostats last'))
;

Plan hash value: 2407482549
 
-------------------------------------------------------------------------------------------
| Id  | Operation                | Name | Starts | E-Rows | A-Rows |   A-Time   | Buffers |
-------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT         |      |      1 |        |  50000 |00:00:00.12 |     208 |
|*  1 |  VIEW                    |      |      1 |      1 |  50000 |00:00:00.12 |     208 |
|*  2 |   WINDOW SORT PUSHED RANK|      |      1 |  99999 |  50000 |00:00:00.11 |     208 |
|   3 |    TABLE ACCESS FULL     | S    |      1 |  99999 |  99999 |00:00:00.02 |     208 |
-------------------------------------------------------------------------------------------
 
Predicate Information (identified by operation id):
---------------------------------------------------
 
   1 - filter("RN"<=1)
   2 - filter(ROW_NUMBER() OVER ( PARTITION BY "N" ORDER BY 
              INTERNAL_FUNCTION("N2") DESC )<=1)

While waiting for the patching to happen, an alternative to bypass the problem is not to use row_number analytic function. By using rank and order by part that won't return competing winners, we get another cardinality estimate. It is as much wrong as the earlier one, but most likely will help with the cartesian join problem.

select /*+gather_plan_statistics*/ n
     , n2
  from (
   select n
        , n2
        , rank()over(partition by n order by n2 desc, rowid) rn
     from s)
 where rn <= 1
;

select * 
  from table(dbms_xplan.display_cursor(format=>'iostats last'))
;

Plan hash value: 2407482549
 
-------------------------------------------------------------------------------------------
| Id  | Operation                | Name | Starts | E-Rows | A-Rows |   A-Time   | Buffers |
-------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT         |      |      1 |        |  50000 |00:00:00.13 |     208 |
|*  1 |  VIEW                    |      |      1 |  99999 |  50000 |00:00:00.13 |     208 |
|*  2 |   WINDOW SORT PUSHED RANK|      |      1 |  99999 |  50000 |00:00:00.11 |     208 |
|   3 |    TABLE ACCESS FULL     | S    |      1 |  99999 |  99999 |00:00:00.02 |     208 |
-------------------------------------------------------------------------------------------
 
Predicate Information (identified by operation id):
---------------------------------------------------
 
   1 - filter("RN"<=1)
   2 - filter(RANK() OVER ( PARTITION BY "N" ORDER BY INTERNAL_FUNCTION("N2") DESC 
              ,ROWID)<=1)

Trip to BGOUG Conference Pravets 10.-13.11.2016

It was a day in this autumn. I had put a marker to my calendar. Bulgarian Oracle User Group conference 2016 call for papers deadline was that day. Toon Koppelaars will be there. I could talk about ideas behind SQL Assertion implementation. Vote here if you have not yet voted yet. 5000 votes already. I did not know how could I fit the trip to my calendar. I did not know it even then but decided to submit. My presentation "Hub Insert ORA-00001" got accepted. And after a while I started booking flights to Sofia. There was a possibility to choose such flight that I did not have to start my journey too early in the morning and I would get in time to reach the speakers dinner.

While preparing my slides I noticed that my abstract I promised to talk about parallelism. Concurrency issues was the ones I was prepared for. Well just need to add content to the presentation. Parallelism on top of SQL calls cause concurrency issues and parallelism under a SQL clause should speed up the execution. Had to figure out a way to fit a parallel SQL execution example to talk about in some minutes. This is the time when you learn and find out new stuff. While gathering information you learn new stuff. This time one thing to unlearn from 11G days was how 12c parallel dml is able to populate same segments from different processes. Further reading in a oracle optimizer blog post.

The morning of leaving from home came. My first flight to München will be delayed. The latter flight had to be moved to evening flight. I will miss the speakers dinner. Arriving to Sofia 10pm. If I had arrived with my planned flight, conference organizer transportation would have been available. Had to take a taxi. I choose the yellow official cab company next to the station. Not a pleasant journey to Pravets. Taxi driver stopping three times. Once on a dark motorway side about should I pay 100EUR to the driver that he would continue the trip. I gave some money and the rest of the 60km journey continued. Meter was not running anymore. Arrived to the hotel. Somehow I was not sleepy even thou travelling 15 hours already. Went for an one hour walk around the hotel.

Morning walk around a lake wearing short pants. While I left home there was five centimeters snow and -8 celsius cold. Seminar registration, breakfast and on with the show. I got to see good talks. Julian Dontcheff tweaking his laptop virtual machine database super_fast "with a parameter". Toon explaining thickdb approach. Flame graphs explained. Network speed and layers influence performance. Somethings about Oracle in the Cloud. Pluggable databases used in development cycle and also what will they brake. Surprisingly not so much fuss about in-memory option. Full rooms for some sessions. My session had plenty of space. Thanks for those who attended. Hopefully you got something out of it.

In addition to the talks I heard one major reason to attend these kind of seminars is to meet people and get to talk about issues. This conference had a well planned schedule for that. At least 15 minutes pause in between each session. Here are some.

Even thou I did not get to see OBIEE presentations by Gianni Ceresa or Christian Berg it was nice meeting you.

I gave some minor feedback to Martin Widlake about hist talks how Oracle works animations.

Michal Šimoník "join talk" got a slide about qube join.

Straight from Nigel Bayliss, Optimizer Product Manager I got encouragement to my suspicion that my struggling with OLTP system parse times will have some other issues also than the known 12c adaptivity problems. Seems like we found something just yesterday a day after the seminar. Maybe a place to another blog post.

I did not get to talk with Toon about SQL assertions. But at the Sunday breakfast changed some words with Bryn Llewellyn. The vote page is not promising assertion word to be implemented. Will there be some more weight to parsing time? And as those men are influencing the implementation will it be supported at first phases behind the scenes of thickdb approach. We will see. Hopefully sooner than later. So vote.

There was talks also with Neil Chandler and many others.

After the presentations and before dinner time I had also time to go to a Finnish Sauna and watch ice hockey. Tappara winning two matches. The flight home was more enjoyable than going there as I had got business class tickets at the same price than economy tickets would have cost. I was quite relaxed. I read almost a full book by Miika Nousiainen that I purchased for the trip.

On a way home nice talks with Tim Hall and Gianni while waiting at airport.

Thank you Milena Gerova and BGOUG staff for letting me to be a speaker in your event. This was the first time I did a presentation in foreign user group other than Ougf.