Hi together I know that issue has been discussed for several times, but we face an insane weird issue on our new SQL Server 2014 SP3 right now, and we just can't get behind.

We switched over from Oracle to MS SQL Server last year. Our mayor DB is quite big, abt. 800GB, big tables, PDM system. +1000 active users. 16 cores, 192 GB Memory, SSD SAN Storage. ESX 6.5

Settings in SQL Server:

  • Create Auto Statistics Enabled
  • optimize for Ad Hoc Queries = true
  • Snapshot Isolation enabled
  • Max Parallel = 4
  • Threshold 50
  • Statistics Updates async in TempDB, normally enabled on our main DBs.

Any how, we have some queries which are handled extremly bad. All leads to the fact that the SQl Server Optimizer creates an execution plan, thinks its okay but on execution it does an inner join (or multiples) with Millions instead of expected 1-2 rows. And of cours millions of logical reads. And I can't get behind whats ongoing in this. These statements ran minutes then.

So basically, we have already 3 databases. All same version and hardware, Prod Test and Dev Environment. My tests can be done pretty easily. All dbs are configured the same and show the same behaviour, but on different queries. Lets say somethimes the test environment does the same statement multipletimes in 47s, the prod db in just a second. Others take seconds in Prod, are instand in Test. What the heck is going on? I always retry the statements muliple times to ensure its cached.

The query is generated by the 4 Tier Application we are using.

1s vs 47s. Both of these DBs share the same SQL Instance. The weird thing is that it might be instant as well if we reset the plan cache. But any how, it slows down again by a complete mess of misscalculation in the optimizer.

Any ideas what could have been wrong here? How to correct single execution plans? How can it be that the sql server doenst learn from that extremly bad statement and to correct it the next time it runs?

I have switched over and rebooted the Test DB with standard trace flags, so no compatibility mode.

Detailed DB settings are: enter image description here

Initially the query runs for 10 minutes... first run. Now it cached instantly as well.


, PFORM t_05 
WHERE ( (  UPPER(t_06.pobject_type)  IN  ( UPPER( 'EN_Item Revision' ) ,  UPPER( 'EN_Item Revision' )  ) 
AND ( ( ( t_07.pvalu_0 = t_02.puid ) 
AND  UPPER(t_02.pname)  IN  ( UPPER( 'V' ) ,  UPPER( 'E9_D' ) ,  UPPER( 'F' ) ,  UPPER( 'E9_to_F' ) ,  UPPER( 'E9_M' )  ) ) 
AND ( ( ( t_08.puid = t_03.rprimary_objectu ) 
AND ( ( t_03.rrelation_typeu = 'gMx8h03uVJFL2B' ) 
AND ( ( t_03.rsecondary_objectu = t_05.puid ) 
AND ( ( t_05.rdata_fileu = t_01.puid ) 
AND ( ( t_01.re9_manufactureru = t_09.puid ) 
AND  ( UPPER(t_09.pe9_id)  =  UPPER( '7300007' )  ) ) ) ) ) ) 
AND ( ( t_08.puid = t_04.rprimary_objectu ) 
AND ( ( t_04.rrelation_typeu = 'gMx8h03uVJFL2B' ) 
AND ( ( t_04.rsecondary_objectu = t_05.puid ) 
AND ( ( t_05.rdata_fileu = t_01.puid ) 
AND  ( UPPER(t_01.phersteller_artikelnr)  =  UPPER( '00100893' )  ) ) ) ) ) ) ) ) AND ( t_06.puid = t_07.puid AND t_07.puid = t_08.puid ) );

but that should be the same like now.

What could it be if the first queries against the DB are lightening fast in the beginning and at later on it slows down by factor 50 as it does a weird execution plan?

With the standard CE in our test environment enabled + a required wipe of the plan cache, the statement above runs instantly on our test environment using the "good plan".

This whole topic, including the test with the older CE in "test", came up due to the fact that if we create a single most likely required index in our production environment, a bad execution plan is created, similar to the bad one above, causing 100% cpu load within three minutes and we had to kill all sessions 3 times during working time now, even as a MS SQL expert created that index in good mood.

The missing index statement was this:

WITH cte_00000000059A53B0_17 AS (
         SELECT t_03.puid AS revPuid
               ,t_04.pdate_released AS revDateRel
               ,t_03.ritems_tagu AS myItem
               ,t_04.puid AS rlsPuid
               ,t_04.pname AS rlsName
               ,PRELEASE_STATUS_LIST t_02
               ,PITEMREVISION t_03
               ,PRELEASESTATUS t_04
         WHERE  (
                                    t_03.ritems_tagu = #PRefbindtag3.puid
                                    AND (t_04.pdate_released <= @P1)
                                AND (t_04.pname = @P2)
                            AND (t_02.pvalu_0 = t_04.puid)
                        AND (t_01.pactive_seq != 0)
                    AND (t_01.puid = t_02.puid AND t_02.puid = t_03.puid)

FROM   cte_00000000059A53B0_17 t_07
           SELECT MAX(t_05.revDateRel) AS rDate
                 ,t_05.myItem AS dtItem
           FROM   cte_00000000059A53B0_17 t_05
           GROUP BY
       ) t_06
               (t_07.myItem = t_06.dtItem)
               AND (t_07.revDateRel = t_06.rDate)
           AND (t_07.rlsName = @P3)
       t_06.rDate DESC


CREATE NONCLUSTERED INDEX [<Name of Missing Index, sysname,>]
ON [dbo].[PRELEASESTATUS] ([pname],[pdate_released])
INCLUDE ([puid])

We need to try the same with a complete DB restart on weekend as we cant see that weird plan in our test environment having this Index enabled. Statistics are a good point, I'll verify if they were all updated. Any ideas on this?


Sadly we dont seem to have a solution yet. What we inestigated the last weeks:

-> we found out that some date / time colums were not migrated correctly from oracle, they had no hour and minutes time stamps so all have been on date and 0:00. The in general high load on our production environment dropped fixing the affected tables. I expect the queries got far less data to compute once the dates have been more atomic.

-> we were pretty optimistic that the highly suggested Index creation, who caused out prod db to stuck, now will be created without any issue on less than 10% Db CPU load.

-> we created it yesterday after working hours so less load on the DB. Within minutes we sould see some active CTE statements having billion, seriously billions of logical reads and they did not shift to resent at all. It seems that it have been always the 4 same process IDs but as said, they crank up some cores to 100% and keep running. We did not do a flush of the whole execution plan cache as we saw the plans they use are new.

-> I could only grab of the active statements and save their execution plan. sadly this should be the expected, not the real execution plan or?



The new index was named: EN_PIPRELEASESTATUS_1

So any how, this Index causes our prod Db to stuck in hefty plans, while the identical test environment runs fine.

ideas? Thanks

Btw: We could drop the indey by


  -- lock table "a" till end of transaction
  SELECT top 1 puid
  -- do some other stuff (including inserting/updating table "a")
  -- release lock


I took a screenshot of some sessions going crazy, that was on the very beginning enter image description here

  • Can you please upload the plans to PasteThePlan.com and replace your screenshots? The screenshots don't contain any of the useful meta data needed to properly diagnose the problem. – George.Palacios Jan 24 at 8:44
  • thx, give me a min and they will be there. – Krautmaster Jan 24 at 8:46
  • I added them in the first post – Krautmaster Jan 24 at 8:52
  • 2
    SELECT DISTINCT after a 9-tables join. The ORMs and their pitfalls ;) – ypercubeᵀᴹ Jan 24 at 8:55
  • 1
    Have you checked that your statistics are up to date? – pacreely Jan 24 at 14:33

Your database settings are DIFFERENT.

Your first plan is produced on database with compatibility level = 120:


And your second plan was produced by using legacy cardinality estimator:


So you can change database compatibility level to 120 on the second database (it will affect all the queries) or use query hint (QUERYTRACEON 2312) (it will affect only this query) as described here: Enable plan-affecting SQL Server query optimizer behavior that can be controlled by different trace flags on a specific-query level

Trace flag 2312 forces the query optimizer to use version 120 (the SQL Server 2014 version) of the cardinality estimator when creating the query plan.

But the main issue that I see in your query is that you used Oracle approach to write it.

You should not use UPPER() unless your database use case sensitive collation.

For example, UPPER(t_06.pobject_type) should be rewritten as simple t_06.pobject_type, when you apply any function to a field you successfully exclude indexes on this field from using them.

SQL Server is NOT Oracle, case-sensitive collations are rarely used.

The second observation regards the number of joins in your query.

When you have more than 5 tables, optimizer has no chance to find good enough plan in small time, so it can stop optimization with the reason StatementOptmEarlyAbortReason="TimeOut" that I found in one of your plans.

Split your query in two, start the first from smallest table (smallest after filtering), PE9_MANUFACTURERSTORAGE t_09 in your case, join non more than 4 tables to it and save the result in temporary table. Then use this temp table to join other tables.


On the last picture that you posted as the answer your db collation IS CASE SENSITIVE. I don't know why did you choose binary collation, it's the fastest collation but if you don't want to use correct case and use UPPER() indexes will not be used. On other side, if you use UPPER() it seems that you don't want this case sensitivity, so why did you choose binary?

  • 1
    Thanks for your detailed answer. Yeah, I am aware of that fact with CardinalityEstimationModelVersion="70" but we tried that on our test system yesterday, its still active but made no difference . We will until this lunch time once nobody is on that test environment. On the statements itself we only have rare influence as they are produced by Siemens. We only request the data by a defined API and they can handle with SQL, no matter if they do good or bad, in most cases I have to deal with the statements they do. – Krautmaster Jan 24 at 9:52
  • I think “please stop to use UPPER” should probably be “please stop using UPPER”? They mean opposite things, so it seems a bit confusing right now... – RBarryYoung Jan 24 at 13:27
  • @RBarryYoung Sorry, my English is bad, I still cannot understand the difference, it seems to me that both the phrases means the same. I wanted to say "You should not use UPPER unless your database use case sensitive collation" – sepupic Jan 24 at 13:32
  • 1
    Yes, I thought that might be the case, just trying to help. Either your last sentence or my suggested change should work. The original phrase is awkward but technically means the reverse of what I think that you intended. – RBarryYoung Jan 24 at 13:37
  • @RBarryYoung Thank you, I'll edit my answer with the "should" phrase :) – sepupic Jan 24 at 13:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.