6

I'm troubleshooting an issue related to SQL Server (Azure SQL Database technically) occasionally choosing a bad execution plan, presumably due to skewed stats. sp_updatestats fixes it every time, until a few hours or days later when a bad plan gets cached again.

Looking at the "bad" plan, I noticed something that strikes me as odd: there is a Clustered Index Seek on a table that currently has about 1.7 million rows. The "Estimated Number of Rows" for this operation is about 1200, which is definitely in line with the average row count I would expect from that operation in this case, but the "Actual Number of Rows" is in excess of 60 million! Following the fat line from this leaf node, various downstream operations such as joins and sorts are being performed on all 60 million, causing excessive slowness, spills to tempdb, and other badness.

I must be misunderstanding what a Clustered Index Seek actually does, because I wouldn't think it's possible for it to "output" more rows than are in the underlying table. What could cause this? And better yet, any pointers on how to fix it?

[Bonus points for including something like "sp_updatestats fixes it every time but can't figure out how to fix it permanently? Go read this article." This has been a general problem for us on a few different fronts lately.]

6

The Seek returns more rows because it is on the inner (bottom) side of a Nested Loop. Every row returned by the outer operation results in a new Seek operation. So you're not getting 60m rows from a single Seek, but from over 9000 of them (number of executions).

Also of note: when looking at estimations, the total number of rows estimated will be Estimated Number of Executions multiplied by Estimated Number of Rows

  • Ah, that makes perfect sense. 60 million / 9000 = a much more reasonable number. :) – Todd Menier Dec 12 '18 at 16:49
  • Thanks. I'm accepting this since it directly answers the "why", but I sure I wish knew where to go from here. I'm joining 9 tables, all of which look to be indexed properly. The plan that goes off the rails is (to simplify) joining tables A and B first, resulting in 60 million rows, and later joining C, which brings it down to zero. I basically want to ensure C is considered earlier, preferably without "hacks" like index hints or forced execution plans. – Todd Menier Dec 12 '18 at 18:41
  • @ToddMenier That might be worth asking as a separate question then. There is a large community here of query tuners who will love to help. Please note that you will want to share the view definition and probably a "good" plan as well. – Forrest Dec 12 '18 at 18:53
  • 1
    @ToddMenier - you need to see if you can get accurate estimates for SELECT * FROM Buyer B JOIN CatalogBuyer CB ON CB.CompanyID = B.BuyerID JOIN Party P ON P.CompanyID = CB.CompanyID WHERE B.SellerID = 10424 AND P.Type IN (2,3) the plan estimates 58 and has actual 9,090 leading into that nested loops join you point out. The most reliable method of getting accurate estimates would be to materialize that result into a temp table and join onto that instead though this would mean you rewriting the query to not use the view. Or try updating stats or creating filtered/multi column stats. – Martin Smith Dec 12 '18 at 19:13

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

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