0

I've got an issue with estimated cost and join predicate :

SELECT c.Id FROM Table_A a
LEFT JOIN Table_B b ON b.Id_A = a.Id
LEFT JOIN Table_C c ON c.Id_B = b.Id
WHERE a.Id = 2500
OPTION (RECOMPILE)

Table_A has 50k rows, Table_B is empty, Table_C has 2M rows The PK and FK indexes exists and the stats are up to date.

But for some reason, SQL Server use a Clustred Index Scan on Table_C. https://www.brentozar.com/pastetheplan/?id=r1G_LnpeU

This is a sample my problem, the prod tables are fare bigger and use too much estimated grant memory because of this issue.

Using a FORCESEEK or change the join to LEFT JOIN Table_C c ON c.Id_B = b.Id AND c.Id_B IS NOT NULL resolve the issue but it's a query (mutliple queries in fact) generate by Entity Framework so I don't have so much control on it.

Is there a way to seek Table_C without changing the query here ?

0

1 Answer 1

1

I was finally able to replicate your issue due to you noting that

LEFT JOIN Table_C c ON c.Id_B = b.Id AND c.Id_B IS NOT NULL 

improved the query.

DDL & DML

CREATE TABLE Table_A (id int not null,
                     constraint ClusteredIndex_A primary key (id))

INSERT INTO dbo.Table_A(id)
SELECT top(50000) ROW_NUMBER() OVER(ORDER BY (SELECT NULL))
FROM MASTER..spt_values spt1
CROSS APPLY MASTER..spt_values spt2;


CREATE TABLE Table_B (id int not null,
                      Id_A int 
                     constraint ClusteredIndex_B primary key (id))


-- constraint FK_C_B FOREIGN KEY (id_B) REFERENCES Table_B(id)
CREATE TABLE Table_C (id int not null,
                      Id_B int 
                     constraint ClusteredIndex_C primary key (id),
                     constraint FK_C_B FOREIGN KEY (id_B) REFERENCES Table_B(id))
INSERT INTO dbo.Table_C(id,Id_B)
SELECT top(2000000) ROW_NUMBER() OVER(ORDER BY (SELECT NULL)),
NULL
FROM MASTER..spt_values spt1
CROSS APPLY MASTER..spt_values spt2;

CREATE INDEX [NonClusteredIndex-B-A] on dbo.Table_B(id_A);
CREATE INDEX [NonClusteredIndex-C-B] on dbo.Table_C(id_B);

And as a result, your query is going slow due to all the NULL values in the Table_C table.

SELECT c.Id FROM Table_A a 
LEFT JOIN Table_B b ON b.Id_A = a.Id 
LEFT  JOIN Table_C c ON c.Id_B = b.Id
WHERE a.Id = 2500 OPTION (RECOMPILE);

Resulting to a starting point close to yours:

enter image description here

The easiest & best solution here would be changing the query as you pointed out.

SELECT c.id
FROM Table_A a 
LEFT JOIN Table_B b ON b.Id_A = a.Id 
LEFT  JOIN Table_C c ON c.Id_B = b.Id and c.Id_B is not null
WHERE a.Id = 2500;

A possible but far from ideal solution would be adding OPTION(loop join) with a plan guide

exec sp_create_plan_guide   
@name = N'Guide_1',  
@stmt = N'SELECT c.id
FROM Table_A a 
LEFT JOIN Table_B b ON b.Id_A = a.Id 
LEFT JOIN Table_C c ON c.Id_B = b.Id 
WHERE a.Id = 2500;',  
@type = N'SQL',  
@module_or_batch = NULL,
@params = NULL,  
@hints = N'OPTION (LOOP JOIN)';  

enter image description here

But the issue with high estimates on dbo.Table_C persists & plan guides are a last resort (if even).

Im gonna leave this up so other people can test & replicate your problem using DDL & DML for now. Maybe statistics or indexing could prove a solution that I was unable to find.

Edit

option(FAST 1) produces single row estimates & low cost (for this single example)

SELECT c.Id FROM Table_A a 
LEFT JOIN Table_B b ON b.Id_A = a.Id 
LEFT  JOIN Table_C c ON c.Id_B = b.Id
WHERE a.Id = 2500 OPTION (RECOMPILE,FAST 1);

enter image description here

Again not ideal since bigger resultsets will be impacted and other issues might arise.

What also works is inserting one row in dbo.Table_B

INSERT INTO dbo.Table_B(id,Id_A)
VALUES(1,NULL);

& update the stats UPDATE STATISTICS dbo.Table_B

So sql server does not have to join table_C to an empty table_B

enter image description here

Or insert -1.

INSERT INTO dbo.Table_B(id,Id_A)
VALUES(-1,NULL);

What you then could do, is generate the statistics with the 1 row, and create these on the empty dbo.Table_B.

By right clicking the database --> tasks --> generate scripts

enter image description here

Choose table_B

enter image description here

On advanced scripting options select script out statistics & histograms:

enter image description here

Save to new query window, next, next, finish.

enter image description here

And for me I got these statistics:

/****** Object:  Statistic [ClusteredIndex_B]    Script Date: 17/01/2020 10:58:36 ******/
UPDATE STATISTICS [dbo].[Table_B]([ClusteredIndex_B]) WITH STATS_STREAM = 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
GO
/****** Object:  Statistic [NonClusteredIndex-B-A]    Script Date: 17/01/2020 10:58:36 ******/
UPDATE STATISTICS [dbo].[Table_B]([NonClusteredIndex-B-A]) WITH STATS_STREAM = 0x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
GO

Deleted all data:

DELETE FROM dbo.Table_B

Reran the two update statistics with stat streams.

To get a better performing execution plan with correct estimates & results

3
  • thanks for the answer, and for the DDL & DML. I will try the plan guide. I continue searching and I suppose the problem is within the statistics : if I have only 1 match between table_C and table_B, we have the same issue except with a UPDATE STATISTICS Table_C WITH FULLSCAN. I suspect SQL Server to always do a SCAN and a MERGE JOIN when the sample in the statistics is empty, but I didn't find any explanation over it.
    – Zobix
    Jan 17, 2020 at 8:43
  • @Zobix Happy to help. Do watch out with the plan guide. The one posted forces nested loop joins for all executions of the queries. It is also something not easily spotted, document the change if you make any and test the results thoroughly. Jan 17, 2020 at 9:23
  • 1
    A big thanks : I just tried your edit and it works well. Even without the streams stats : Doing UPDATE STATISTICS [dbo].[Table_B]([ClusteredIndex_B]) WITH ROWCOUNT = 2, PAGECOUNT = 1 AND UPDATE STATISTICS [dbo].[Table_B]([NonClusteredIndex-B-A]) WITH ROWCOUNT = 1, PAGECOUNT = 1 solve the problem.
    – Zobix
    Jan 17, 2020 at 10:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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