Sql-server – can anyone help me with this awful query plan

execution-planperformancequery-performancesql serversql-server-2017

The query:

    SELECT Object1.Column1, Object2.Column2 AS Column3, Object2.Column4 AS Column5, 
Object3.Column6, Object3.Column7,Object1.Column8, Object1.Column9, 
Object1.Column10, Object1.Column11, Object1.Column12, Object1.Column13, 
Object1.Column14, Object1.Column15 as Column15, Object1.Column16, 
Object4.Column4 AS Column17, Object4.Column2 AS Column18, Object1.Column19,
 Object1.Column20, Object1.Column21, Object1.Column22, Object1.Column23, 
Object1.Column24, Object1.Column25, Object1.Column26, Object5.Column4, 
Object1.Column27, Object1.Column28, Object1.Column29, Object3.Column30, 
Object3.Column1 as Column31, Object3.Column32 as Column33, Object1.Column34 
as Column34, ? AS Column35 , Object3.Column36 as Column37 
FROM Object6 AS Object1 
INNER JOIN Object7 AS Object3 ON Object1.Column38 = Object3.Column1 
INNER JOIN Object8 AS Object2 ON Object3.Column30 = Object2.Column1 
LEFT JOIN Object9 AS Object4 ON Object1.Column16 = Object4.Column2 
LEFT JOIN Object10 AS Object5 ON Object1.Column9 = Object5.Column2 
WHERE Object2.Column1 <> ? AND Object1.Column8 = ? 
AND ( coalesce(Column16,?)= ? ) 
AND EXISTS (
         SELECT ? 
         FROM Object11 
         WHERE Column39 = ? 
         AND Column30 = Object3.Column30) 
ORDER BY Column7 desc
OFFSET ? ROWS FETCH FIRST ? ROWS ONLY

here is the query plan

I know that I should maybe add an index on this:

Database1.Schema1.Object7.Column30, Database1.Schema1.Object7.Column36, Database1.Schema1.Object7.Column6, Database1.Schema1.Object7.Column32

but one of this columns is a varchar 4000 and it can't be created cause of the big dimension of the field.

I noticed that it takes 25 second only if the rows returned are fewer than the fetch first number

Best Answer

enter image description here

The execution plan accesses Object7 first using a non covering index in Column7 order. It then does some key lookups on that table and nested loops joins against the other tables with the final joined resulting arriving at the TOP operator still ordered by Column7.

Once this has received enough rows to satisfy the OFFSET ... FETCH it can stop requesting any more rows from downstream operators. SQL Server estimates that it will only need to read 2419 rows from the initial index on Object7.Column7 before this point is arrived at.

This estimate is not at all correct. In fact it ends up reading the entirety of Object7 and likely runs out of rows before the OFFSET ... FETCH is satisfied.

The semi join on Object11 reduces the rowcount by almost half but the killer is the join on Object6 and predicate on the same table. Together these reduce the 9,753,116 rows coming out of the semijoin to 2.

You could try spending some time looking at statistics on the tables involved to try and get the cardinality estimates from these joins to be more accurate or alternatively you could add OPTION (USE HINT ('DISABLE_OPTIMIZER_ROWGOAL') ) so the plan is costed without the assumption that it can stop early due to the OFFSET ... FETCH - this will certainly give you a different plan.