Sample Header Ad - 728x90

How to get cross apply to operate row by row on a view?

7 votes
1 answer
2102 views
We have a view that is optimized for single item queries (200ms no parallelism): select * from OptimizedForSingleObjectIdView e2i where ObjectId = 3374700 It also works on small sets of static ids (~5). select * from OptimizedForSingleObjectIdView e2i where ObjectId in (3374700, 3374710, 3374720, 3374730, 3374740); However if the objects come from an external source, then it generates a slow plan. The execution plan shows that the execution branch for the view part is ignoring the predicate on ObjectId whereas in the original case it uses them to perform index seeks. select v.* from ( select top 1 ObjectId from Objects where ObjectId % 10 = 0 order by ObjectId ) o join OptimizedForSingleObjectIdView v -- (also tried inner loop join) on v.ObjectId = o.ObjectId; We don't wish to invest in "dual" optimizing the view for non-singular cases. Rather, the solution we "seek" is to repetitively call the view once per object **without resorting to an SP**. Most of the time the following solution calls the view row by row. However not this time and not even for just 1 object: select v.* from ( select top 1 ObjectId from Objects where ObjectId % 10 = 0 -- non-trivial predicate order by ObjectId ) o cross apply ( select top 2000000000 * from OptimizedForSingleObjectIdView v_ where ObjectId = o.ObjectId order by v_.SomeField ) v; At one time I thought there was a claim that cross apply was guaranteed for row by row execution when it calls a UDF but this also failed: create function FunctionCallingView(@pObjectId bigint) returns table as return select * from OptimizedForSingleObjectIdView where ObjectId = @pObjectId; select v.* from ( select top 1 ObjectId from Objects where ObjectId % 10 = 0 order by ObjectId ) o cross apply FunctionCallingView(o.ObjectId) v Adding option(force order) did not help--however there are two hash hints in the view already. Temporarily removing them did not help and slows down the single case. Here is a snippet of the estimated plan for the function based slow case. The estimation of 1 row is correct. Far off to the right (not shown) is where there is a seek predicate that does not include that top 1 result. This seems similar to other cases we have where singular probed values from table seeks are not used as seek predicates elsewhere. enter image description here
Asked by crokusek (2110 rep)
Jan 25, 2019, 01:15 AM
Last activity: Jan 26, 2019, 10:12 AM