Nested Loop in Where Statement killing performance - sql

Question

I am having serious performance issues when using a nested loop in a WHERE clause.
When I run the below code as is, it takes several minutes. The trick is I'm using the WHERE clause to pull ALL data if the report_id is NULL, but only certain report_id's if I set them in the parameter string.
The function [fn_Parse_List] turns a VARCHAR string such as '123,456,789' into a table where each row is each number in integer form, which is then used in the IN clause.
When I run the code below with report_id = '456' (the dashed out portion), the code takes seconds, but passing the temporary table and using the SELECT statement in the WHERE clause kills it.
alter procedure dbo.p_revenue
(#report_id varchar(max) = NULL)
as
select cast(value as int) Report_ID
into #report_ID_Temp
from [fn_Parse_List] (#report_id)
SELECT *
FROM BIGTABLE
where #report_id is null
or a.report_id in (select Report_ID from #report_ID_Temp)
--Where #report_id is null or a.report_id in (456)
exec p_revenue #report_id = '456'
Is there a way to optimize this? I tried a JOIN with the table #report_ID_Temp, but it still takes just as long and doesn't work when the report_id is NULL.

Answer 1

You're breaking three different rules.
If you want two query plans, you need two queries: OR does not give you two query plans. IF does.
If you have a temporary table, make sure it has a primary key and any appropriate indexes. In your case, you need an ALTER TABLE statement to add the primary key clustered index. Or you can CREATE TABLE to declare the structure in the first place.
If you think fn_Parse_List is a good idea, you haven't read enough Sommarskog

Answer 2

If I were to write the Stored Procedure for your case, I would use a Table Valued Parameter (TVP) instead of passing multiple values as a comma-seperated string.
Something like the following:
-- Create a type for the TVP
CREATE TYPE REPORT_IDS_PAR AS TABLE(
report_id INT
);
GO
-- Use the TVP type instead of VARCHAR
CREATE PROCEDURE dbo.revenue
#report_ids REPORT_IDS_PAR READONLY
AS
BEGIN
SET NOCOUNT ON;
IF NOT EXISTS(SELECT 1 FROM #report_ids)
SELECT
*
FROM
BIGTABLE;
ELSE
SELECT
*
FROM
#report_ids AS ids
INNER JOIN BIGTABLE AS bt ON
bt.report_id=ids.report_id;
-- OPTION(RECOMPILE) -- see remark below
END
GO
-- Execute the Stored Procedure
DECLARE #ids REPORT_IDS_PAR;
-- Empty table for all rows:
EXEC dbo.revenue #ids;
-- Specific report_id's for specific rows:
INSERT INTO #ids(report_id)VALUES(123),(456),(789);
EXEC dbo.revenue #ids;
GO
If you run this procedure with a TVP with a lot of rows or a wildly varying number of rows, I suggest you add the option OPTION(RECOMPILE) to the query.

Answer 3

I see 2 possible things that could help improve performance. Depends on which part is taking the longest. First off, SELECT INTO is a single threaded operation until SQL Server 2014. If this is taking a long time, create an explicitly defined temp table with CREATE TABLE. Secondly, depending on the number of records inserted into the temp table, you probably need an index on the Report_ID column. That can all be done in the body of the stored procedure. If you do end up using an explicitly defined temp table, I would create the index after the data is loaded.
If that doesn't help, first check that the report_id column on the BIGTABLE is indexed. Then try splitting the select into 2 and combining with a UNION ALL like this:
ALTER PROCEDURE dbo.p_revenue
(
#report_id VARCHAR(MAX) = NULL
)
AS
SELECT CAST(value AS INT) Report_ID
INTO #report_ID_Temp
FROM fn_Parse_List(#report_id);
SELECT *
FROM BIGTABLE
WHERE #report_id IS NULL
UNION ALL
SELECT *
FROM BIGTABLE
WHERE a.report_id IN ( SELECT Report_ID
FROM #report_ID_Temp );
GO
EXEC p_revenue #report_id = '456';

Answer 4

Are you saying I should have two queries, one where it pulls if the report_id doesn't exists and one where there is a list of report_ids?
Yes, yes, yes. The fact, that it somehow works when You enter the numbers directly, distracts You from the core problem. You need table scan when #report_id is null and index seek when it is not and You can not have both in one execution plan. The performance would inevitably have to suffer, one way or another.
I would prefer not to, as the table i'm pulling from is actually a
view with 800 lines with an additional parameter not shown above.
I do not see where is the problem, SELECT * FROM BIGTABLE and SELECT * FROM BIGVIEW seems the same. If You need parameters You can use inline table valued function. If You have more parameters with variable selectivity like #report_id, I guess You would end up with dynamic sql anyway, sooner or later.
UNION ALL as proposed by #db_brad would help, but one of those subquery is executed even when there is no need for it.
As a quick patch You can append OPTION(RECOMPILE) to the SELECT and have table scan one time and index seek the other time, but recompiling every time would induce nontrivial overhead.