Rewrite SQL Query- I need to replace NOT IN with Join - sql-server

Question

I have a query in my production environment which is taking long time to execute. I did not write this query but I must find a way to make it quicker since it is causing a big performance issue at the moment. I need to replace NOT IN with Left Join but not sure how to rewrite it. It looks like following at the moment
SELECT TOP 1 IT.ITEMID
FROM (SELECT CAST(ITEMID AS NUMERIC) + 1 ITEMID
FROM Items
WHERE ISNUMERIC(ITEMID) = 1
AND CAST(ITEMID AS NUMERIC) >= 50000) IT
WHERE IT.ITEMID NOT IN (SELECT CAST(ITEMID AS NUMERIC) ITEMID
FROM Items
WHERE ISNUMERIC(ITEMID) = 1)
ORDER BY IT.ITEMID
Kindly suggest how am I supposed to rewrite it using Left Join for better performance. Any help/guidance is greatly appreciated.

Answer 1

Try this one -
;WITH cte AS
(
SELECT DISTINCT ITEMID =
CASE WHEN ISNUMERIC(ITEMID) = 1
THEN ITEMID
END
FROM Items
)
SELECT TOP 1 ITEMID = ITEMID + 1
FROM cte t
WHERE ITEMID >= 50000
AND NOT EXISTS(
SELECT 1
FROM cte t2
WHERE t.ITEMID + 1 = t2.ITEMID
)
ORDER BY t.ITEMID

Answer 2

As mentioned in the comments, the NOT EXISTS version of the query is usually faster in SQLServer than the LEFT JOIN - for completeness, here's both versions:
Left join variant of existing query:
with cte as
(SELECT CAST(it.ITEMID AS NUMERIC) ITEMID
FROM Items
WHERE ISNUMERIC(ITEMID) = 1)
select top 1 i.ITEMID + 1 ITEMID
FROM cte i
LEFT JOIN cte ni ON i.ITEMID + 1 = ni.ITEMID
WHERE i.ITEMID >= 50000 AND ni.ITEMID IS NULL
Not exists variant of existing query:
with cte as
(SELECT CAST(it.ITEMID AS NUMERIC) ITEMID
FROM Items
WHERE ISNUMERIC(ITEMID) = 1)
select top 1 i.ITEMID + 1 ITEMID
FROM cte i
WHERE i.ITEMID >= 50000 AND NOT EXISTS
(SELECT NULL
FROM cte ni
WHERE i.ITEMID + 1 = ni.ITEMID)

Answer 3

As #gbn pointed at the comments, the CAST and functions on predicates which invalidates index use anyway, so there is no point in converting this from NOT IN to LEFT JOIN / IS NULL or to NOT EXISTS. And NOT EXISTS usually performs better than LEFT NULL in SQL-Server.
NOT IN is not advised due to the problems (wrong, unexpected results) when there are nulls (in the compared columns or produced by the expressions) and the inefficient plans because of the nullability of the columns/expessions.
And ISNUMERIC() is not doing always what you think it does (as # Damien_The_Unbeliever noted in another comment.) There are cases where the IsNumeric result is 1 but the cast fails.
So, the sane thing to do would be - in my opinion - to add another column in the table and convert (the values that can be converted) to numeric and store them in that column. Then you could write the query without casting and an index on that column could be used.
If you cannot alter the tables in any way (by adding a new column or a materialized view), then you can try and test the various rewritings the other answers offer.

Answer 4

I agree with #ypercube that the sane thing to do is to fix your schema.
If for some reason this is not an option maybe materialising the whole thing into an indexed temporary table at runtime would make the best of a bad job.
CREATE TABLE #T
(
ITEMID NUMERIC(18,0) PRIMARY KEY
WITH ( IGNORE_DUP_KEY = ON)
)
INSERT INTO #T
SELECT CASE WHEN ISNUMERIC(ITEMID) = 1 THEN ITEMID END
FROM Items
WHERE CASE WHEN ISNUMERIC(ITEMID) = 1 THEN ITEMID END >= 50000
SELECT TOP 1 ITEMID+1
FROM #T T1
WHERE NOT EXISTS (SELECT * FROM #T T2 WHERE T2.ITEMID = T1.ITEMID +1)
ORDER BY ITEMID