NOT IN vs NOT EXISTS

I always default to NOT EXISTS.

The execution plans may be the same at the moment but if either column is altered in the future to allow NULLs the NOT IN version will need to do more work (even if no NULLs are actually present in the data) and the semantics of NOT IN if NULLs are present are unlikely to be the ones you want anyway.

When neither Products.ProductID or [Order Details].ProductID allow NULLs the NOT IN will be treated identically to the following query.

SELECT ProductID,
       ProductName
FROM   Products p
WHERE  NOT EXISTS (SELECT *
                   FROM   [Order Details] od
                   WHERE  p.ProductId = od.ProductId) 

The exact plan may vary but for my example data I get the following.

A reasonably common misconception seems to be that correlated sub queries are always “bad” compared to joins. They certainly can be when they force a nested loops plan (sub query evaluated row by row) but this plan includes an anti semi join logical operator. Anti semi joins are not restricted to nested loops but can use hash or merge (as in this example) joins too.

/*Not valid syntax but better reflects the plan*/ 
SELECT p.ProductID,
       p.ProductName
FROM   Products p
       LEFT ANTI SEMI JOIN [Order Details] od
         ON p.ProductId = od.ProductId 

If [Order Details].ProductID is NULL-able the query then becomes

SELECT ProductID,
       ProductName
FROM   Products p
WHERE  NOT EXISTS (SELECT *
                   FROM   [Order Details] od
                   WHERE  p.ProductId = od.ProductId)
       AND NOT EXISTS (SELECT *
                       FROM   [Order Details]
                       WHERE  ProductId IS NULL) 

The reason for this is that the correct semantics if [Order Details] contains any NULL ProductIds is to return no results. See the extra anti semi join and row count spool to verify this that is added to the plan.

If Products.ProductID is also changed to become NULL-able the query then becomes

SELECT ProductID,
       ProductName
FROM   Products p
WHERE  NOT EXISTS (SELECT *
                   FROM   [Order Details] od
                   WHERE  p.ProductId = od.ProductId)
       AND NOT EXISTS (SELECT *
                       FROM   [Order Details]
                       WHERE  ProductId IS NULL)
       AND NOT EXISTS (SELECT *
                       FROM   (SELECT TOP 1 *
                               FROM   [Order Details]) S
                       WHERE  p.ProductID IS NULL) 

The reason for that one is because a NULL Products.ProductId should not be returned in the results except if the NOT IN sub query were to return no results at all (i.e. the [Order Details] table is empty). In which case it should. In the plan for my sample data this is implemented by adding another anti semi join as below.

The effect of this is shown in the blog post already linked by Buckley. In the example there the number of logical reads increase from around 400 to 500,000.

Additionally the fact that a single NULL can reduce the row count to zero makes cardinality estimation very difficult. If SQL Server assumes that this will happen but in fact there were no NULL rows in the data the rest of the execution plan may be catastrophically worse, if this is just part of a larger query, with inappropriate nested loops causing repeated execution of an expensive sub tree for example.

This is not the only possible execution plan for a NOT IN on a NULL-able column however. This article shows another one for a query against the AdventureWorks2008 database.

For the NOT IN on a NOT NULL column or the NOT EXISTS against either a nullable or non nullable column it gives the following plan.

When the column changes to NULL-able the NOT IN plan now looks like

It adds an extra inner join operator to the plan. This apparatus is explained here. It is all there to convert the previous single correlated index seek on Sales.SalesOrderDetail.ProductID = <correlated_product_id> to two seeks per outer row. The additional one is on WHERE Sales.SalesOrderDetail.ProductID IS NULL.

As this is under an anti semi join if that one returns any rows the second seek will not occur. However if Sales.SalesOrderDetail does not contain any NULL ProductIDs it will double the number of seek operations required.