Tag Archive for Set Theory

Using Set Theory Instead of ISNULL To Filter Data

I have written dozens upon dozens of reports for my users. I love the challenge of finding the data and wrangling it into a report in a way that doesn’t take a hundred years to run when they decide to return a year’s worth of data. One of the common requests that I get is to provide a parameter that will allow them to choose a single client or all clients. At first blush this is a very simple request and can be accomplished by using the ISNULL function.

ISNULL - ISNULL Code

Unfortunately there are performance implications using the ISNULL function. The ISNULL function is non-sargable. This means that when the function is used as part of a predicate, it can’t utilize the necessary index in an optimal way to filter the data. In this example, the Execution Plan shows that a full index scan was used to return one row. This equated to 80 Logical Reads for my dataset. This is the same number of Logical Reads whether one row was returned or all rows were returned. Eighty Logical Reads may not be that big a deal, but what about on a table that has hundreds of thousands of rows?

ISNULL - ISNULL Execution Plan
Execution Plan

ISNULL - ISNULL Reads 1
Statistics on the Table I/O Tab in SQL Sentry Plan Explorer Pro

 

An Alternate Universe

There is an alternate way of accomplishing this same request and it uses Set Theory to solve the problem. I can accomplish the same request by separating out the ISNULL logic into two different SQL statements and using UNION ALL to join the two result sets into one result set.  Even though I’m using two different SQL statements, only one will return data during the execution of the logic.

ISNULL - Union All Code

The first SELECT statement returns rows when the @ClientID variable has a value other than NULL. The second SELECT statement returns rows when @ClientID is NULL. By using UNION ALL instead of UNION we forgo the task of checking if there are duplicate rows in the second SELECT statement.

The Execution Plan now contains the ability to handle both cases, and the Logical Reads changes based on the value of @ClientID. There are still 80 Logical Reads when @ClientID is NULL because the whole table is being returned, but the Logical Reads are reduced to two when @ClientID is set to a single client number.

 

ISNULL - Union All Execution Plan 1
Execution Plan when @ClientID =7890

ISNULL - Union All Reads 1

 

Statistics when @ClientID = 7890
ISNULL - Union All Execution Plan 2
Execution plan when @ClientID IS NULL

ISNULL - Union All Reads 2
Statistics when @ClientID IS NULL

 

In this example I was able to use UNION ALL to filter on one or all clients, but it isn’t always the best solution. If the table being filtered fits on one page, then a table scan will occur regardless of how the predicate is setup. In that case, the ISNULL statement is easier to maintain so it would be the better solution.

Conclusion

It is always best to take a look at the Execution Plan of a SQL statement to see if there are any Index Scans on filtered data. If there are, you should take a look to see if there is another, set based approach to solve the problem.

%d bloggers like this: