Tuesday, April 07, 2015

SQL Performance Basics: Part 4, The Order of Predicates

Did you know that the order in which you code your predicates can have an impact on query performance? It is usually a minimal impact, but it may buy you a couple of microseconds for a very performance-critical query. In order to use predicate ordering to your advantage however, you  need to be armed with some basic information on how DB2 evaluates predicates as it processes your SQL.

So, before we continue, let's review the order in which DB2 evaluates predicates at execution time. DB2 will evaluate indexable predicates first: matching predicates before non-matching. Then, Stage 1 predicates, and finally Stage 2 predicates. Within each of these four groups, DB2 will evaluate equal predicates, then BETWEEN and NOT NULL predicates, and finally, any other predicates. If more than one predicate exists within a group, then DB2 will evaluate them in the physical order in which they are coded in the SQL statement. 

The re-ordering of predicates to take advantage of this situation should be considered only as a last resort. When implemented, the technique will usually shave only a little bit from the query's execution time. It is also important to note that predicate order will not impact a query's access path: it will remain unchanged (as shown in the PLAN_TABLE).

Now, how can we use this to our advantage

Consider the following query:


For the purposes of this discussion, no index exists for either of the columns coded in the predicates. They are therefore the same type: stage 1 and equal predicates. Furthermore, we know our data - in our organization, there is approximately a 50-50 split between males and females, and 15% of all employees are managers.

To optimize this query then, we can swap the two predicates to achieve better performance. So the query becomes:


Why should this query outperform the previous version? Well, assume we have 100,000 employees. If DB2 retrieves 50% of the rows (SEX = 'M') and then retrieves 15% of those 50%, we will have processed 57,500 rows:

( 100000 * 0.5 ) + ( ( 100000 * 0.5 ) * 0.15 ) = 57,500

But, if instead, DB2 were to retrieve 15% of the rows (TITLE = 'MANAGER') and then 50% of those, we will have processed only 22,500 rows:

( 100000 * 0.15) + ( ( 100000 * 0.15 ) * 0.5 ) = 22,500

Obviously, it is better for fewer rows to qualify early, thereby reducing the answer set and the number of rows that will have to be subsequently scanned.

No comments: