In order to tune a query, you need to know two things:
– can it be tuned, and if yes, then by how much
– which part of the query (which operation, or which data object) is most promising from the tuning point of view.
Currently existing tuning methods don’t really answer these questions. You either focus on the most expensive operation(s), and hope that you can eliminate them (or transform them into something less cosly), or you focus on the ones where you see a large discrepancy between actual rowcounts and optimizer predictions (cardinality feedback tuning). Either way, you can’t be sure that you’ve set your priorities right. It could well be the case that the cost of the most expensive operation cannot be reduced by much, but you can win back enough performance elsewhere. With the cardinality feedback tuning, you also don’t have any guarantee that improving accuracy of optimizer estimates would eventually transform into acceptable level of performance.
Of course, if the plan only contains a few operations, this is not a big issue, and after a few trials you will usually get to the bottom of the problem. However, when dealing with very complex plans, hundreds operations long, this is not really an option. When dealing with such plans a few months back, I developed for myself a simple tuning method that allows to evaluate with high accuracy potential tuning benefits of plan operations, using rowsource stats and optimizer as input. In this post, I’m sharing this method, as well as a script that implements it.