Abstract

Query optimizers are complex subsystems of database management systems. Modifying query optimizers to admit new algorithms or storage structures is quite difficult, but partly alleviated by extensible approaches to optimizer construction. Rule-based optimizers are a step in that direction, but from our experience, the rule sets of such optimizers are rather monolithic and brittle. Conceptually minor changes often require wholesale modifications to a rule set. Consequently, much can be done to improve the extensibility of rule-based optimizers. As a remedy, we present a tool called Prairie that is based on an algebra of layered optimizers. This algebra naturally leads to a building-blocksapproach to rule-set construction. Defining customized rule sets and evolving previously defined rule sets is accomplished by composing building-blocks. We explain an implementation of Prairie and present experimental results that show how classical relational optimizers can be synthesized from building-blocks, and that the efficiency of query optimization is not sacrificed. 1

Citations