Feedback-Directed Query Optimization

Abstract

Current database systems employ static heuristics for estimating the access time of a particular query. These heuristics are based on several parameters, such as relation size and number of tuples. Yet these parameters are only updated intermittently, and the heuristics themselves are hand-tuned. As trends in database systems aim toward self-tuning systems, we can apply the experience of the feedback-directed compiler world to provide robust, self-tuning query optimizers. This paper presents the design and evaluation of a feedback-directed query optimization infrastructure. Using trace-driven simulation, we conclude that dynamic feedback can be quite effective at improving the accuracy of a query optimizer, and adapting to predictable query overhead.