StatAdvisor: recommending statistical views

Database statistics are crucial to cost-based optimizers for estimating the execution cost of a query plan. Using traditional basic statistics on base tables requires adopting unrealistic assumptions to estimate the cardinalities of intermediate results, which usually causes large estimation errors that can be several orders of magnitude. Modern commercial database systems support statistical or sample views, which give more accurate statistics on intermediate results and query sub-expressions. While previous research focused on creating and maintaining these advanced statistics, only little effort has been done towards automatically recommending the most beneficial statistical views to construct. In this paper, we present StatAdvisor , a system for recommending statistical views for a given SQL workload. The StatAdvisor addresses the special characteristics of statistical views with respect to view matching and benefit estimation, and introduces a novel plan-based candidate enumeration method, and a benefit-based analysis to determine the most useful statistical views. We present the basic concepts, architecture, and key features of StatAdvisor , and demonstrate its validity and benefits through an extensive experimental study using a prototype that we built in the IBM® DB2® database system as part of the DB2 Design Advisor tools.