HASE: A Hybrid Approach to Selectivity Estimation for Conjunctive Predicates

Abstract

Current methods for selectivity estimation fall into two broad categories, synopsis-based and sampling-based. Synopsis-based methods, such as histograms, incur minimal overhead at query optimization time and thus are widely used in commercial database systems. Sampling-based methods are more suited for ad-hoc queries, but often involve high I/O cost because of random access to the underlying data. Though both methods serve the same purpose of selectivity estimation, their interaction in the case of selectivity estimation for conjuncts of predicates on multiple attributes is largely unexplored. Our work aims at taking the best of both worlds, by making consistent use of synopses and sample information when they are both present. To achieve this goal, we propose HASE, a novel estimation scheme based on a powerful mechanism called generalized raking. We formalize selectivity estimation in the presence of single attribute synopses and sample information as a constrained optimization problem. By solving this problem, we obtain a new set of weights associated with the sampled tuples, which has the nice property of reproducing the known selectivities when applied to individual predicates. We discuss different variants of the optimization problem and provide algorithms for solving it. We also provide asymptotic error bounds on the estimate. Extensive experiments are performed on both synthetic and real data, and the results show that HASE significantly outperforms both synopsis-based and sampling-based methods.