Stability-conscious Query Optimization

Modern query optimizers choose their execution plans primarily on a cost-minimization basis, assuming that the inputs to the costing process, such as relational selectivities, are accurate. However, in practice, these inputs are subject to considerable run-time variation relative to their compile-time estimates, often leading to poor plan choices that cause inflated response times. We present in this paper a parametrized family of online plan generation and selection algorithms that substitute, whenever feasible, the optimizer’s solely cost-conscious choice with an alternative plan that is (a) guaranteed to be near-optimal in the absence of selectivity estimation errors, and (b) likely to deliver comparatively stable performance in the presence of arbitrary errors. The proposed algorithms have been implemented within the PostgreSQL optimizer, and their performance evaluated on a rich spectrum of TPC-H and TPC-DS-based query templates in a variety of database environments. Our experimental results indicate that it is indeed possible to identify robust plan choices that substantially curtail the adverse effects of erroneous selectivity estimates. In fact, the plan selection quality provided by our online algorithms is often competitive with those obtained through apriori knowledge of the plan search and optimality spaces. Further, the additional optimization overheads incurred by our algorithms are miniscule in comparison to the expected savings in query execution times. Finally, we also demonstrate that with appropriate parameter choices, it is feasible to directly produce anorexic plan diagrams, a potent objective in query optimizer design.