Handling Estimation Inaccuracy in Query Optimization

Cost-Based Optimizers choose query execution plans using a cost model. The latter relies on the accuracy of estimated statistics. Unfortunately, compile-time estimates often differ significantly from runtime values, leading to a suboptimal plan choices. In this paper, we propose a compile-time strategy, wherein the optimization process is fully aware of the estimation inaccuracy. This is ensured by the use of intervals of estimates rather than single-point estimates of error-prone parameters. These intervals serve to identify plans that provide stable performance in several run-time conditions, so called robust. Our strategy relies on a probabilistic approach to decide which plan to choose to start the execution. Our experiments show that our proposal allows a considerable improvement of the ability of a query optimizer to produce a robust execution plan in case of large estimation errors. The produced plan is also competitive with those obtained using existing optimization methods when errors are small.