Sampling issues in parallel database systems

Abstract

Sampling has proven useful in database systems in applications including query size estimation, and most recently, probabilistic parallel query evaluation algorithms. In order to apply the full power of modern multiprocessor database systems, sampling techniques must (1) distribute the sampling workload evenly among the processors in the system, and (2) make use of all the data on the pages brought into main memory during the course of the sampling. In this paper we show how to achieve these two goals by proving that for query size estimation, (1) stratified random sampling guarantees perfect load balancing without reducing the accuracy of the estimate, and that (2) for a given number of I/O operations, page level sampling always produces a more accurate estimate than tuple level sampling. For probabilistic parallel query evaluation algorithms, high performance requires tight boundsxon the expected skew in the allocation of work to processors as a function of the number of samples. Toward this end we prove a new bound on this skew, and show that our new bound is better than previously known bounds.