Workload-Based Wavelet Synopses

This paper introduces workload-based wavelet synopses, which exploit query workload information to significantly boost accuracy in approximate query processing. We show that wavelet synopses can adapt effectively to workload information, and that they have significant advantages over previous approaches. An important aspect of our approach is optimizing synopses constructions toward error metrics defined by workload information, rather than based on some uniform metrics. We present an adaptive greedy algorithm which is simple and efficient. It is run-time competitive to previous, non-workload based algorithms, and constructs workload-based wavelet synopses that are significantly more accurate than previous synopses. The algorithm also obtains improved accuracy for non-workload case when the error metric is the mean relative error. We also present a self-tuning algorithm that adapts the workload-based synopses to changes in the workload. All algorithms are extended to workload-based multidimensional wavelet synopses with improved performance over previous algorithms. Experimental results demonstrate the effectiveness of workload-based wavelet synopses for different types of data sets and query workloads, and show significant improvement in accuracy even with very small training sets.