アイテム詳細

要旨

This paper aims to improve the accuracy of query result-size estimations in query optimizers by leveraging the dynamic feedback obtained from observations on the executed query workload. To this end, an approximate "synopsis" of data-value distributions is devised that combines histograms with parametric curve fitting, leading to a specific class of linear splines. The approach reconciles the benefits of histograms, simplicity and versatility, with those of parametric techniques especially the adaptivity to statistically biased and dynamically evolving query workloads. The paper presents efficient algorithms for constructing the linear-spline synopsis for data-value distributions from a moving window of the most recent observations on (the most critical) query executions. The approach is worked out in full detail for capturing frequency as well as density distributions of data values, and it is shown how result size estimations are inferred for exact-match and range queries as well as projections and grouping. To a large extent, the developed methods can be generalized to multi-dimensional distributions, thus bearing the ability to capture correlations among attributes as well. Experimental studies underline the accuracy of the developed estimation methods, outperforming the best known classes of histograms.