Smooth Interpolating Histograms with Error Guarantees

Accurate selectivity estimations are essential for query optimization decisions where they are typically derived from various kinds of histograms which condense value distributions into compact representations. The estimation accuracy of existing approaches typically varies across the domain, with some estimations being very accurate and some quite inaccurate. This is in particular unfortunate when performing a parametric search using these estimations, as the estimation artifacts can dominate the search results. We propose the usage of linear splines to construct histograms with known error guarantees across the whole continuous domain. These histograms are particularly well suited for using the estimates in parameter optimization. We show by a comprehensive performance evaluation using both synthetic and real world data that our approach clearly outperforms existing techniques.