A Spline-based Cost Model for Metric Trees

Abstract

Whenever two (or more) access methods are alternatives for the execution of a query, how to choose which one is the best for the task? Such a decision is made by the DBMS optimizer module, which models the query costs according to the distribution of the data space. Cost modeling of similarity searches, however, requires the representation of distances’ rather than data distribution. In this paper, we propose the Stockpile model for cost estimation of similarity queries on metric trees by using pivot-based distance histograms that represent the local densities around the query elements. By combining the local densities to the probability of traversing the tree nodes, Stockpile provides a fair estimation of both disk accesses (I/O costs) and distance calculations (CPU costs). We compared Stockpile and two literature models regarding similarity queries in real-world data sources and our model was up to 85% more precise than the competitors.