Cost models for overlapping and multiversion structures

Abstract

Overlapping and multiversion techniques are two popular frameworks that transform an ephemeral index into a multiple logical-tree structure in order to support versioning databases. Although both frameworks have produced numerous efficient indexing methods, their performance analysis is rather limited; as a result there is no clear understanding about the behavior of the alternative structures and the choice of the best one, given the data and query characteristics. Furthermore, query optimization based on these methods is currently impossible. These are serious problems due to the incorporation of overlapping and multiversion techniques in several traditional (e.g., financial) and emerging (e.g., spatiotemporal) applications. In this article, we reduce performance analysis of overlapping and multiversion structures to that of the corresponding ephemeral structures, thus simplifying the problem significantly. This reduction leads to accurate cost models that predict the sizes of the trees, the node/page accesses, and selectivity of queries. Furthermore, the models offer significant insight into the behavior of the structures and provide guidelines about the selection of the most appropriate method in practice. Extensive experimentation proves that the proposed models yield errors below 5 and 15% for uniform and nonuniform data, respectively.