Workload-independent data-driven vertical partitioning

Abstract

Vertical partitioning is a well-explored area of automatic physical database design. The classic approach is as follows: Derive an optimal vertical partitioning scheme for a given database and a workload. The workload describes queries, their frequencies, and involved attributes. In this paper we identify a novel class of vertical partitioning algorithms. The algorithms of this class do not rely on knowledge of the workload, but instead use data properties that are contained in the workload itself. We propose such algorithm that uses a logical scheme represented by functional dependencies, which are derived from stored data. In order to discover functional dependencies we use TANE — a popular functional dependency extraction algorithm. We evaluate our algorithm using an industrial DBMS (PostgreSQL) on number of workloads. We compare the performance of an unpartitioned configuration with partitions produced by our algorithm and several state-of-the-art workload-aware algorithms.