Automatic relationship discovery in self-managing database systems

Abstract

In this paper, we describe CORDS, an algorithm that automatically discovers correlations and soft functional dependencies (FDs) between pairs of columns and, based on these relationships, determines a set of statistics to maintain. This data-driven technology is an essential complement to query-driven approaches such as LEO, helping to ensure acceptable performance during slow learning periods. CORDS focuses on column pairs because this greatly simplifies the algorithms, and experiments have shown that the marginal benefit of capturing n-way dependencies for n > 2 is relatively small.