Hypertree Decompositions for Query Optimization

The database community has investigated many structure-driven methods, which guarantee that large classes of queries may be answered in (input-output) polynomial-time. However, despite their very nice computational properties, these methods are not currently used for practical applications, since they do not care about output variables and aggregate operators, and do not exploit quantitative information on the data. In fact, none of these methods has been implemented inside any available DBMS. This paper aims at filling this gap between theory and practice. First, we define an extension of the notion of hypertree decomposition, which is currently the most powerful structural method. This new version, called query-oriented hypertree decomposition, is a suitable relaxation of hypertree decomposition designed for query optimization, and such that output variables and aggregate operators can be dealt with. Based on this notion, a hybrid optimizer is implemented, which can be used on top of available DBMSs to compute query plans. The prototype is also integrated into the well-known open-source DBMS PostgreSQL. Finally, we validate our proposal with a thorough experimental activity, conducted on PostgreSQL and on a commercial DBMS, which shows that both systems may significantly benefit from using hypertree decompositions for query optimization.