Investigating Learning Join Order Optimization Strategies for Rule-based Data Engines

Abstract

A recent trend in data management research investigates whether machine learning techniques could improve or replace core components of traditional database architectures, such as the query optimizer or selectivity and cardinality cost estimators. The preliminary approaches leverage cost-based optimizers and cost models to avoid a cold-start as they train and build learning models. In this work, we investigate whether learning could also be beneficial in rule-based optimizers, which instead of driving query execution decisions via a cost model they rely on a set of fixed rules and pre-defined heuristics. Our experimental testbed employs MonetDB, an open-source, column-store analytics data engine, and explore whether a learning model using Graph Neural Networks (GNNs) that is trained on a cost-based engine, such as PostgreSQL, could improve MonetDB optimizer’s decisions. Our initial findings reveal that our approach could improve significantly MonetDB’s query execution plans, especially as the query complexity increases whet it involves many join operators.