MINCE: Dialect-Aware SQL Decomposition for Federated Query Execution

Abstract

The increasing adoption of specialized database systems has led to the rise of heterogeneous data environments. While having multiple engines in a data infrastructure enables opportunities for workload optimization, SQL dialect incompatibility makes workload migration difficult. To address this challenge, we develop MINCE (Multi-dialect INtegration and Crossengine Execution), a technique that decomposes SQL queries into parts to enable federated execution across engines with differing SQL dialects. MINCE uses a rule-based method to partition a query into executable components that are assigned to different database systems. To evaluate different execution strategies, MINCE further implements a cost model that incorporates both on-engine query execution time and inter-system data transfer overhead. We evaluate MINCE on a TPC-H-based workload augmented with PostgreSQL-specific functions unsupported in Amazon Redshift. Experimental results show that MINCE produces the fastest execution strategy among our baselines for 72.1% of queries using estimated cardinality, achieving a 2× speedup over single-engine baselines. With perfect cardinality information available to our cost model, this value increases to 88.4%, with an average 2.8× speedup. These results demonstrate that our system not only enables more flexible federated query execution, but also reliably identifies performant execution strategies.