MFTune: An Efficient Multi-fidelity Framework for Spark SQL Configuration Tuning
International Conference on Very Large Data Bases (VLDB), 2026.
* Equal contribution.
Overview
MFTune uses representative SQL subsets as low-cost proxies for full-workload evaluations. It combines query-based fidelity partitioning, density-based configuration-space compression, transfer learning, and a two-phase warm start to make Spark SQL tuning more efficient.