MODT: Multi-Objective Database Tuner Using Hierarchical Reinforcement Learning

Index recommendation and knob tuning are two important database tuners. Despite substantial progress in each of them, how these tuners together affect the overall database performance is still an open question. There exists a critical research gap in addressing integrated optimization of these tuners especially with additional consideration of resource utilization. Only a few works have focused on this, with challenges including high-dimensional search space, difficulty in model fitting, and delayed evaluation bias. To address these issues, we propose MODT, a novel Multi-Objective Database Tuning framework, which combines hierarchical reinforcement learning (HRL) with a two-level recursive structure to automatically provide sequential configuration of indexes and knobs based on workload characteristics and database status. Compared with state-of-the-art integrated optimization approaches on TPC-H, TPC-DS, and Join Order Benchmark (JOB), MODT can find competitive index-knob configurations and outperforms competitors in reducing execution time and resource utilization.