ML-Driven Performance Tuning Framework for Heterogeneous Unified Databases

Abstract

Modern organizations increasingly operate in heterogeneous database environments encompassing relational systems such as Oracle, SQL Server, and PostgreSQL, alongside NoSQL platforms such as MongoDB and Cassandra. This diversity introduces significant challenges in performance tuning because each engine exhibits distinct workload behaviors, tuning parameters, and performance indicators. Traditional manual tuning approaches are limited in terms of scalability and effectiveness, particularly under dynamic, high-volume workloads. This study proposes a unified, Python-based machine-learning framework designed to automate performance tuning across heterogeneous databases. The framework incorporates cross-database metric collection, feature engineering, supervised and unsupervised learning models for bottleneck prediction, and an intelligent recommendation engine that generates actionable strategies. Experiments conducted using benchmark workloads and real-world performance traces demonstrated notable reductions in query latency, improvements in throughput, and increased system stability across all supported database engines. Comparative analysis indicates that the ML-driven approach outperforms vendor-native recommendations and manual heuristics in terms of both accuracy and tuning impact. The findings highlight the feasibility and effectiveness of adopting a unified ML-driven tuning architecture for diverse database environments. Future extensions include integrating reinforcement learning, expanding support for additional database paradigms, and developing automated rollback-safe execution workflows.