A Systematic Review of Modern Machine Learning in Query Optimization

Abstract

Traditional query optimization struggles with modern, heterogeneous workloads. This PRISMA-guided survey examines 27 recent studies (2018–2025) on machine learning for query optimization, addressing two research questions: (RQ1) how learning-based approaches intervene across the classical optimization pipeline and what trade-offs characterize each stage, and (RQ2) what limitations persist in generalization, benchmark standardization, and deployment readiness. We synthesize the literature across five intervention points: component-level estimation, learned enumeration, hint steering, end-to-end optimization, and emerging trends. Results show a paradigm shift from isolated predictors to integrated optimizers, while gaps in generalization and standardized evaluation persist. We conclude with a roadmap for robust and reproducible data-driven query optimization.