AI-Augmented Query Optimization: A Technical Survey of Foundations, Systems, and Open Challenges

Query optimization has been a cornerstone of database management systems for over four decades, yet traditional cost-based optimizers struggle with complex analytical workloads, userdefined functions, and correlated data distributions. Recent advances in artificial intelligence and machine learning offer a transformative approach: learning optimization strategies from data rather than relying solely on hand-crafted heuristics. This survey provides a comprehensive analysis of AI-augmented query optimization, examining the evolution from rule-based systems to learned components that adapt to workload characteristics and system dynamics. We present a systematic taxonomy of learned optimization approaches spanning four major categories: learned cardinality estimation, learned cost models, reinforcement learning for plan enumeration, and endto-end learned optimizers. Through detailed comparative analysis of systems including Bao, Neo, Lero, and recent LLM-based approaches, we demonstrate that learned optimizers achieve 1.5-3× geometric mean speedups on standard benchmarks while reducing cardinality estimation errors by an order of magnitude. However, challenges remain in generalization, interpretability, training data requirements, and production deployment. We critically examine evaluation methodologies, identifying reproducibility concerns and inconsistent benchmarking practices that hinder progress. Our analysis of integration strategies reveals that hybrid architectures combining traditional and learned components offer the most promising path to production adoption. Finally, we outline future directions including neuro-symbolic optimization, federated learning, transfer learning across databases, and the development of standardized MLOps pipelines. This survey serves as both a comprehensive reference for researchers and a practical guide for practitioners considering AI-augmented optimization in production systems.