Lequa: A Learning-Based Query-Aware Framework for Selective Query Optimization

Learning-based query optimizers have shown promise in improving database performance. However, existing learning-based methods often fail to consistently improve end-toend performance across all queries due to inference overhead and variability in both query characteristics and data distributions. To address these challenges, we propose Lequa, a learningbased query-aware framework that selects appropriate optimization strategies for each query based on expected performance gain and query complexity. Lequa addresses different queries through three components. The Unprofitable Query Recognizer component identifies queries unlikely to benefit from learned optimization and routes them directly to the default optimizer, avoiding unnecessary inference. The Struct Planner component uses a lightweight model based on dataset-agnostic features to handle queries whose optimal plans can be inferred from general structural patterns. For more complex queries where general patterns fall short, the Detail Planner component leverages datasetspecific features to produce higher-quality plans. Together, these components enable Lequa to generalize across diverse workloads while selectively applying fine-grained optimization where it matters most. Experiments on real-world workloads show that our approach consistently outperforms state-of-the-art baselines, achieving up to 55% reduction in end-to-end latency compared to PostgreSQL.