Attention-based model for reliable and active benefit estimation in index tuning

Automatic index tuning is crucial for optimizing database query performance, but quickly achieving the optimal configuration is challenging due to difficulties in accurately estimating benefits. Traditional methods, such as using what-if tools, are often inefficient and imprecise. To overcome these limitations, several learning-based models have been proposed: (1) extracting features from query plans and index configurations to estimate benefits, (2) filtering configurations based on what-if results, and (3) enhancing model stability through uncertainty quantification. While these methods differ in feature extraction and prediction techniques, they lack comprehensive analysis regarding their impact on tuning results. Moreover, none achieve optimal levels of accuracy, efficiency, and stability in benefit estimation. This paper compares and analyzes these techniques, integrating their strengths into a new model. We design a novel feature extractor with an attention-based encoder for query plans and index configurations, improving feature extraction and accuracy. Additionally, we refine the uncertainty quantification framework, increasing the precision of confidence measurements and applying it to active training data selection, thus reducing data collection and training costs. Our experiments validate the model's accuracy, efficiency, and uncertainty quantification across six benchmarks, achieving optimal accuracy in 80% of metrics and outperforming five other methods in index tuning.