Robust Index Benefit Estimation via Hierarchical and Two-Dimensional Feature Representation

In recent years, machine learning-based index advisors have gained success as they can estimate the benefit of a given index without actually evaluating it via what-if optimizers. However, existing methods often fail to capture indexrelevant features adequately, leading to limited accuracy and poor adaptability to changes in workload, schema, or data. To address these challenges, we propose Eddie, a novel index benefit estimation approach based on Hierarchical and Twodimensional Encoding. Our method encodes columns according to their positions in queries and indexes, consolidates indexrelated features into a compact representation, and leverages twodimensional attention to model both query plan structure and index interactions. Finally, Eddie can be seamlessly integrated with existing index advisors in real-world systems. 11https://github.com/quanjnq/Eddie Extensive evaluations demonstrate that it significantly outperforms stateof-the-art estimators in both accuracy and robustness.