Abstract

We study the performance evaluation and analysis of Flood, a learned index designed to process multi-dimensional queries over relational data. Unlike traditional indexing methods such as KD-Trees and R-Trees, which rely on static partitioning strategies, Flood leverages machine learning techniques to dynamically adapt its grid structure and data layout based on data distribution and query workloads. We identify and evaluate the core components of the Flood framework—including grid-based partitioning, learned layout optimization, and refinement steps—and assess the impact of each component on overall system performance. We compare Flood’s efficiency and scalability against baselines such as KD-Trees, Z-order indexing, and brute-force scan across multiple datasets and query workloads. Our experiments reveal that scan emerges as the dominant bottleneck, with layout tuning and component configuration introducing significant overhead. To gain a deeper understanding of Flood and explore opportunities for improvement, we developed a modular prototype implementation from scratch. This modular design enabled a systematic, in-depth performance study by isolating components and allowing for alternative configurations. We also refined the cost model calibration and proposed a new optimization strategy for guiding layout selection. Our work contributes to more effective configuration and tuning of learned indexes by offering insights into the trade-offs, limitations, and opportunities of using learned indexes as an alternative or complementary solution for supporting multi-dimensional queries in relational database systems.