Making In-Memory Learned Indexes Efficient on Disk

Abstract

Learned indexes have been demonstrated to outperform traditional ones in memory-resident scenarios. However, recent studies show that they fail to outperform B+tree when extended to disks directly. In this paper, we argue that it is feasible to create efficient disk-based learned indexes by applying a set of general transformations and optimizations to existing in-memory ones. Through theoretical analysis and controlled experiments, we propose six transformation guidelines applicable to various state-of-the-art learned index structures to fully leverage the characteristics of disk storage. Our evaluation shows that the indexes developed by applying our guidelines achieve a Pareto improvement in both throughput and space efficiency compared to the traditional B+tree and previous implementations of disk-based learned indexes.