Benchmarking Learned Indexes

Ryan Marcus, Andreas Kipf, Alexander van Renen, Mihail Stoian, Sanchit Misra, Alfons Kemper, Thomas Neumann, Tim Kraska

Abstract: Recent advancements in learned index structures propose replacingexisting index structures, like B-Trees, with approximate learnedmodels. In this work, we present a unified benchmark that com-pares well-tuned implementations of three learned index structuresagainst several state-of-the-art “traditional” baselines. Using fourreal-world datasets, we demonstrate that learned index structurescan indeed outperform non-learned indexes in read-only in-memoryworkloads over a dense array. We investigate the impact of caching,pipelining, dataset size, and key size. We study the performanceprofile of learned index structures, and build an explanation for whylearned models achieve such good performance. Finally, we investi-gate other important properties of learned index structures, such astheir performance in multi-threaded systems and their build times.