Indexing meets machine learning: a systematic literature review of learned indexes
TL;DR
A persistent research-practice gap is identified in which algorithmic innovation has outpaced system-level integration, underscoring the infrastructural work still required for production adoption and contributing a comprehensive architectural taxonomy, an evidence-based performance framework, and a prioritized research roadmap.
Abstract
Learned index structures use machine learning models, rather than traditional algorithmic data structures, to optimize database indexing performance. This systematic literature review applies the PICOC framework to 49 peer-reviewed articles (2017–2025) on learned index effectiveness versus traditional methods, using systematic multi-venue searches and strict, empirically grounded inclusion criteria. Three evolutionary phases emerge from the corpus: foundational development (2017–2018), rapid diversification (2019–2022), and systematic optimization (2023–2025). In terms of performance, learned indexes typically achieve 1.4–5× lookup speedups on standard CPU platforms. In contrast, substantially larger improvements (up to 174×) have only been reported under GPU-parallel batch inference on benchmark datasets. Likewise, reported memory reductions and training-time improvements depend on the specific workload, data distribution, and evaluation environment. Despite this progress, critical gaps remain in dynamic workload support, database integration, and reliability guarantees. Learned indexes perform best in read-heavy analytic and multi-dimensional workloads but degrade substantially under write-intensive conditions. This review contributes a comprehensive architectural taxonomy, an evidence-based performance framework, and a prioritized research roadmap, and in doing so identifies a persistent research-practice gap in which algorithmic innovation has outpaced system-level integration, underscoring the infrastructural work still required for production adoption.