HIRE

A Hybrid Learned Index for Robust and Efficient Performance under Mixed Workloads (SIGMOD '26)

HIRE: A Hybrid Learned Index for Robust and Efficient Performance under Mixed Workloads To appear at ACM SIGMOD 2026 · Xinyi Zhang, Liang Liang, Anastasia Ailamaki, Jianliang Xu

Motivation

Learned indexes substantially outperform traditional structures for point lookups by modeling the CDF of data. However, they frequently suffer from high tail latency, suboptimal range-query performance, and inconsistent effectiveness across diverse workloads — limiting their adoption in production systems.

Design

HIRE is a hybrid in-memory index that combines the structural robustness of traditional indexes with the predictive power of learned models:

  1. Hybrid leaf nodes that adapt to varying data distributions and workloads.
  2. Model-accelerated internal nodes augmented with log-based updates for efficient writes.
  3. A non-blocking, cost-driven recalibration mechanism for dynamic data.
  4. An inter-level bulk-loading algorithm that jointly accounts for leaf and internal-node errors.

Results

  • Up to 41.7× higher throughput under mixed workloads vs. SOTA learned / traditional indexes.
  • Up to 98% reduction in tail latency across varying scenarios.
  • Consistently robust range-query and overall performance.

[paper]