Building Real-Time Feature Stores: Architecture Patterns and Trade-offs at Scale

Abstract

Feature stores have emerged as a critical component of production machine learning infrastructure, providing a layer of abstraction between raw data and model training and serving that addresses the dual challenges of feature reuse and online/offline consistency. Despite their growing adoption, the design space for feature stores is poorly mapped in the academic literature, with most detailed discussions confined to engineering blog posts and conference talks from technology companies. This technical communication systematises the feature store design space through a taxonomy of five architectural patterns — Lambda architecture, Kappa architecture, pre-computed offline-only, on-demand online-only, and hybrid push-pull — and provides quantitative comparisons of each pattern across four operational dimensions: latency, throughput, freshness, and operational complexity. We describe our experience migrating from a Lambda-based feature store to a hybrid push-pull architecture serving over 800 models and 15,000 feature definitions at Rakuten Group, report on the specific engineering trade-offs encountered, and provide a decision framework for architecture selection based on feature type and serving latency requirements.

Keywords: feature store; MLOps; real-time ML; Lambda architecture; Kappa architecture; feature serving; ML infrastructure