Feature Stores: A Hierarchy of Needs

Eugene Yan · 5 years ago

Feature stores are becoming increasingly common, with AWS launching SageMaker Feature Store in December 2020 and other platforms following, as managing features remains a significant bottleneck in productionizing ML models. The article organizes feature store functionality into a five-level hierarchy of needs: access (reducing duplication and reusability), serving (real-time feature deployment), integrity (consistency between training and production), convenience (ease of use), and autopilot (automation). Companies like Uber, Airbnb, and GoJek have built feature stores to address these needs, with the first two levels (access and serving) and parts of the third level (train-serve skew) providing the most immediate value for most organizations.