Feature Store: Definition and Examples

A Feature Store is a platform dedicated to managing the lifecycle of features in machine learning. Features, also called predictive variables, are the transformed data that feed AI models. Without a Feature Store, each data team recalculates its own features in isolation, leading to duplication, inconsistencies, and considerable waste of time and resources.

The Feature Store solves this problem by offering a centralized catalog where features are defined, computed, stored, and served uniformly. It ensures consistency between the training environment (offline) and the production environment (online), eliminating the infamous "training-serving skew" problem that can silently degrade a model's performance in production.

Concretely, a Feature Store manages two types of storage: an offline store (often based on a data lake or warehouse) for training on historical data, and an online store (low-latency database like Redis or DynamoDB) for serving features in real-time during inference. Automated pipelines synchronize the two environments.

Solutions like Feast (open source), Tecton, Hopsworks, or the Feature Stores integrated into cloud platforms (Vertex AI Feature Store from Google, SageMaker Feature Store from AWS) have popularized this approach. The Feature Store is today considered an essential component of any mature MLOps infrastructure, alongside model versioning or production monitoring.