Feature Store
Also known as: ML feature platform
A feature store is a centralized system for managing the features used by machine learning models — providing consistent definitions, training/serving parity, freshness guarantees, and reusability across teams.
Detailed explanation
Feature stores solve a recurring ML production problem: features computed for training drift from features computed at inference, models share features but redefine them, and freshness requirements vary by use case. A feature store provides a single definition of each feature with documented owner, schema, and freshness SLA.
Common implementations include open-source Feast and Hopsworks, plus managed offerings from Tecton, Databricks, Vertex AI, and SageMaker. The store typically maintains an offline store for training and an online store for low-latency serving, with backfill and consistency guarantees between them.
Feature stores are most valuable for teams running multiple models in production that share underlying signals. For a single-model team they are usually overkill — a well-versioned dbt project plus careful inference code goes a long way.