A feature store is a system that centralizes, stores and serves the features (variables computed from raw data) used by machine learning models, both for training and for predictions in production.
A feature is, for example, “number of purchases by a customer over 30 days”. Without a feature store, each team recomputes it its own way, and the computation used for training often differs from the one used in production. The model then sees different data from what it learned on: this is training/serving skew, a classic source of silent degradation.
The feature store solves this with a single definition of each feature, historical (offline) storage for training and fast (online) storage to serve predictions within milliseconds.
Common solutions are Feast (open source), Hopsworks, Tecton, and the built-in feature stores of Databricks and Google Vertex AI. The tool is mostly justified when several models run in production.
Why it matters when hiring
It is a topic for a senior MLOps Engineer or ML Engineer. The good signal: a candidate who can explain a concrete case of mismatch between training and production and how they detected it. Be wary of a project that wants to set up a feature store before having a first model in production: it often means putting tooling before the need.
