Feature stores
A feature store computes and serves the inputs a model reads, the same way in training and in production. These explainers cover what it is, how online and offline features differ, where a data warehouse stops, and why the freshness of a feature decides what it can act on.
In this topic
What is a feature store?
A feature store is the system that computes and serves the features a model uses, in training and in production, from one definition. It exists so the values a model learns from match the values it reads when it runs.
Online vs offline features
Offline features are computed in bulk over historical data and used to train a model. Online features are the current values served to that model when it runs. The two describe the same thing at different points in time.
What is training-serving skew?
Training-serving skew is the gap between the features a model was trained on and the features it reads in production. It makes a model perform worse than it did in testing, with no visible failure.
Feature store vs data warehouse
A data warehouse holds large volumes of history and answers analytical queries. A feature store serves individual feature values to a model with low latency. A warehouse is where offline features are computed; it is not where online features are served.
What is an online feature store?
An online feature store is the part that serves fresh feature values to a model at request time. It holds the current value of each feature, ready to read within the latency budget of a live prediction.
What is feature freshness?
Feature freshness is the lag between something happening and the served feature reflecting it. The fresher a feature, the more recent the behavior a model can act on. Stale features make a model answer for a customer who has already moved on.
Do you need a feature store?
Not every team needs a feature store. The case for one grows with the number of models and features you run, the need to reuse features across them, and the cost of training-serving skew. For a single simple model, it can be overkill.
Start with the free tier
Read the ideas, then put them to work: add the SDK, define one attribute, and see an attribute updating against your own traffic in the first session. The free tier runs on the real engine: no card, no sales call.