What is a feature store?
A feature store is a system that defines, computes, stores, and serves the features that machine learning models use as inputs. Each feature, such as a customer’s order count over the last 30 days, is defined once, with an owner, a description, and its source tables. An offline store keeps the full history of feature values for training. An online store holds the latest value of each feature for fast lookup during real-time inference.
The offline history is point-in-time correct, so a training example from a given day sees only the values known on that day. This prevents information leakage, where a model learns from facts that arrived after the event it predicts.
Why a feature store matters
A feature store prevents training-serving skew, where a feature is computed one way for training and another way in production. It also lets teams reuse features across models instead of repeating the feature engineering for each project.
Feature store in practice
An online retailer saw its checkout fraud model perform worse live than in testing. The data team’s return-rate feature excluded exchanges, while the checkout application’s version counted them. Once the fraud model and the weekly churn model both read return rate from the feature store, one definition served training and live scoring alike.
In Kirk Tech Solutions engagements, features usually begin as documented tables in the client’s lakehouse. An online store is added only when a model has to score requests in real time.