Feature Store

A feature store is a centralised system for managing the features — the input variables — that machine learning models use for both training and prediction. It acts as a shared repository where data teams can store, discover, and reuse pre-computed features across different models and projects.

The problem feature stores solve

In machine learning, features are the processed data inputs that models learn from. For example, a customer churn model might use features like "average monthly spend," "days since last purchase," and "number of support tickets." Creating these features requires data engineering work — querying databases, aggregating data, handling missing values, and computing derived metrics.

Without a feature store, this work is duplicated across teams and projects. The marketing team computes "average monthly spend" one way; the finance team computes it differently. The features used to train a model differ subtly from those used in production, causing "training-serving skew" — a major source of production failures.

How a feature store works

A feature store typically provides:

• Feature registry: A catalogue of all available features with metadata — who created them, what they mean, how they are computed.
• Offline store: Stores historical feature values for training models. Answers the question: "What was this customer's average monthly spend six months ago?"
• Online store: Serves feature values in real time for production predictions. Answers the question: "What is this customer's average monthly spend right now?"
• Feature pipelines: Automated processes that compute and update features on a schedule.

Why consistency matters

The most critical function of a feature store is ensuring that the exact same feature definitions are used during training and production. If a model learns that "average monthly spend" means the mean of the last 12 months during training, but receives the mean of the last 6 months in production, its predictions will be unreliable.

Popular feature stores

• Feast: An open-source feature store that integrates with common data infrastructure.
• Tecton: A managed feature store built by the creators of Uber's Michelangelo ML platform.
• Databricks Feature Store: Integrated with the Databricks lakehouse platform.
• AWS SageMaker Feature Store: Amazon's managed offering for AWS users.

When you need a feature store

Feature stores are most valuable when:

• Multiple teams build models that share common features
• Models require real-time predictions with consistent feature computation
• Feature engineering represents a significant portion of development time
• Training-serving skew has caused production issues