What is a feature store?

Definition: A feature store is a central system for defining, computing, storing and serving machine learning features, so the same feature values are used consistently in training and in production predictions.

How a feature store works

Features are model inputs derived from raw data, such as a customer's transaction count in the last 30 days. A feature store keeps an offline store of historical values for training and an online store with low-latency lookups for real-time inference. Examples include Feast, Tecton and the feature stores in Databricks, SageMaker and Vertex AI.

Why it matters

When training and serving compute features with different code, models see different data in production than in training, a problem known as training-serving skew. A feature store prevents that, produces point-in-time correct training sets, and lets teams reuse features across models.

When to use one

It pays off when several models share features or need real-time predictions, such as fraud detection or recommendations. A single batch model may not need one.

How Ryz helps

Ryz Databricks engineers and MLOps engineers build feature pipelines on your data platform, working on your team.

Related

• MLOps services
• Hire Databricks engineers
• MLOps
• Data pipeline