A shared place where the calculated inputs to models live, so that the same definition is used when training and when running live. The classic failure without one is that the training pipeline and the production system compute ‘customer tenure’ slightly differently, and nobody notices until performance is poor.
Technical view
Provides an offline store for training data with point-in-time correctness, ensuring features reflect only information available at the historical prediction moment, and an online store for low-latency serving, both generated from a single definition to prevent training-serving skew. Includes versioning, backfill and monitoring per feature. It is heavy infrastructure and warranted mainly where many models share features or point-in-time correctness is difficult to get right by hand.
Related
Part of AI Terminology Index, section: Running It in Production.