Amazon SageMaker Feature Store
A central, consistent repository for ML features.
❓ What is it?
A purpose-built store for ML features with an online store (low-latency reads at inference) and an offline store (historical data in S3 for training), keeping both in sync from one ingestion path.
💡 Why does it exist?
Training/serving skew — computing a feature one way in the training pipeline and another way in production — is a classic source of invisible model bugs. A shared feature store makes features computed once, reused everywhere, and discoverable across teams.
⏱️ When should you use it?
Adopt it when multiple models or teams reuse the same features, or when a real-time model needs millisecond feature lookups that must exactly match what it was trained on.
🗺️ Where does it fit?
Between your data pipelines and your models: ingestion writes feature groups; training jobs read point-in-time-correct data from the offline store; endpoints read the freshest values from the online store.
🔌 How do you integrate it?
Define a feature group (schema + record identifier + event time), ingest via the SDK or streaming, query the offline store with Athena for training sets, and call GetRecord for online inference lookups.
🧩 Commonly integrated with
🎯 Exam angle (AIF-C01)
- Online store = real-time inference reads; offline store = training/batch. Questions test which store serves which phase.
- It is the fix when a scenario mentions inconsistent features between training and production.