ML platform
SageMaker Feature Store
Stores, shares, and serves reusable ML features.
Key points
- Provides a managed repository for machine learning features.
- Supports reuse of consistent features across training and inference workflows.
- Helps reduce training-serving skew by using the same feature definitions.
- Supports online and offline feature access patterns for low-latency inference and training datasets.
- Improves collaboration by making curated features discoverable and governed.
When to use it
- Choose Feature Store when multiple models or teams need shared feature definitions.
- Use it when training and inference must use consistent feature values.
- Use it for real-time prediction systems that need low-latency feature lookup.
Exam tips
- Feature Store stores features; Data Wrangler creates and transforms features.
- A feature store is not a vector database for RAG embeddings in Bedrock Knowledge Bases.
- The online store supports inference use cases, while offline data supports training and analysis.
- Consistency between training and inference is a major exam clue.