Managing features and building datasets with SageMaker Feature Store
Until now, we've engineered our training and validation features in a notebook or in a SageMaker Processing script, before storing them as S3 objects. Then, we used these objects as-is to train and evaluate models. This is a perfectly reasonable workflow. However, the following questions may arise as your machine learning workflows grow and mature:
- How can we apply a well-defined schema to our features?
- How can we select a subset of our features to build different datasets?
- How can we store and manage different feature versions?
- How can we discover and reuse feature engineering by other teams?
- How can we access engineered features at prediction time?
SageMaker Feature Store is designed to answer these questions. Let's add it to the classification training workflow we built with BlazingText and Amazon Reviews in Chapter 6, Training Natural Language Processing Models.
