PRODUCTION SCENARIO
A company computes aggregate customer features in a nightly AWS Glue job for training and recomputes them in application code at inference. The two implementations have drifted. Two other teams want to reuse the same features, and the fraud endpoint reads them on the request path.
Which approach keeps the definitions consistent and serves the endpoint with the LEAST ongoing maintenance?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Ingest the features into SageMaker Feature Store and read the online store at inference
Feature Store computes each feature once, serves the online store at millisecond latency on the request path and the offline store for training, and makes the definitions discoverable to the other teams. A shared library and a Processing job both leave two runtimes to keep in step, and transforming inside the inference container fixes one endpoint and no one else's models.
Official Source: https://docs.aws.amazon.com/sagemaker/latest/dg/feature-store.html ↗