PRODUCTION SCENARIO
A fraud endpoint must read the most recent value of each customer feature in single-digit milliseconds. The data science team needs the full history of those same features to build training datasets. The feature group is configured with both an online store and an offline store.
Which read path meets both requirements?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Read the online store at inference and the offline store for training
The online store retains only the latest record per identifier and is built for millisecond reads, so it belongs on the inference path. The offline store is an append-only history in Amazon S3, so it belongs on training. Reversing them puts an Athena-speed query in front of a fraud decision, online-only leaves training with no history, and offline-only misses the latency budget.
Official Source: https://docs.aws.amazon.com/sagemaker/latest/dg/feature-store.html ↗