Question

A fraud detection team maintains a SageMaker Feature Store feature group with both online and offline stores enabled. A real-time transaction service must write newly computed features (e.g., rolling transaction counts) so they are immediately available for low-latency inference lookups, while also being automatically persisted for future model retraining. Which approach satisfies both requirements with the least custom engineering?

Accepted Answer

Call the PutRecord API on the feature group from the transaction service; the online store is updated immediately and records are automatically buffered and written to the offline store.

Answer

Write the features directly to the offline S3 store as Parquet files and run a scheduled Glue job to sync them into the online store.

Answer

Use a Glue ETL job triggered per transaction to batch-ingest the features into the feature group's offline store, then query it with Athena at inference time.

Answer

Write records only to the online store with PutRecord and schedule a nightly export job that copies the online store contents to S3 for retraining.