FinCorp runs a fraud-detection pipeline in AWS. You created multiple SageMaker Feature Groups with offline store enabled to persist historical feature values to S3. Data scientists need to generate training datasets that join transaction records (CSV in S3 with a transaction_timestamp) to the correct historical feature values as of each transaction time using SQL in Athena. Which configuration and step will reliably enable performant, correct joins on historical feature timestamps for large training exports?

This option is correct because enabling the offline store with DataCatalogConfig (Glue table) for each FeatureGroup allows Athena to directly query the historical feature data stored in S3. Specifying recordidentifiercolumn and eventtimefeaturename during FeatureGroup creation is crucial for correct point-in-time joins, as it defines how records are uniquely identified and timestamped. The Athena SQL join condition eventtime <= transactiontimestamp correctly retrieves features as they appeared at each transaction time, and partition pruning optimizes performance for large datasets. Incorrect options: Using only the online store and GetRecord calls for training data generation is inefficient and costly for large datasets due to individual API calls. SageMaker Batch Transform calling the featurestore-runtime API is designed for inference, not for generating large historical training datasets, and would be complex and inefficient for this use case. Relying solely on a Glue crawler without DataCatalogConfig during FeatureGroup creation is less reliable. The crawler might not correctly infer the specific eventtime partitioning schema used by Feature Store, leading to incorrect or inefficient joins.