A bank ingests raw transaction data from an Amazon Kinesis data stream. The solution must compute rolling averages from the stream, store those features in SageMaker Feature Store, and serve them to models in near real time. Which design meets these requirements?

The correct answer leverages Kinesis Data Analytics for real-time stream processing and Lambda for integration. Kinesis Data Analytics SQL applications are designed for continuous queries on streaming data, making them ideal for computing rolling averages in near real-time. The resulting stream contains the computed averages, which a Lambda function can then efficiently ingest into SageMaker Feature Store as an online feature group for low-latency retrieval by models. The other options are less suitable: Sending to S3 via Firehose and then using SageMaker Processing introduces latency and is not a near real-time solution for continuous aggregation. Feature Store's GetRecord API is for retrieving existing features, not for computing aggregates within the store itself. Storing as an offline feature group and computing at query time does not meet the "near real-time" requirement for serving.