MLA-C01 Practice Question 31: Streaming Feature Ingestion
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 4 · question 1 of 10
A payments company scores every card transaction with a fraud model hosted on a SageMaker AI real-time endpoint. The model needs a feature called txn_count_5min, the number of transactions on the same card in the previous five minutes. Transactions already arrive on an Amazon Kinesis data stream. The feature must be available to the scoring service within a few seconds of each transaction, and it is served from a feature group that has the online store enabled.
Which ingestion design meets the freshness requirement?
Show answer and explanation
Answer:
A rolling five-minute count is a *streaming feature*: it has to be recomputed continuously as events arrive. Managed Service for Apache Flink runs stateful stream processing, including windowed aggregations, directly on the Kinesis stream. Feature Store's PutRecord API is built for this case. It takes one record at a time from a streaming source and writes it straight to the online store, where the endpoint can look it up with low latency (option B).
- A is wrong because Firehose buffers records before it delivers them, and a Glue job every 15 minutes adds more delay on top. The feature would be minutes out of date, which defeats a five-minute window used for fraud detection. - C is wrong because the offline store is append-only historical data in Amazon S3, meant for building training sets. An Athena query per transaction is far too slow for synchronous scoring, and data reaches the offline store with its own delay. - D is wrong because an hourly export followed by a batch job leaves the count up to an hour stale. Batch ingestion suits features that change slowly, not a five-minute rolling count.
Reference: Data sources and ingestion - Amazon SageMaker Feature Store