AIF-C01 Practice Question 22: Batch vs Real Time Inference
AWS Certified AI Practitioner (AIF-C01) · Test 3 · question 2 of 10
Every night an insurance company scores roughly 40 million policy records with a churn model. The results are only read the next morning by an analytics team, so no user is waiting on any individual prediction, and the company wants to avoid paying for compute during the 23 hours a day when nothing is being scored. Which inference option fits these requirements at the lowest cost?
Show answer and explanation
Answer:
Batch transform is the right fit when you need inferences from a large dataset, no persistent endpoint is required, and results can be written to Amazon S3 rather than returned in a request. SageMaker AI starts the compute instances, distributes the dataset across them, writes the output files, and then shuts the instances down, so you pay only for the length of the job, not for 24 hours of idle endpoint (option B).
- A is wrong because a real-time endpoint runs, and bills, continuously, including the 23 hours a day when nothing is being scored. Forty million individual HTTP calls is also a slow and expensive way to move a bulk workload. - C is wrong for the same reason, and makes it worse: a larger instance raises the hourly cost of an endpoint that is idle most of the day. Auto scaling reduces instances under low load but does not remove the always-on baseline. - D is wrong because serverless inference is designed for intermittent, low-volume, latency-sensitive traffic. Invoking it 40 million times in one window is a bulk job wearing a request/response shape, and it does not beat batch transform on cost here.
Reference: Batch transform for inference with Amazon SageMaker AI