MLA-C01 Practice Question 38: Serverless Intermittent Traffic
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 4 · question 8 of 10
An HR team uses an internal tool that calls a small scikit-learn model (about 300 MB) to suggest a job category for each new posting. Recruiters use it a few dozen times a day, in unpredictable bursts, and often not at all for several days. Each request carries a small JSON payload and needs an answer while the recruiter waits, though an occasional delay of a few seconds is acceptable. The model currently runs on an ml.m5.large real-time endpoint that sits idle almost all the time.
Which deployment option minimizes cost while still meeting these requirements?
Show answer and explanation
Answer:
Serverless inference is built for traffic with idle periods between bursts where cold starts are acceptable. It provisions compute on demand, scales to zero when there are no requests, and bills only for the compute used to process requests, so days without traffic cost nothing. A 300 MB CPU-only model with small payloads fits comfortably within its limits (up to 6 GB of memory, no GPU needed), and the recruiters' tolerance for a few seconds' delay covers the cold start (option C).
- A is wrong because a smaller instance lowers the hourly rate but still bills every hour of every day, including the many days with no requests. With traffic this sparse, paying per request costs far less than paying for any instance that is always on. - B is wrong because asynchronous inference queues requests and writes results to Amazon S3 for the caller to collect later. Scaling up from zero instances takes minutes, and the workflow does not suit a recruiter waiting on screen for an answer. - D is wrong because each batch transform job has to launch instances before it can run, which takes minutes. It is built for large offline datasets, not one posting at a time while a person waits.
Reference: Deploy models with Amazon SageMaker Serverless Inference