MLA-C01 Practice Question 40: Inference Recommender Load Test
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 4 · question 10 of 10
A team registered a new model version in the SageMaker Model Registry and must choose the lowest-cost instance type and count for its real-time endpoint. The product requirements are specific: P99 model latency must stay under 120 ms at a peak of 18,000 invocations per minute, and traffic ramps up in steps each morning. The platform team supports only three instance types, so the choice must come from those.
How should the team get a recommendation that reflects these requirements?
Show answer and explanation
Answer:
Inference Recommender has two job types. A default job runs a standard set of load tests, finishes in under an hour, and gives an initial recommendation. An *advanced* job (a custom load test) is built for concrete production requirements. You choose up to 10 instance types to test, define the traffic pattern (phases, or stairs that add users at intervals), and set stopping conditions such as MaxInvocations and ModelLatencyThresholds with a percentile like P99. The results give an instance type, the instance count needed to meet the invocation target, and the cost per hour for each configuration, which answers the question directly (option D).
- A is wrong because a default job runs a standard load test. It cannot reproduce the stepped morning ramp or plan for a peak of 18,000 invocations per minute, so the cheapest configuration from a generic benchmark is not evidence that it holds P99 under 120 ms at that peak. - B is wrong because, for SageMaker AI endpoints, AWS Compute Optimizer gives only *idle* recommendations, flagging endpoints that are unused. It does not compare instance types against a P99 latency target at a given invocation rate. The plan would also expose real customers to weeks of latency experiments. - C is wrong because starting from the largest instance is a guess, not a recommendation. Auto scaling adds capacity but does not tell the team which instance type meets P99 at the lowest cost.
Reference: Run a custom load test - Amazon SageMaker Inference Recommender