AWS Certified Machine Learning Engineer - Associate MLA-C01 Practice Test 37
Question 360
An endpoint is only used for a short daily burst. The team considers SageMaker Serverless Inference but relies on very large GPU model weights, and the model requires a GPU for acceptable latency. Which consideration makes this choice unsuitable?
- Changing the SageMaker Model Registry approval state enables serverless GPU instances.
- Serverless Inference does not offer GPU-backed execution, so evaluate compatible real-time or asynchronous GPU hosting instead.
- Serverless Inference always has a larger supported payload than asynchronous inference.
- Serverless Inference can keep a dedicated GPU warmed continuously at no charge.
Explanation
Correct Answer
B:GPU-bound workloads need a hosting mode that exposes supported GPU instances, unlike the serverless inference mode.
Incorrect Answers
A:Registry governance cannot add an unsupported hardware class to the hosting service.
C:Payload limits alone do not supply the required GPU compute or establish suitability.
D:There is no supported always-on dedicated GPU in the serverless inference offering.
Question 361
A data-science group wants to compare two real-time endpoint configurations for cost per thousand predictions while requiring p99 latency below a fixed SLO. What minimum experimental setup would make the comparison fair?
- Use the same representative request distribution, concurrency, model artifact and observation window, then report latency percentiles plus actual compute costs.
- Use one endpoint under peak traffic and the other under quiet off-hours, then compare mean latency only.
- Assume that the configuration with the cheapest instance-hour will always have the lowest total serving cost.
- Compare only the package version labels from Model Registry.
Explanation
Correct Answer
A:Controlled load testing isolates hosting differences and prevents traffic-mix changes from confounding price-performance.
Incorrect Answers
B:Different workloads and missing tail latency make the performance comparison misleading.
C:A slower instance might require more replicas or cause unacceptable SLA failures.
D:Registry labels do not measure latency or billable serving capacity.
Question 362
An AI research team launches hyperparameter tuning with hundreds of trial jobs even though several poor candidates clearly underperform after early epochs. What cost-aware control should be checked first?
- Increase every trial's maximum epoch count to ensure the full quota is spent.
- Remove validation reporting so the tuner cannot compare trials.
- Pick trials solely by training runtime, regardless of model quality.
- Use supported SageMaker automatic tuning early stopping based on consistently reported objective metrics.
Explanation
Correct Answer
D:Early stopping can end underperforming trials and reduce compute consumed by unlikely winning candidates.
Incorrect Answers
A:Extra epochs waste capacity on consistently weak candidates instead of reducing unnecessary trial cost.
B:Lack of objective metrics prevents evidence-based early termination.
C:Fastest duration alone is not the specified optimization objective and can select a poor model.
Question 363
A SageMaker Managed Spot training job has its instances interrupted regularly. Each restart wastes four hours of training despite significant cost savings from Spot prices. Which cost-control combination is appropriate?
- Always purchase the most expensive on-demand instance before investigating recoverability.
- Keep training the same model longer after each restart without saving its state.
- Persist reproducible checkpoints to S3 and choose interruption-aware retry/time limits for the training job.
- Disable checkpoints so every interruption restarts at epoch zero.
Explanation
Correct Answer
C:Checkpoint recovery retains model progress when capacity disappears, improving usable training per paid compute hour.
Incorrect Answers
A:Changing purchase model can avoid interruptions but ignores the requested cost-effective recovery control.
B:Longer runs do not preserve work lost when instances are reclaimed.
D:Rework increases total training work and can erase the savings from Spot pricing.
Question 364
An operational dashboard reports only the number of failed invocations per hour, while total request volume varies tenfold between day and night. Which normalization helps compare reliability fairly?
- Use only the maximum CPU clock rate as the error denominator.
- Calculate failure rate as failed invocations divided by valid total invocations over matched periods.
- Compare absolute failure counts without reporting overall traffic.
- Divide errors by the number of stored model packages.
Explanation
Correct Answer
B:A denominator based on actual requests separates worsening reliability from normal traffic growth.
Incorrect Answers
A:Compute frequency does not represent the invocation population.
C:Counts can rise with demand even when the probability of failure is constant.
D:Model registry inventory is unrelated to how many requests could have failed.
Question 365
A team receives an endpoint error alarm after one failed request during an interval containing only one invocation; the alarm also triggers continuously during traffic bursts that include occasional harmless retries. What is the better signal?
- Use a minimum request volume plus a rolling error-rate threshold aligned with the service SLO.
- Alert on every single 4xx or 5xx without regard to origin or sample size.
- Use only the number of running EC2 instances in the VPC.
- Raise the alarm threshold to an undefined infinity.
Explanation
Correct Answer
A:A rate with sufficient sample size reduces noise and reflects whether a meaningful fraction of requests fails.
Incorrect Answers
B:An unqualified count can page on client mistakes and statistically tiny samples.
C:Instance inventory cannot establish request success or a meaningful error budget.
D:A signal that can never fire fails to detect sustained real service failure.
Question 366
A model endpoint runs in two AWS Regions. Engineers want to compare customer-facing latency without mixing Region-level network distance with the model container's processing duration. What metric split is useful?
- Use only training job duration in both Regions.
- Assume a single global endpoint mean captures all regional behavior.
- Measure only the SageMaker ModelLatency metric and assume it already includes all client-to-Region network time.
- Report ModelLatency and OverheadLatency separately, and correlate them with end-to-end client measurements by Region.
Explanation
Correct Answer
D:Separating model processing from service overhead helps identify where time is spent before accounting for client network transit.
Incorrect Answers
A:Training duration cannot describe individual production invocation latency.
B:Aggregating Regions conceals distinct latency populations and network paths.
C:ModelLatency describes model-side processing, so it cannot on its own explain client network transit differences across Regions.
Question 367
During a compliance review, engineers find that an ML endpoint's request logs include full customer financial records even though performance monitoring needs only duration, status and a request correlation ID. What is a better instrumentation design?
- Turn off all request correlation information to avoid retaining any evidence.
- Retain encrypted complete request bodies for every inference call, even when only latency telemetry is required.
- Record minimal structured operational fields and protect any necessary sensitive logs under approved retention and access controls.
- Enable full raw feature logging for every request indefinitely.
Explanation
Correct Answer
C:Least-data operational telemetry preserves diagnosability without copying unnecessary private feature content.
Incorrect Answers
A:Removing all identifiers makes incident investigation and per-request tracing unnecessarily difficult.
B:Encryption reduces exposure but does not justify unnecessary collection of complete financial inputs for routine performance metrics.
D:Overcollection increases privacy and storage risk without being required for latency analysis.
Question 368
CloudWatch log retention for short-lived processing jobs is set to 'Never expire', and storage charges keep increasing. The organization must retain these diagnostic logs for 90 days. What adjustment fits?
- Reduce the training model's batch size to change logging retention.
- Set a 90-day retention policy for the appropriate log groups after validating legal and operational exceptions.
- Delete every new log event immediately after ingestion.
- Keep indefinite retention for every test log group regardless of policy.
Explanation
Correct Answer
B:A defined retention period stops routine diagnostics growing without limit while meeting the specified policy.
Incorrect Answers
A:Model batch size is unrelated to CloudWatch Logs retention.
C:Immediate deletion conflicts with the mandated 90-day ability to investigate issues.
D:Unbounded retention continues the avoidable storage cost.
Question 369
A production SageMaker endpoint serves two models as separate variants; one gets 95% of traffic and one 5% for a controlled test. Billing is high despite low traffic to the test model. Which resource-level detail matters?
- Each variant's configured instance capacity is provisioned, even when its traffic weight is small.
- The 5% traffic weight guarantees that its instance cost is only 5% of the primary variant's.
- Evaluation metric thresholds define each variant's instance-hour price.
- Removing training data from S3 will stop costs for the currently provisioned test variant.
Explanation
Correct Answer
A:Traffic routing percentages and allocated instance capacity are separate controls, so a low-weight variant can still incur instance-hours.
Incorrect Answers
B:Weighted routing does not automatically divide the infrastructure invoice in matching proportions.
C:Model accuracy thresholds do not set the host instance rate or capacity.
D:Stored training input does not determine whether a deployed variant has active replicas.
Question 370
A company has a batch of image predictions every quarter and wants to optimize both batch runtime and spending. One instance family is faster per hour but substantially more expensive. How should the team compare options?
- Choose only the instance with the lowest published hourly rate without testing throughput.
- Choose only the highest network throughput instance without measuring image model runtime.
- Average the two hourly prices and choose whichever matches the mean.
- Compare end-to-end cost per completed batch including total job duration and parallelism, while enforcing the batch completion deadline.
Explanation
Correct Answer
D:The fastest instance-hour and the lowest total workload cost need not coincide; measure the complete job.
Incorrect Answers
A:A cheaper instance may take enough extra time to cost more overall or miss the deadline.
B:Unused bandwidth does not establish inference price-performance for this workload.
C:The arithmetic midpoint does not describe the cost or speed of either complete workflow.
