AWS Certified Machine Learning Engineer - Associate MLA-C01 Practice Test 35
Question 340
A model uses numerical features from a seasonal retail system. Its drift monitor alarms every holiday even when this pattern occurred in past years. What improvement is reasonable?
- Train the model using only holiday records regardless of operating population.
- Use seasonally comparable baselines or segmented thresholds and validate the business tolerance for expected shifts.
- Disable monitoring permanently after one false-positive holiday alert.
- Set every feature threshold to infinity.
Explanation
Correct Answer
B:Season-aware baselines distinguish recurring patterns from unexpected changes.
Incorrect Answers
A:Training on one period may worsen performance outside that season.
C:Removing all detection loses coverage of genuine new shifts.
D:Unbounded thresholds cannot flag any meaningful deviation.
Question 341
An inference pipeline emits predictions and a separate fraud outcomes service supplies truth. Repeated outcome messages make recall calculations inconsistent. What control should the label join have?
- Deduplicate outcome events by stable prediction/outcome identity and apply a documented correction/version policy.
- Treat every retry as a new independent fraud event.
- Drop prediction IDs before outcomes are ingested.
- Use arrival order alone to match labels to predictions.
Explanation
Correct Answer
A:Idempotent label ingestion prevents double-counting truth events or stale corrections.
Incorrect Answers
B:Duplicate labels inflate denominators and change confusion-matrix counts.
C:Missing linkage keys prevents correct label assignment.
D:Out-of-order and duplicate messages break positional joins.
Question 342
A binary model's predicted positive rate stays stable after release, but delayed ground truth shows a higher false-positive fraction. What should be evaluated?
- Only the number of feature columns in the model package.
- Only the AWS Region where the endpoint runs.
- Only the image hash of the model server.
- Precision and calibration on the newly labeled cohort, not only the raw positive alert rate.
Explanation
Correct Answer
D:Unchanged volume of positive predictions cannot establish that those positives remain correct.
Incorrect Answers
A:Model schema size is not observed prediction correctness.
B:Region alone does not compute real alert quality.
C:Artifact identity cannot substitute for true outcomes.
Question 343
A real-time model has a training baseline showing 99% non-null for a customer-age feature. Live traffic has 70% nulls, but no endpoint errors because a fallback runs. What should alert?
- Only the current container memory footprint.
- Only the last model training epoch number.
- A completeness-drift signal that compares the live non-null fraction with the approved baseline.
- Only an HTTP 5xx count, which remains zero.
Explanation
Correct Answer
C:Successful fallback predictions do not remove the underlying loss of feature data.
Incorrect Answers
A:Container memory consumption cannot reveal the shift from 99% to 70% non-null customer-age inputs despite successful fallback scoring.
B:Training history cannot measure new input completeness.
D:The fallback can conceal data-quality degradation from error-only monitoring.
Question 344
A model drift alert is sent by email with no feature name, time range or baseline reference, and engineers cannot investigate it. What additional alert metadata is most useful?
- Replace the alert with a training-loss screenshot from months earlier.
- Include the feature, monitoring window, observed statistic, baseline threshold and report location.
- Only include the number of SageMaker endpoints in the account.
- Only include a generic message stating something happened.
Explanation
Correct Answer
B:Actionable drift alerts identify what changed, when it changed and where supporting evidence resides.
Incorrect Answers
A:Old training loss cannot explain the current feature deviation.
C:Fleet size does not explain which feature violated a constraint.
D:An unscoped message provides no diagnostic path.
Question 345
A SageMaker endpoint meets its mean latency target, but customers report intermittent slow responses at peak traffic. The team wants an alert reflecting the slowest five percent of requests rather than the mean. Which measurement should drive the alarm?
- The p95 ModelLatency or end-to-end latency statistic over an appropriate observation period.
- The average latency across the same interval with no percentile breakdown.
- The minimum request latency in each observation period.
- The number of training jobs created in the account that day.
Explanation
Correct Answer
A:A p95 percentile exposes a slow tail that an average can conceal when most requests are fast.
Incorrect Answers
B:The mean can remain acceptable even when a meaningful minority of calls is very slow.
C:The fastest request does not establish service quality at peak load.
D:Training-job volume does not measure the serving latency experienced by customers.
Question 346
An inference API occasionally receives zero requests for hours. Its CloudWatch alarm on high error rate currently enters ALARM whenever metric data is missing, paging engineers on healthy idle periods. What is the first alarm-design correction?
- Force every missing period to count as an application error regardless of traffic volume.
- Increase the endpoint's minimum instance count solely to generate metrics during idle time.
- Delete the alarm and never measure inference failures again.
- Choose a missing-data treatment consistent with no-traffic periods and require actual error evidence for paging.
Explanation
Correct Answer
D:For a rate metric with no traffic, missing observations need an explicit handling policy rather than automatically implying a failure.
Incorrect Answers
A:That policy treats absence of requests as a real service failure and reproduces false alarms.
B:Capacity changes do not fix the alarm's interpretation of missing rate data and add unnecessary cost.
C:Removing the monitor hides genuine problems instead of correcting its missing-data semantics.
Question 347
An asynchronous inference endpoint has a low average queue length, yet some individual document jobs wait forty minutes before processing. Which additional CloudWatch metric directly exposes this experience?
- The number of daily new SageMaker Studio notebooks.
- The mean training loss from the most recent model build.
- ApproximateAgeOfOldestRequest for the endpoint's queued work.
- The count of all model packages stored in the model registry.
Explanation
Correct Answer
C:The oldest-request age tracks wait time directly and can reveal starvation hidden by average queue size.
Incorrect Answers
A:Notebook provisioning does not represent queue delay for deployed async inference.
B:Training loss cannot report the time already queued inference requests have waited.
D:Package count is not a measure of waiting inference requests.
Question 348
A SageMaker asynchronous inference endpoint writes completed prediction artifacts into a dedicated S3 output prefix. Application clients need the outputs for at most 30 days, no compliance hold applies, and storage costs keep growing because old results remain indefinitely. Which targeted change meets the data-retention need and removes expired output objects?
- Enable S3 Intelligent-Tiering for the output prefix but keep all results indefinitely, with no object expiration rule.
- Create an S3 Lifecycle expiration rule for completed objects in the dedicated asynchronous-output prefix after 30 days.
- Configure only an S3 rule to abort incomplete multipart uploads, leaving all successfully written inference outputs untouched.
- Lower the endpoint maximum concurrent inference count while keeping every completed S3 prediction output without a deletion policy.
Explanation
Correct Answer
B:An S3 Lifecycle rule can expire objects in the scoped output prefix after the retention period; SageMaker async outputs are written to S3, so the rule addresses old result files.
Incorrect Answers
A:Intelligent-Tiering may optimize storage class charges but does not enforce the requested 30-day deletion boundary for completed inference outputs.
C:Aborting incomplete multipart uploads cleans failed in-progress transfers, not completed prediction outputs stored in the bucket.
D:Endpoint concurrency affects invocation processing capacity, not storage retention of completed output artifacts in S3.
Question 349
A business asks for a monthly cost report for each internal ML product, but hundreds of SageMaker training and processing jobs have no consistent project attribution. What should the platform team implement?
- Use enforced project/owner resource tags and activate applicable cost allocation tags for billing reports.
- Use only human-readable job names without a stable tagging policy.
- Assume that all SageMaker charges can be divided evenly across projects.
- Use a CloudWatch latency alarm as the sole accounting method.
Explanation
Correct Answer
A:Consistent cost allocation tags let the cost-management tools associate charges with the responsible products.
Incorrect Answers
B:Names can be inconsistent and do not establish a robust billing allocation dimension.
C:Even division ignores how much compute and storage each product consumed.
D:Latency observations do not assign cloud spend to business owners.
Question 350
A research team created a new managed endpoint and wants to be warned before this month's actual and forecast ML spending exceeds its approved budget. Which AWS capability best fits?
- A single endpoint Inference Recommender job, without any cost threshold or notification rule.
- Only a CloudWatch metric on model precision.
- An AWS Config rule that tests only whether all S3 buckets have versioning enabled.
- An AWS Budgets cost budget with configured actual/forecast threshold notifications.
Explanation
Correct Answer
D:AWS Budgets can compare usage-derived costs with planned limits and alert before or after threshold crossings.
Incorrect Answers
A:A performance recommendation cannot by itself enforce the monthly spending notification requirement.
B:Predictive accuracy does not measure total account expenditures or future budget variance.
C:S3 configuration compliance is not a spending forecast or budget alarm.
