AWS Certified Machine Learning Engineer - Associate MLA-C01 Practice Test 3
Question 20
A retailer needs predictions on an S3 archive of 40 million records once per month. There is no interactive application, and keeping a model-serving endpoint active between jobs is unnecessary. Which deployment mode is most suitable?
- A permanently provisioned real-time SageMaker endpoint.
- SageMaker Batch Transform on the available input objects.
- An asynchronous endpoint that queues every row independently all month.
- A serverless endpoint invoked manually once per record in a tight single-thread loop.
Explanation
Correct Answer
B:Batch Transform is designed for large offline datasets when no persistent serving endpoint is required.
Incorrect Answers
A:A persistent endpoint would incur unnecessary serving overhead when predictions are only generated as an offline job.
C:An async endpoint is useful for queued inference requests, not a natural replacement for a single bulk offline transform job.
D:Per-record serverless calls add invocation complexity and are not the straightforward bulk S3 inference mechanism.
Question 21
A small application receives unpredictable bursts of requests separated by long idle periods. Each payload is 100 KB and takes a few hundred milliseconds; there is no GPU requirement. The team wants to avoid managing provisioned instances. Which option is most aligned?
- Serverless Inference for short, intermittent requests within its supported limits.
- Batch Transform because the caller requires a synchronous response.
- A persistent GPU real-time endpoint with a fixed minimum instance count of four.
- Asynchronous Inference solely because it always has lower response latency than serverless.
Explanation
Correct Answer
A:Serverless Inference manages capacity and charges for usage, making it appropriate for intermittent small, short-duration invocations.
Incorrect Answers
B:Batch Transform is offline and does not give an immediate interactive prediction per request.
C:A fixed multi-GPU endpoint would be costly during idle periods and is not justified by the workload size or duration.
D:Async inference involves queue processing and is not guaranteed to minimize per-request latency for fast synchronous interactions.
Question 22
A SageMaker asynchronous endpoint is configured with minimum capacity zero. The first job arrives while capacity is zero, but the usual scaling target is a backlog of ten requests and one queued request remains stuck. Which scaling addition specifically addresses this cold queue?
- Enable a HasBacklogWithoutCapacity policy so a nonempty queue can trigger capacity from zero.
- Set the model package approval status to PendingManualApproval.
- Increase only the training job MaxRuntimeInSeconds.
- Disable asynchronous request queuing for the model and lose incoming jobs.
Explanation
Correct Answer
A:The HasBacklogWithoutCapacity metric/policy addresses the no-capacity-plus-backlog condition even before the usual higher queue target is reached.
Incorrect Answers
B:Model approval status governs promotion of a model package; it does not change the inference endpoint’s queue scaling signal.
C:Training runtime limits do not control inference instance scale-out and cannot drain an async backlog.
D:Removing queues defeats the long-running request delivery behavior rather than restoring the missing inference capacity.
Question 23
An ML team repeatedly creates inference endpoints by clicking through a console wizard; test and production environments diverge in instance type and variant settings. What is the best change for reproducible infrastructure?
- Write and version an AWS CloudFormation or AWS CDK definition for endpoint-related infrastructure, with reviewed environment parameters.
- Add more operators to the console so all environments can be edited at the same time.
- Export endpoint invocation logs and treat them as the source of infrastructure settings.
- Only save the model weights to S3 and rely on operators to recreate hosting details from memory.
Explanation
Correct Answer
A:Versioned infrastructure-as-code definitions describe provisioned resources and configuration consistently across environments.
Incorrect Answers
B:More operators do not provide a stable specification; they can amplify drift when using manual point-and-click changes.
C:Inference logs describe traffic and events, not a full declarative source for deploying the correct endpoint resources.
D:Model weights alone omit the hosting instance count, networking, variants and other reproducibility-critical configuration.
Question 24
An asynchronous inference container handles only one request at a time even though it has multiple workers and spare CPU. The endpoint configuration shows MaxConcurrentInvocationsPerInstance set to 1. Which setting should the engineer evaluate first?
- Increase the asynchronous endpoint concurrency limit after verifying the model server can safely process parallel requests.
- Lower the model-training learning rate to reduce inference queue delays.
- Increase the S3 artifact object retention period for the deployed model.
- Remove the SageMaker execution role from the model to avoid endpoint authorization overhead.
Explanation
Correct Answer
A:The configured async concurrent-invocation limit caps how many requests reach each inference container, regardless of internal worker count.
Incorrect Answers
B:The training learning rate controls optimization during model fitting, not invocation dispatch concurrency for a deployed endpoint.
C:S3 artifact retention does not raise the request-concurrency limit enforced for the serving container.
D:Removing the execution role can break resource access and does not provide a supported method of increasing request concurrency.
Question 25
A fraud model training pipeline must reject a candidate if validation AUC falls below 0.88 and must avoid registering an unacceptable model. Which SageMaker Pipelines design implements this automatically?
- Register every trained model first and wait for users to discover which version scored poorly.
- Add a condition step reading the evaluation report, and register the model only on the passing branch.
- Send the training data to a second real-time endpoint and let its latency determine model accuracy.
- Omit the evaluation job and promote the version with the largest model artifact.
Explanation
Correct Answer
B:A Pipelines condition step can branch on evaluation output and route only acceptable candidates into a registration step.
Incorrect Answers
A:Registering everything first ignores the specified pre-registration quality gate and creates unacceptable candidate packages.
C:Endpoint latency measures serving behavior; it is not the stated held-out validation AUC threshold.
D:Model file size is not evidence of classification quality and cannot replace the objective evaluation condition.
Question 26
A model has passed an offline validation suite but cannot be deployed until a release manager approves its version. The CI/CD system watches SageMaker Model Registry statuses. Which event should it use as the authorized promotion signal?
- A newly uploaded S3 object with a similar file name to a candidate model.
- An increase in the training dataset object count.
- The selected model package version changing to Approved after review.
- A lower average endpoint CPU utilization for the previous production model.
Explanation
Correct Answer
C:The Model Registry supports explicit model-version approval states, which can trigger authorized CI/CD deployment.
Incorrect Answers
A:Unreviewed S3 uploads do not enforce the model-version approval workflow required by the stated release policy.
B:Dataset counts do not indicate whether the particular model package has been approved for deployment.
D:Performance of the existing endpoint does not authorize promotion of an unrelated new candidate model.
Question 27
A training pipeline needs to run a Python data-preparation job, train a model, evaluate it, and then conditionally register a passing version. What orchestration design best expresses the dependencies?
- Independent cron jobs with no shared run identifier and no execution dependencies.
- A single long-running notebook that an engineer must execute manually after each source refresh.
- A SageMaker Pipeline with processing, training, evaluation, condition and model-registration steps.
- Four differently named endpoint variants with no upstream processing workflow.
Explanation
Correct Answer
C:A SageMaker Pipeline represents the dependent processing, training, evaluation and conditional registration operations as steps.
Incorrect Answers
A:Uncoordinated jobs cannot reliably ensure that evaluation and registration use the correct data and specific training run.
B:A manually started notebook lacks the managed execution graph and repeatable pipeline controls in the requirement.
D:Endpoint variants route serving traffic and do not establish an upstream data-processing and model-training dependency graph.
Question 28
A fraud-scoring endpoint already captures incoming feature values for SageMaker Model Monitor. The team wants an alert when a numeric input feature drifts from its training baseline. Which monitoring setup fits?
- Create baseline statistics and constraints, schedule data-quality monitoring against captured inference inputs, and inspect violations.
- Monitor only endpoint invocation count and latency, without inspecting captured feature values.
- Schedule model-quality monitoring using delayed fraud labels, but do not create any input-feature baseline.
- Compare only endpoint CPU utilization against its predeployment average.
Explanation
Correct Answer
A:Data quality monitoring compares captured production inputs against an established feature baseline and can report constraint violations.
Incorrect Answers
B:Request volume and latency reveal serving load, not whether input feature distributions have shifted.
C:Model-quality monitoring uses outcomes to assess prediction correctness; it does not replace input feature statistics and constraints for drift.
D:Resource utilization measures infrastructure behavior, not the distribution of the input feature.
Question 29
A lender wants to compute the real-world recall of a deployed fraud model, not merely the distribution of its input features. Fraud outcomes become known three weeks after each prediction. What additional information is essential?
- Only production input-feature histograms and data-quality distribution-violation counts, with no confirmed fraud outcomes.
- The eventual actual fraud outcomes matched to each earlier production prediction after the three-week labeling delay.
- Only predicted fraud probabilities and the deployed decision threshold, without any outcome labels.
- Only the offline training-validation confusion matrix from before deployment, with no subsequent production labels.
Explanation
Correct Answer
B:Production recall is TP/(TP+FN), so the delayed ground-truth fraud results must be joined to their corresponding deployment-time predictions to identify missed fraud.
Incorrect Answers
A:Input-feature drift measures whether traffic changes; it cannot count production false negatives without the actual fraud outcome for each case.
C:Predicted scores and thresholds describe what the model estimated, not which real fraud events it missed; they cannot establish recall without outcomes.
D:A training-validation confusion matrix estimates pre-deployment behavior, not the actual false negatives and positives after deployment and evolving traffic.
Question 30
A financial institution uses SageMaker Clarify with a deployed lending model. Auditors need to compare the contributions of income and credit history to individual loan scores. Which Clarify output fits?
- Endpoint request count and p95 model latency for the scoring period.
- A group-level fairness report comparing approval rates, without individual feature contributions.
- SHAP-based feature attribution values for the scored records, interpreted relative to a documented baseline.
- A single global feature-importance ranking averaged across the entire evaluation dataset.
Explanation
Correct Answer
C:SHAP attributions quantify contributions of model inputs to predictions relative to a reference baseline, matching the explanation need.
Incorrect Answers
A:Serving throughput and latency describe operational performance, not the variables behind a particular score.
B:Group fairness metrics can expose disparities but do not explain an individual record’s predicted score.
D:Global importance summarizes behavior across many records; it does not identify contributions for the specific loans under review.
