Amazon AWS Certified Machine Learning Engineer - Associate MLA-C01 Practice Test 39
Question 380
As a regional data-pipeline rollout is validated against approved AWS storage and streams, the following condition is observed: A training job is dominated by extreme transaction-amount magnitudes, although most records have modest nonnegative amounts. The data team must preserve event-time meaning and lineage. Which observation most directly identifies the root cause of the reported behavior?
- Inspect event identifiers, object versions or ETags, processing ledger, and downstream record keys for duplicate deliveries.
- Inspect feature scale distributions, model sensitivity, and whether fitted scaler statistics are shared across serving and training.
- Inspect quantiles, heavy-tail plots, outliers, and the target model's response to feature scale.
- Inspect the stream retention setting, consumer checkpoint timestamps, and the oldest available sequence or time position.
Explanation
Correct Answer
C:Inspect quantiles, heavy-tail plots, outliers, and the target model's response to feature scale. This directly tests skewed numerical feature log transformation in the reported AWS ML workflow. A monotonic log transform can reduce heavy-tail influence, but invalid domains and train-serving parity must be handled. It distinguishes the problem before choosing a change. The related targeted response is: Apply a justified log1p transformation with documented treatment of invalid negatives and package it for serving. Reported signal: A training job is dominated by extreme transaction-amount magnitudes, although most records have modest nonnegative amounts. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
Incorrect Answers
A:Inspect event identifiers, object versions or ETags, processing ledger, and downstream record keys for duplicate deliveries. This is a possible diagnostic for s3 object event at-least-once processing, but it does not directly test the skewed numerical feature log transformation failure reported here. Prioritize: Inspect quantiles, heavy-tail plots, outliers, and the target model's response to feature scale. The relevant case is: A training job is dominated by extreme transaction-amount magnitudes, although most records have modest nonnegative amounts. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
B:Inspect feature scale distributions, model sensitivity, and whether fitted scaler statistics are shared across serving and training. This is a possible diagnostic for standardization of numerical features, but it does not directly test the skewed numerical feature log transformation failure reported here. Prioritize: Inspect quantiles, heavy-tail plots, outliers, and the target model's response to feature scale. The relevant case is: A training job is dominated by extreme transaction-amount magnitudes, although most records have modest nonnegative amounts. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
D:Inspect the stream retention setting, consumer checkpoint timestamps, and the oldest available sequence or time position. This is a possible diagnostic for kinesis consumer retention gap, but it does not directly test the skewed numerical feature log transformation failure reported here. Prioritize: Inspect quantiles, heavy-tail plots, outliers, and the target model's response to feature scale. The relevant case is: A training job is dominated by extreme transaction-amount magnitudes, although most records have modest nonnegative amounts. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
Question 381
During an approved rollout of an AWS ML inference workload, the following condition is observed: A document embedding application has payloads too large for its synchronous inference path and some requests need minutes of processing. The team must meet its documented availability and rollback constraints. Which engineering response resolves the actual issue without masking the symptoms?
- Compile the supported model for the specific edge target using SageMaker Neo or choose a compatible conversion route.
- Configure a justified endpoint variant auto scaling policy using a workload-relevant supported metric such as invocations per instance.
- Define the endpoint resources and parameters in versioned CloudFormation or CDK code, review a change set, and reconcile drift.
- Use SageMaker asynchronous inference with S3 input/output and suitable queue and notification settings.
Explanation
Correct Answer
D:Use SageMaker asynchronous inference with S3 input/output and suitable queue and notification settings. This is the appropriate controlled remedy for asynchronous endpoint for large payloads. Asynchronous inference decouples request submission from long-running processing and can work with larger payload contracts. Verify the cause with: Inspect payload sizes, model processing duration, caller timeout limits, and request concurrency. Then confirm: Large inputs complete without client request timeout and clients receive tracked result locations or failure notifications. Reported signal: A document embedding application has payloads too large for its synchronous inference path and some requests need minutes of processing. Operational setting: During an approved rollout of an AWS ML inference workload.
Incorrect Answers
A:Compile the supported model for the specific edge target using SageMaker Neo or choose a compatible conversion route. This addresses a different issue involving sagemaker neo edge compiler compatibility. It does not correct the stated asynchronous endpoint for large payloads mechanism. The targeted response is: Use SageMaker asynchronous inference with S3 input/output and suitable queue and notification settings. Case evidence: A document embedding application has payloads too large for its synchronous inference path and some requests need minutes of processing. Operational setting: During an approved rollout of an AWS ML inference workload.
B:Configure a justified endpoint variant auto scaling policy using a workload-relevant supported metric such as invocations per instance. This addresses a different issue involving sagemaker endpoint auto scaling metric selection. It does not correct the stated asynchronous endpoint for large payloads mechanism. The targeted response is: Use SageMaker asynchronous inference with S3 input/output and suitable queue and notification settings. Case evidence: A document embedding application has payloads too large for its synchronous inference path and some requests need minutes of processing. Operational setting: During an approved rollout of an AWS ML inference workload.
C:Define the endpoint resources and parameters in versioned CloudFormation or CDK code, review a change set, and reconcile drift. This addresses a different issue involving cloudformation stack parameter drift. It does not correct the stated asynchronous endpoint for large payloads mechanism. The targeted response is: Use SageMaker asynchronous inference with S3 input/output and suitable queue and notification settings. Case evidence: A document embedding application has payloads too large for its synchronous inference path and some requests need minutes of processing. Operational setting: During an approved rollout of an AWS ML inference workload.
Question 382
At a production AWS ML data lake supporting customer-facing predictions, the following condition is observed: A customer-support labeling dataset includes raw email addresses and account identifiers that external annotators do not need. Engineers must avoid changing unrelated source partitions. What should the ML engineer inspect FIRST to distinguish this failure from unrelated issues?
- Compare source schemas by partition, Glue Data Catalog types, and failed cast rows for the evolving field.
- Inspect AWS Glue Data Quality rule evaluations, missing-key rates, source partitions, and prior accepted baselines.
- Inspect source fields, sensitivity classification, Ground Truth task input, and downstream annotation permissions.
- Inspect quantiles, heavy-tail plots, outliers, and the target model's response to feature scale.
Explanation
Correct Answer
C:Inspect source fields, sensitivity classification, Ground Truth task input, and downstream annotation permissions. This directly tests pii masking before model labeling in the reported AWS ML workflow. Data minimization applies before external labeling; encryption in transit does not remove unnecessary sensitive content from the task. It distinguishes the problem before choosing a change. The related targeted response is: Remove or pseudonymize unnecessary PII before annotation and restrict access to the token mapping under separate controls. Reported signal: A customer-support labeling dataset includes raw email addresses and account identifiers that external annotators do not need. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Incorrect Answers
A:Compare source schemas by partition, Glue Data Catalog types, and failed cast rows for the evolving field. This is a possible diagnostic for glue schema evolution with mixed input types, but it does not directly test the pii masking before model labeling failure reported here. Prioritize: Inspect source fields, sensitivity classification, Ground Truth task input, and downstream annotation permissions. The relevant case is: A customer-support labeling dataset includes raw email addresses and account identifiers that external annotators do not need. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
B:Inspect AWS Glue Data Quality rule evaluations, missing-key rates, source partitions, and prior accepted baselines. This is a possible diagnostic for glue data quality non-null contract, but it does not directly test the pii masking before model labeling failure reported here. Prioritize: Inspect source fields, sensitivity classification, Ground Truth task input, and downstream annotation permissions. The relevant case is: A customer-support labeling dataset includes raw email addresses and account identifiers that external annotators do not need. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
D:Inspect quantiles, heavy-tail plots, outliers, and the target model's response to feature scale. This is a possible diagnostic for skewed numerical feature log transformation, but it does not directly test the pii masking before model labeling failure reported here. Prioritize: Inspect source fields, sensitivity classification, Ground Truth task input, and downstream annotation permissions. The relevant case is: A customer-support labeling dataset includes raw email addresses and account identifiers that external annotators do not need. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Question 383
During production evaluation of a SageMaker AI model candidate, the following condition is observed: A house-price model's RMSE increases sharply after a few exceptionally expensive properties enter the validation set, although typical absolute errors remain stable. The release is blocked until the underlying model behavior is understood. Which engineering response resolves the actual issue without masking the symptoms?
- Register immutable model package versions with lineage and require explicit approval before promotion to production.
- Use a smaller learning rate, optionally freeze early layers, and assess staged unfreezing against a held-out set.
- Compare RMSE with MAE or robust error summaries and choose the metric matching the application's penalty for large misses.
- Use appropriate SageMaker Debugger or framework instrumentation; adjust learning rate or gradient clipping after confirming instability.
Explanation
Correct Answer
C:Compare RMSE with MAE or robust error summaries and choose the metric matching the application's penalty for large misses. This is the appropriate controlled remedy for rmse versus mae outlier sensitivity. Squaring residuals gives large errors disproportionately high weight; MAE weights magnitudes linearly. Verify the cause with: Inspect squared versus absolute residual contributions, target distribution, and large-error business costs. Then confirm: The evaluation report shows both typical error and extreme-error impact with a justified primary metric. Reported signal: A house-price model's RMSE increases sharply after a few exceptionally expensive properties enter the validation set, although typical absolute errors remain stable. Operational setting: During production evaluation of a SageMaker AI model candidate.
Incorrect Answers
A:Register immutable model package versions with lineage and require explicit approval before promotion to production. This addresses a different issue involving sagemaker model registry candidate approval. It does not correct the stated rmse versus mae outlier sensitivity mechanism. The targeted response is: Compare RMSE with MAE or robust error summaries and choose the metric matching the application's penalty for large misses. Case evidence: A house-price model's RMSE increases sharply after a few exceptionally expensive properties enter the validation set, although typical absolute errors remain stable. Operational setting: During production evaluation of a SageMaker AI model candidate.
B:Use a smaller learning rate, optionally freeze early layers, and assess staged unfreezing against a held-out set. This addresses a different issue involving fine-tuning pretrained model learning-rate stability. It does not correct the stated rmse versus mae outlier sensitivity mechanism. The targeted response is: Compare RMSE with MAE or robust error summaries and choose the metric matching the application's penalty for large misses. Case evidence: A house-price model's RMSE increases sharply after a few exceptionally expensive properties enter the validation set, although typical absolute errors remain stable. Operational setting: During production evaluation of a SageMaker AI model candidate.
D:Use appropriate SageMaker Debugger or framework instrumentation; adjust learning rate or gradient clipping after confirming instability. This addresses a different issue involving model debugger exploding gradient diagnosis. It does not correct the stated rmse versus mae outlier sensitivity mechanism. The targeted response is: Compare RMSE with MAE or robust error summaries and choose the metric matching the application's penalty for large misses. Case evidence: A house-price model's RMSE increases sharply after a few exceptionally expensive properties enter the validation set, although typical absolute errors remain stable. Operational setting: During production evaluation of a SageMaker AI model candidate.
Question 384
During a controlled replay of the data-preparation workflow in a separate ML staging account, the following condition is observed: A medical-image classifier uses random row splitting so multiple scans from the same patient occur in training and validation. The dataset contract must remain reproducible for a backfill. What targeted change should be implemented after verifying the stated cause?
- Use rolling-origin or chronological validation with strictly past-derived features and appropriate forecast horizons.
- Split by stable patient identity before training and reserve a disjoint held-out group for evaluation.
- Fit an appropriate scaler on training data and package the same transformation with inference preprocessing.
- Compact the partitioned dataset into appropriately sized Parquet objects without changing the record keys or schema.
Explanation
Correct Answer
B:Split by stable patient identity before training and reserve a disjoint held-out group for evaluation. This is the appropriate controlled remedy for group-leakage train and validation split. Correlated records from one entity across splits can inflate performance without improving true generalization. Verify the cause with: Inspect shared entity identifiers across partitions and compare patient-level versus image-level evaluation results. Then confirm: No patient identifier crosses split boundaries and evaluation reflects generalization to unseen patients. Reported signal: A medical-image classifier uses random row splitting so multiple scans from the same patient occur in training and validation. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
Incorrect Answers
A:Use rolling-origin or chronological validation with strictly past-derived features and appropriate forecast horizons. This addresses a different issue involving chronological split for forecast evaluation. It does not correct the stated group-leakage train and validation split mechanism. The targeted response is: Split by stable patient identity before training and reserve a disjoint held-out group for evaluation. Case evidence: A medical-image classifier uses random row splitting so multiple scans from the same patient occur in training and validation. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
C:Fit an appropriate scaler on training data and package the same transformation with inference preprocessing. This addresses a different issue involving standardization of numerical features. It does not correct the stated group-leakage train and validation split mechanism. The targeted response is: Split by stable patient identity before training and reserve a disjoint held-out group for evaluation. Case evidence: A medical-image classifier uses random row splitting so multiple scans from the same patient occur in training and validation. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
D:Compact the partitioned dataset into appropriately sized Parquet objects without changing the record keys or schema. This addresses a different issue involving columnar file compaction for small objects. It does not correct the stated group-leakage train and validation split mechanism. The targeted response is: Split by stable patient identity before training and reserve a disjoint held-out group for evaluation. Case evidence: A medical-image classifier uses random row splitting so multiple scans from the same patient occur in training and validation. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
Question 385
At a production AWS ML data lake supporting customer-facing predictions, the following condition is observed: A medical-image classifier uses random row splitting so multiple scans from the same patient occur in training and validation. Engineers must avoid changing unrelated source partitions. Which TWO complementary choices provide the decisive diagnosis AND corrective response? Select TWO.
- Split by stable patient identity before training and reserve a disjoint held-out group for evaluation.
- Encode hour with sine and cosine cyclic features and validate with a time-aware split.
- Inspect shared entity identifiers across partitions and compare patient-level versus image-level evaluation results.
- Perform point-in-time correct joins using the most recent eligible pre-prediction feature record, not the globally latest value.
- Measure file-size distribution, object request count, Spark task scheduling overhead, and read throughput.
Explanation
Correct Answers
A:Split by stable patient identity before training and reserve a disjoint held-out group for evaluation. This is the appropriate controlled remedy for group-leakage train and validation split. Correlated records from one entity across splits can inflate performance without improving true generalization. Verify the cause with: Inspect shared entity identifiers across partitions and compare patient-level versus image-level evaluation results. Then confirm: No patient identifier crosses split boundaries and evaluation reflects generalization to unseen patients. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A medical-image classifier uses random row splitting so multiple scans from the same patient occur in training and validation. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
C:Inspect shared entity identifiers across partitions and compare patient-level versus image-level evaluation results. This directly tests group-leakage train and validation split in the reported AWS ML workflow. Correlated records from one entity across splits can inflate performance without improving true generalization. It distinguishes the problem before choosing a change. The related targeted response is: Split by stable patient identity before training and reserve a disjoint held-out group for evaluation. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A medical-image classifier uses random row splitting so multiple scans from the same patient occur in training and validation. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Incorrect Answers
B:Encode hour with sine and cosine cyclic features and validate with a time-aware split. This selection concerns time-series cyclical calendar feature rather than the required diagnostic-and-remedy pair for group-leakage train and validation split. Use evidence: Inspect shared entity identifiers across partitions and compare patient-level versus image-level evaluation results. and the matching correction: Split by stable patient identity before training and reserve a disjoint held-out group for evaluation. Signal: A medical-image classifier uses random row splitting so multiple scans from the same patient occur in training and validation. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
D:Perform point-in-time correct joins using the most recent eligible pre-prediction feature record, not the globally latest value. This selection concerns point-in-time feature data leakage rather than the required diagnostic-and-remedy pair for group-leakage train and validation split. Use evidence: Inspect shared entity identifiers across partitions and compare patient-level versus image-level evaluation results. and the matching correction: Split by stable patient identity before training and reserve a disjoint held-out group for evaluation. Signal: A medical-image classifier uses random row splitting so multiple scans from the same patient occur in training and validation. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
E:Measure file-size distribution, object request count, Spark task scheduling overhead, and read throughput. This selection concerns columnar file compaction for small objects rather than the required diagnostic-and-remedy pair for group-leakage train and validation split. Use evidence: Inspect shared entity identifiers across partitions and compare patient-level versus image-level evaluation results. and the matching correction: Split by stable patient identity before training and reserve a disjoint held-out group for evaluation. Signal: A medical-image classifier uses random row splitting so multiple scans from the same patient occur in training and validation. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Question 386
During a production MLOps reliability and security review, the following condition is observed: An ML pipeline in one account is allowed to write model predictions to an unapproved external bucket because its role has overly broad s3:PutObject access. The owner needs an auditable resolution without broadening privileges. What should the ML engineer inspect FIRST to distinguish this failure from unrelated issues?
- Inspect S3 object encryption, key policy, grant permissions, IAM role, and denied KMS decrypt events.
- Inspect captured request IDs, ground-truth keys, and the quality monitor dataset schema.
- Inspect identity and destination bucket policies, resource ARNs, organization restrictions, and CloudTrail data-event evidence.
- Inspect online versus offline feature computation code, event-time cutoffs, freshness, and sampled entity values.
Explanation
Correct Answer
C:Inspect identity and destination bucket policies, resource ARNs, organization restrictions, and CloudTrail data-event evidence. This directly tests cross-account s3 bucket exfiltration protection in the reported AWS ML workflow. Least privilege must constrain the destination resource; an unrestricted object-write policy can enable data exfiltration. It distinguishes the problem before choosing a change. The related targeted response is: Restrict writes to approved bucket prefixes and apply appropriate resource and organizational guardrails. Reported signal: An ML pipeline in one account is allowed to write model predictions to an unapproved external bucket because its role has overly broad s3:PutObject access. Operational setting: During a production MLOps reliability and security review.
Incorrect Answers
A:Inspect S3 object encryption, key policy, grant permissions, IAM role, and denied KMS decrypt events. This is a possible diagnostic for s3 sse-kms cross-account model artifact access, but it does not directly test the cross-account s3 bucket exfiltration protection failure reported here. Prioritize: Inspect identity and destination bucket policies, resource ARNs, organization restrictions, and CloudTrail data-event evidence. The relevant case is: An ML pipeline in one account is allowed to write model predictions to an unapproved external bucket because its role has overly broad s3:PutObject access. Operational setting: During a production MLOps reliability and security review.
B:Inspect captured request IDs, ground-truth keys, and the quality monitor dataset schema. This is a possible diagnostic for model monitor ground truth association, but it does not directly test the cross-account s3 bucket exfiltration protection failure reported here. Prioritize: Inspect identity and destination bucket policies, resource ARNs, organization restrictions, and CloudTrail data-event evidence. The relevant case is: An ML pipeline in one account is allowed to write model predictions to an unapproved external bucket because its role has overly broad s3:PutObject access. Operational setting: During a production MLOps reliability and security review.
D:Inspect online versus offline feature computation code, event-time cutoffs, freshness, and sampled entity values. This is a possible diagnostic for training-serving feature skew incident, but it does not directly test the cross-account s3 bucket exfiltration protection failure reported here. Prioritize: Inspect identity and destination bucket policies, resource ARNs, organization restrictions, and CloudTrail data-event evidence. The relevant case is: An ML pipeline in one account is allowed to write model predictions to an unapproved external bucket because its role has overly broad s3:PutObject access. Operational setting: During a production MLOps reliability and security review.
Question 387
As a regulated ML workload prepares for an automated regional rollout, the following condition is observed: A long training workflow uses only On-Demand capacity even though checkpoints and flexible completion time permit interruptions, and its compute cost breaches target. The team must demonstrate a repeatable monitoring or security control. What targeted change should be implemented after verifying the stated cause?
- Monitor calibration against delayed truth and retrain or recalibrate on representative permitted data when drift is confirmed.
- Identify the component contributing to total request time and optimize network or overhead separately from actual model inference.
- Use supported managed Spot Training with durable checkpoints for interruptible phases and retain On-Demand capacity when deadlines demand it.
- Use stable user-cohort assignment or an appropriate experiment design while keeping exposed groups measurable.
Explanation
Correct Answer
C:Use supported managed Spot Training with durable checkpoints for interruptible phases and retain On-Demand capacity when deadlines demand it. This is the appropriate controlled remedy for spot training cost versus interruption recovery. Spot savings depend on interruption handling; a lower hourly rate is not a guarantee of lower total job cost. Verify the cause with: Inspect training interruption tolerance, persistent checkpoints, allowed instance families, and deadline requirements. Then confirm: Measured training cost drops while interrupted test jobs resume from checkpoints within the delivery window. Reported signal: A long training workflow uses only On-Demand capacity even though checkpoints and flexible completion time permit interruptions, and its compute cost breaches target. Operational setting: As a regulated ML workload prepares for an automated regional rollout.
Incorrect Answers
A:Monitor calibration against delayed truth and retrain or recalibrate on representative permitted data when drift is confirmed. This addresses a different issue involving serving output confidence calibration drift. It does not correct the stated spot training cost versus interruption recovery mechanism. The targeted response is: Use supported managed Spot Training with durable checkpoints for interruptible phases and retain On-Demand capacity when deadlines demand it. Case evidence: A long training workflow uses only On-Demand capacity even though checkpoints and flexible completion time permit interruptions, and its compute cost breaches target. Operational setting: As a regulated ML workload prepares for an automated regional rollout.
B:Identify the component contributing to total request time and optimize network or overhead separately from actual model inference. This addresses a different issue involving cloudwatch invocations and modellatency distinction. It does not correct the stated spot training cost versus interruption recovery mechanism. The targeted response is: Use supported managed Spot Training with durable checkpoints for interruptible phases and retain On-Demand capacity when deadlines demand it. Case evidence: A long training workflow uses only On-Demand capacity even though checkpoints and flexible completion time permit interruptions, and its compute cost breaches target. Operational setting: As a regulated ML workload prepares for an automated regional rollout.
D:Use stable user-cohort assignment or an appropriate experiment design while keeping exposed groups measurable. This addresses a different issue involving champion challenger a/b assignment consistency. It does not correct the stated spot training cost versus interruption recovery mechanism. The targeted response is: Use supported managed Spot Training with durable checkpoints for interruptible phases and retain On-Demand capacity when deadlines demand it. Case evidence: A long training workflow uses only On-Demand capacity even though checkpoints and flexible completion time permit interruptions, and its compute cost breaches target. Operational setting: As a regulated ML workload prepares for an automated regional rollout.
Question 388
During an operational load and release rehearsal for SageMaker-hosted inference, the following condition is observed: A new recommendation model has uncertain online behavior but the team plans to replace 100 percent of production traffic immediately. Change control requires measurable endpoint or pipeline acceptance evidence. Which corrective action most directly addresses this ML workflow problem?
- Use a measured canary or linear traffic-shift deployment with alarms and immediate rollback conditions.
- Add schema, integration and sample endpoint inference tests to the ML build and deployment pipeline.
- Implement the SageMaker inference container
/ping and /invocations expectations or use a supported serving toolkit. - Compile the supported model for the specific edge target using SageMaker Neo or choose a compatible conversion route.
Explanation
Correct Answer
A:Use a measured canary or linear traffic-shift deployment with alarms and immediate rollback conditions. This is the appropriate controlled remedy for canary variant rollout for model changes. Progressive exposure limits blast radius; immediate cutover obscures whether regressions occurred during ramp-up. Verify the cause with: Inspect production variant weights, guardrail metrics, rollback procedure, and representativeness of staged exposure. Then confirm: A limited slice serves the new version while metrics are monitored before any broad rollout. Reported signal: A new recommendation model has uncertain online behavior but the team plans to replace 100 percent of production traffic immediately. Operational setting: During an operational load and release rehearsal for SageMaker-hosted inference.
Incorrect Answers
B:Add schema, integration and sample endpoint inference tests to the ML build and deployment pipeline. This addresses a different issue involving continuous integration unit and endpoint contract tests. It does not correct the stated canary variant rollout for model changes mechanism. The targeted response is: Use a measured canary or linear traffic-shift deployment with alarms and immediate rollback conditions. Case evidence: A new recommendation model has uncertain online behavior but the team plans to replace 100 percent of production traffic immediately. Operational setting: During an operational load and release rehearsal for SageMaker-hosted inference.
C:Implement the SageMaker inference container /ping and /invocations expectations or use a supported serving toolkit. This addresses a different issue involving bring-your-own-inference-container contract. It does not correct the stated canary variant rollout for model changes mechanism. The targeted response is: Use a measured canary or linear traffic-shift deployment with alarms and immediate rollback conditions. Case evidence: A new recommendation model has uncertain online behavior but the team plans to replace 100 percent of production traffic immediately. Operational setting: During an operational load and release rehearsal for SageMaker-hosted inference.
D:Compile the supported model for the specific edge target using SageMaker Neo or choose a compatible conversion route. This addresses a different issue involving sagemaker neo edge compiler compatibility. It does not correct the stated canary variant rollout for model changes mechanism. The targeted response is: Use a measured canary or linear traffic-shift deployment with alarms and immediate rollback conditions. Case evidence: A new recommendation model has uncertain online behavior but the team plans to replace 100 percent of production traffic immediately. Operational setting: During an operational load and release rehearsal for SageMaker-hosted inference.
Question 389
During production evaluation of a SageMaker AI model candidate, the following condition is observed: A deep model fits on one GPU but training takes too long over a very large dataset; the team suggests splitting individual layers across devices without evidence of model-memory pressure. The release is blocked until the underlying model behavior is understood. Which corrective action most directly addresses this ML workflow problem?
- Prefer data parallel training with synchronized gradients when the model fits per device and the bottleneck is dataset throughput.
- Select a decision threshold using recall or cost-aware precision/recall targets and measure on the untouched holdout.
- Select a regression approach with suitable loss and evaluation metrics instead of forcing continuous amounts into unrelated classes.
- Configure a supported SageMaker checkpoint location and resume training state on managed Spot interruption recovery.
Explanation
Correct Answer
A:Prefer data parallel training with synchronized gradients when the model fits per device and the bottleneck is dataset throughput. This is the appropriate controlled remedy for data parallelism versus model parallelism. Data parallelism shards data across replicas; model parallelism is more relevant when the model itself cannot fit on one device. Verify the cause with: Profile model memory, throughput, communication overhead, and per-device utilization. Then confirm: Multiple workers reduce epoch time with acceptable convergence and measured communication overhead. Reported signal: A deep model fits on one GPU but training takes too long over a very large dataset; the team suggests splitting individual layers across devices without evidence of model-memory pressure. Operational setting: During production evaluation of a SageMaker AI model candidate.
Incorrect Answers
B:Select a decision threshold using recall or cost-aware precision/recall targets and measure on the untouched holdout. This addresses a different issue involving precision versus recall fraud threshold. It does not correct the stated data parallelism versus model parallelism mechanism. The targeted response is: Prefer data parallel training with synchronized gradients when the model fits per device and the bottleneck is dataset throughput. Case evidence: A deep model fits on one GPU but training takes too long over a very large dataset; the team suggests splitting individual layers across devices without evidence of model-memory pressure. Operational setting: During production evaluation of a SageMaker AI model candidate.
C:Select a regression approach with suitable loss and evaluation metrics instead of forcing continuous amounts into unrelated classes. This addresses a different issue involving regression versus classification problem formulation. It does not correct the stated data parallelism versus model parallelism mechanism. The targeted response is: Prefer data parallel training with synchronized gradients when the model fits per device and the bottleneck is dataset throughput. Case evidence: A deep model fits on one GPU but training takes too long over a very large dataset; the team suggests splitting individual layers across devices without evidence of model-memory pressure. Operational setting: During production evaluation of a SageMaker AI model candidate.
D:Configure a supported SageMaker checkpoint location and resume training state on managed Spot interruption recovery. This addresses a different issue involving managed spot training checkpoint recovery. It does not correct the stated data parallelism versus model parallelism mechanism. The targeted response is: Prefer data parallel training with synchronized gradients when the model fits per device and the bottleneck is dataset throughput. Case evidence: A deep model fits on one GPU but training takes too long over a very large dataset; the team suggests splitting individual layers across devices without evidence of model-memory pressure. Operational setting: During production evaluation of a SageMaker AI model candidate.
Question 390
As a regulated ML workload prepares for an automated regional rollout, the following condition is observed: A SageMaker endpoint role has s3:GetObject for a model artifact, but download fails because the artifact is encrypted under a KMS key in a different account. The team must demonstrate a repeatable monitoring or security control. What targeted change should be implemented after verifying the stated cause?
- Set budget alerts and bounded tuning job quotas before launching experiments, and enforce an owner-approved stopping policy.
- Enable appropriate CloudTrail S3 data events for the sensitive dataset with protected storage and retention.
- Use a consistent versioned feature definition and test offline/online values at the same observation cutoff.
- Authorize the specific serving role to decrypt with the relevant KMS key under a restricted key policy and S3 access.
Explanation
Correct Answer
D:Authorize the specific serving role to decrypt with the relevant KMS key under a restricted key policy and S3 access. This is the appropriate controlled remedy for s3 sse-kms cross-account model artifact access. S3 object permission does not imply access to the KMS encryption key used on the object. Verify the cause with: Inspect S3 object encryption, key policy, grant permissions, IAM role, and denied KMS decrypt events. Then confirm: The endpoint loads the intended encrypted artifact while unrelated principals cannot decrypt it. Reported signal: A SageMaker endpoint role has s3:GetObject for a model artifact, but download fails because the artifact is encrypted under a KMS key in a different account. Operational setting: As a regulated ML workload prepares for an automated regional rollout.
Incorrect Answers
A:Set budget alerts and bounded tuning job quotas before launching experiments, and enforce an owner-approved stopping policy. This addresses a different issue involving aws budgets spend threshold for experimentation. It does not correct the stated s3 sse-kms cross-account model artifact access mechanism. The targeted response is: Authorize the specific serving role to decrypt with the relevant KMS key under a restricted key policy and S3 access. Case evidence: A SageMaker endpoint role has s3:GetObject for a model artifact, but download fails because the artifact is encrypted under a KMS key in a different account. Operational setting: As a regulated ML workload prepares for an automated regional rollout.
B:Enable appropriate CloudTrail S3 data events for the sensitive dataset with protected storage and retention. This addresses a different issue involving cloudtrail training data access audit coverage. It does not correct the stated s3 sse-kms cross-account model artifact access mechanism. The targeted response is: Authorize the specific serving role to decrypt with the relevant KMS key under a restricted key policy and S3 access. Case evidence: A SageMaker endpoint role has s3:GetObject for a model artifact, but download fails because the artifact is encrypted under a KMS key in a different account. Operational setting: As a regulated ML workload prepares for an automated regional rollout.
C:Use a consistent versioned feature definition and test offline/online values at the same observation cutoff. This addresses a different issue involving training-serving feature skew incident. It does not correct the stated s3 sse-kms cross-account model artifact access mechanism. The targeted response is: Authorize the specific serving role to decrypt with the relevant KMS key under a restricted key policy and S3 access. Case evidence: A SageMaker endpoint role has s3:GetObject for a model artifact, but download fails because the artifact is encrypted under a KMS key in a different account. Operational setting: As a regulated ML workload prepares for an automated regional rollout.
