Amazon AWS Certified Machine Learning Engineer - Associate MLA-C01 Practice Test 11
Question 100
During a controlled replay of the data-preparation workflow in a separate ML staging account, the following condition is observed: A fraud endpoint fetches the latest online customer feature while offline training joins an earlier feature snapshot, creating training-serving skew. The dataset contract must remain reproducible for a backfill. Select TWO: the appropriate diagnostic check and the targeted implementation change.
- Compare online Feature Store record values and offline event-time keyed feature history for the same entity and model snapshot.
- Split first; fit the imputer only on the training partition and apply those learned statistics to validation and test data.
- Define point-in-time offline joins, freshness SLAs, and aligned feature version logic for online retrieval.
- Keep restricted records and dependent training artifacts in approved regions and deny unauthorized replication or cross-region exports.
- Inspect event identifiers, object versions or ETags, processing ledger, and downstream record keys for duplicate deliveries.
Explanation
Correct Answers
A:Compare online Feature Store record values and offline event-time keyed feature history for the same entity and model snapshot. This directly tests sagemaker feature store online offline consistency in the reported AWS ML workflow. Online current state and offline historical records answer different temporal queries; blind latest-value joins cause leakage. It distinguishes the problem before choosing a change. The related targeted response is: Define point-in-time offline joins, freshness SLAs, and aligned feature version logic for online retrieval. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A fraud endpoint fetches the latest online customer feature while offline training joins an earlier feature snapshot, creating training-serving skew. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
C:Define point-in-time offline joins, freshness SLAs, and aligned feature version logic for online retrieval. This is the appropriate controlled remedy for sagemaker feature store online offline consistency. Online current state and offline historical records answer different temporal queries; blind latest-value joins cause leakage. Verify the cause with: Compare online Feature Store record values and offline event-time keyed feature history for the same entity and model snapshot. Then confirm: Historical evaluation reproduces features available at prediction time, and production freshness is monitored against the same contract. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A fraud endpoint fetches the latest online customer feature while offline training joins an earlier feature snapshot, creating training-serving skew. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
Incorrect Answers
B:Split first; fit the imputer only on the training partition and apply those learned statistics to validation and test data. This selection concerns median imputation fitted on training split rather than the required diagnostic-and-remedy pair for sagemaker feature store online offline consistency. Use evidence: Compare online Feature Store record values and offline event-time keyed feature history for the same entity and model snapshot. and the matching correction: Define point-in-time offline joins, freshness SLAs, and aligned feature version logic for online retrieval. Signal: A fraud endpoint fetches the latest online customer feature while offline training joins an earlier feature snapshot, creating training-serving skew. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
D:Keep restricted records and dependent training artifacts in approved regions and deny unauthorized replication or cross-region exports. This selection concerns model training data residency boundary rather than the required diagnostic-and-remedy pair for sagemaker feature store online offline consistency. Use evidence: Compare online Feature Store record values and offline event-time keyed feature history for the same entity and model snapshot. and the matching correction: Define point-in-time offline joins, freshness SLAs, and aligned feature version logic for online retrieval. Signal: A fraud endpoint fetches the latest online customer feature while offline training joins an earlier feature snapshot, creating training-serving skew. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
E:Inspect event identifiers, object versions or ETags, processing ledger, and downstream record keys for duplicate deliveries. This selection concerns s3 object event at-least-once processing rather than the required diagnostic-and-remedy pair for sagemaker feature store online offline consistency. Use evidence: Compare online Feature Store record values and offline event-time keyed feature history for the same entity and model snapshot. and the matching correction: Define point-in-time offline joins, freshness SLAs, and aligned feature version logic for online retrieval. Signal: A fraud endpoint fetches the latest online customer feature while offline training joins an earlier feature snapshot, creating training-serving skew. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
Question 101
As a regional data-pipeline rollout is validated against approved AWS storage and streams, the following condition is observed: A missing-value imputer calculates its median on the complete labeled dataset before train/test splitting, inflating apparent validation performance. The data team must preserve event-time meaning and lineage. Which TWO steps best establish the evidence and correct the identified root cause? Select TWO.
- Measure file-size distribution, object request count, Spark task scheduling overhead, and read throughput.
- Split first; fit the imputer only on the training partition and apply those learned statistics to validation and test data.
- Split by stable patient identity before training and reserve a disjoint held-out group for evaluation.
- Revise instructions, route ambiguous samples for expert adjudication, and monitor label quality per class.
- Inspect preprocessing fit order, imputation statistics, dataset split boundaries, and pipeline serialization.
Explanation
Correct Answers
B:Split first; fit the imputer only on the training partition and apply those learned statistics to validation and test data. This is the appropriate controlled remedy for median imputation fitted on training split. Preprocessing parameters learned from held-out records leak information into model selection. Verify the cause with: Inspect preprocessing fit order, imputation statistics, dataset split boundaries, and pipeline serialization. Then confirm: A held-out split uses the training-fitted median and independent evaluation no longer incorporates held-out distribution values. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A missing-value imputer calculates its median on the complete labeled dataset before train/test splitting, inflating apparent validation performance. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
E:Inspect preprocessing fit order, imputation statistics, dataset split boundaries, and pipeline serialization. This directly tests median imputation fitted on training split in the reported AWS ML workflow. Preprocessing parameters learned from held-out records leak information into model selection. It distinguishes the problem before choosing a change. The related targeted response is: Split first; fit the imputer only on the training partition and apply those learned statistics to validation and test data. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A missing-value imputer calculates its median on the complete labeled dataset before train/test splitting, inflating apparent validation performance. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
Incorrect Answers
A: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 median imputation fitted on training split. Use evidence: Inspect preprocessing fit order, imputation statistics, dataset split boundaries, and pipeline serialization. and the matching correction: Split first; fit the imputer only on the training partition and apply those learned statistics to validation and test data. Signal: A missing-value imputer calculates its median on the complete labeled dataset before train/test splitting, inflating apparent validation performance. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
C:Split by stable patient identity before training and reserve a disjoint held-out group for evaluation. This selection concerns group-leakage train and validation split rather than the required diagnostic-and-remedy pair for median imputation fitted on training split. Use evidence: Inspect preprocessing fit order, imputation statistics, dataset split boundaries, and pipeline serialization. and the matching correction: Split first; fit the imputer only on the training partition and apply those learned statistics to validation and test data. Signal: A missing-value imputer calculates its median on the complete labeled dataset before train/test splitting, inflating apparent validation performance. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
D:Revise instructions, route ambiguous samples for expert adjudication, and monitor label quality per class. This selection concerns ground truth labeling agreement review rather than the required diagnostic-and-remedy pair for median imputation fitted on training split. Use evidence: Inspect preprocessing fit order, imputation statistics, dataset split boundaries, and pipeline serialization. and the matching correction: Split first; fit the imputer only on the training partition and apply those learned statistics to validation and test data. Signal: A missing-value imputer calculates its median on the complete labeled dataset before train/test splitting, inflating apparent validation performance. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
Question 102
While reproducing the alert in an isolated AWS ML integration account, the following condition is observed: A model is deployed to a large general-purpose instance based only on notebook habits, with no comparison of lower-cost inference-optimized choices. The proposed change must preserve logging and least privilege. Which engineering response resolves the actual issue without masking the symptoms?
- Use structured logging and scoped CloudWatch Logs Insights queries to aggregate failures by training job and time.
- Restrict writes to approved bucket prefixes and apply appropriate resource and organizational guardrails.
- Use SageMaker Inference Recommender or equivalent controlled load tests to rightsize the serving instance family.
- Remove embedded credentials, rotate the exposed secret, and retrieve current credentials through an approved managed secret service.
Explanation
Correct Answer
C:Use SageMaker Inference Recommender or equivalent controlled load tests to rightsize the serving instance family. This is the appropriate controlled remedy for inference recommender benchmarking instance family. Rightsizing depends on workload measurements rather than the training machine or instance size alone. Verify the cause with: Measure memory, throughput, p95 latency and cost across supported candidate endpoint configurations. Then confirm: A chosen instance meets throughput and latency targets at lower measured cost per thousand predictions. Reported signal: A model is deployed to a large general-purpose instance based only on notebook habits, with no comparison of lower-cost inference-optimized choices. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
Incorrect Answers
A:Use structured logging and scoped CloudWatch Logs Insights queries to aggregate failures by training job and time. This addresses a different issue involving cloudwatch logs insights error correlation. It does not correct the stated inference recommender benchmarking instance family mechanism. The targeted response is: Use SageMaker Inference Recommender or equivalent controlled load tests to rightsize the serving instance family. Case evidence: A model is deployed to a large general-purpose instance based only on notebook habits, with no comparison of lower-cost inference-optimized choices. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
B:Restrict writes to approved bucket prefixes and apply appropriate resource and organizational guardrails. This addresses a different issue involving cross-account s3 bucket exfiltration protection. It does not correct the stated inference recommender benchmarking instance family mechanism. The targeted response is: Use SageMaker Inference Recommender or equivalent controlled load tests to rightsize the serving instance family. Case evidence: A model is deployed to a large general-purpose instance based only on notebook habits, with no comparison of lower-cost inference-optimized choices. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
D:Remove embedded credentials, rotate the exposed secret, and retrieve current credentials through an approved managed secret service. This addresses a different issue involving secret rotation for inference container credentials. It does not correct the stated inference recommender benchmarking instance family mechanism. The targeted response is: Use SageMaker Inference Recommender or equivalent controlled load tests to rightsize the serving instance family. Case evidence: A model is deployed to a large general-purpose instance based only on notebook habits, with no comparison of lower-cost inference-optimized choices. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
Question 103
While reproducing a training incident in an independent SageMaker experiment, the following condition is observed: A pretrained image encoder loses useful general features after fine-tuning on a small labeled dataset using an excessively high learning rate. The experiment must be compared fairly against its prior baseline. Which corrective action most directly addresses this ML workflow problem?
- Select a regression approach with suitable loss and evaluation metrics instead of forcing continuous amounts into unrelated classes.
- Use a smaller learning rate, optionally freeze early layers, and assess staged unfreezing against a held-out set.
- Use a time-series forecasting approach with past context and seasonal features, evaluating all intended future horizons.
- Adapt the custom training container to SageMaker training job conventions and verify locally against a representative channel layout.
Explanation
Correct Answer
B:Use a smaller learning rate, optionally freeze early layers, and assess staged unfreezing against a held-out set. This is the appropriate controlled remedy for fine-tuning pretrained model learning-rate stability. Large updates on small data can overwrite pretrained features; careful transfer learning limits destructive forgetting. Verify the cause with: Compare pretrained validation metrics, fine-tuning learning curves, and layer-wise parameter updates. Then confirm: The fine-tuned model improves task accuracy without collapsing pretrained representations. Reported signal: A pretrained image encoder loses useful general features after fine-tuning on a small labeled dataset using an excessively high learning rate. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
Incorrect Answers
A: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 fine-tuning pretrained model learning-rate stability mechanism. The targeted response is: Use a smaller learning rate, optionally freeze early layers, and assess staged unfreezing against a held-out set. Case evidence: A pretrained image encoder loses useful general features after fine-tuning on a small labeled dataset using an excessively high learning rate. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
C:Use a time-series forecasting approach with past context and seasonal features, evaluating all intended future horizons. This addresses a different issue involving time-series forecast horizon and seasonality. It does not correct the stated fine-tuning pretrained model learning-rate stability mechanism. The targeted response is: Use a smaller learning rate, optionally freeze early layers, and assess staged unfreezing against a held-out set. Case evidence: A pretrained image encoder loses useful general features after fine-tuning on a small labeled dataset using an excessively high learning rate. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
D:Adapt the custom training container to SageMaker training job conventions and verify locally against a representative channel layout. This addresses a different issue involving training image compatibility for byoc. It does not correct the stated fine-tuning pretrained model learning-rate stability mechanism. The targeted response is: Use a smaller learning rate, optionally freeze early layers, and assess staged unfreezing against a held-out set. Case evidence: A pretrained image encoder loses useful general features after fine-tuning on a small labeled dataset using an excessively high learning rate. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
Question 104
At a production AWS ML data lake supporting customer-facing predictions, the following condition is observed: A regulated ML pipeline copies EU-only customer records into an unapproved cross-region training bucket for convenience. Engineers must avoid changing unrelated source partitions. What should the ML engineer inspect FIRST to distinguish this failure from unrelated issues?
- Inspect event identifiers, object versions or ETags, processing ledger, and downstream record keys for duplicate deliveries.
- Inspect sampler placement, duplicate neighbor IDs, fold membership, and class-conditioned metric changes.
- Inspect event timestamps, ingest timestamps, watermark configuration, and window outputs for delayed sample records.
- Inspect S3 replication rules, region-specific bucket policies, Glue job destinations, KMS keys, and applicable residency controls.
Explanation
Correct Answer
D:Inspect S3 replication rules, region-specific bucket policies, Glue job destinations, KMS keys, and applicable residency controls. This directly tests model training data residency boundary in the reported AWS ML workflow. Encrypting copied data does not automatically satisfy geographic residency constraints; placement and access must be controlled. It distinguishes the problem before choosing a change. The related targeted response is: Keep restricted records and dependent training artifacts in approved regions and deny unauthorized replication or cross-region exports. Reported signal: A regulated ML pipeline copies EU-only customer records into an unapproved cross-region training bucket for convenience. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
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 model training data residency boundary failure reported here. Prioritize: Inspect S3 replication rules, region-specific bucket policies, Glue job destinations, KMS keys, and applicable residency controls. The relevant case is: A regulated ML pipeline copies EU-only customer records into an unapproved cross-region training bucket for convenience. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
B:Inspect sampler placement, duplicate neighbor IDs, fold membership, and class-conditioned metric changes. This is a possible diagnostic for class imbalance handled inside cross-validation, but it does not directly test the model training data residency boundary failure reported here. Prioritize: Inspect S3 replication rules, region-specific bucket policies, Glue job destinations, KMS keys, and applicable residency controls. The relevant case is: A regulated ML pipeline copies EU-only customer records into an unapproved cross-region training bucket for convenience. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
C:Inspect event timestamps, ingest timestamps, watermark configuration, and window outputs for delayed sample records. This is a possible diagnostic for streaming feature rolling window correctness, but it does not directly test the model training data residency boundary failure reported here. Prioritize: Inspect S3 replication rules, region-specific bucket policies, Glue job destinations, KMS keys, and applicable residency controls. The relevant case is: A regulated ML pipeline copies EU-only customer records into an unapproved cross-region training bucket for convenience. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Question 105
During an operational load and release rehearsal for SageMaker-hosted inference, the following condition is observed: A small CPU inference model receives only occasional requests, but a permanently provisioned endpoint stays idle for most hours. Change control requires measurable endpoint or pipeline acceptance evidence. Which observation most directly identifies the root cause of the reported behavior?
- Profile request arrival intervals, cold-start tolerance, memory needs, and compatible serverless inference limits.
- Inspect invocation cadence, acceptable completion window, per-record latency requirement, and idle endpoint cost.
- Compare invocation count per instance, model latency, GPU utilization, and the behavior of current scaling policy.
- Compare template parameters, stack drift results, change sets, and deployed endpoint settings.
Explanation
Correct Answer
A:Profile request arrival intervals, cold-start tolerance, memory needs, and compatible serverless inference limits. This directly tests serverless inference for intermittent requests in the reported AWS ML workflow. Serverless capacity trades some cold-start behavior and resource limits for reduced idle endpoint commitment. It distinguishes the problem before choosing a change. The related targeted response is: Use SageMaker Serverless Inference where the model and latency envelope fit, or justify provisioned capacity if it does not. Reported signal: A small CPU inference model receives only occasional requests, but a permanently provisioned endpoint stays idle for most hours. Operational setting: During an operational load and release rehearsal for SageMaker-hosted inference.
Incorrect Answers
B:Inspect invocation cadence, acceptable completion window, per-record latency requirement, and idle endpoint cost. This is a possible diagnostic for real-time versus batch inference decision, but it does not directly test the serverless inference for intermittent requests failure reported here. Prioritize: Profile request arrival intervals, cold-start tolerance, memory needs, and compatible serverless inference limits. The relevant case is: A small CPU inference model receives only occasional requests, but a permanently provisioned endpoint stays idle for most hours. Operational setting: During an operational load and release rehearsal for SageMaker-hosted inference.
C:Compare invocation count per instance, model latency, GPU utilization, and the behavior of current scaling policy. This is a possible diagnostic for sagemaker endpoint auto scaling metric selection, but it does not directly test the serverless inference for intermittent requests failure reported here. Prioritize: Profile request arrival intervals, cold-start tolerance, memory needs, and compatible serverless inference limits. The relevant case is: A small CPU inference model receives only occasional requests, but a permanently provisioned endpoint stays idle for most hours. Operational setting: During an operational load and release rehearsal for SageMaker-hosted inference.
D:Compare template parameters, stack drift results, change sets, and deployed endpoint settings. This is a possible diagnostic for cloudformation stack parameter drift, but it does not directly test the serverless inference for intermittent requests failure reported here. Prioritize: Profile request arrival intervals, cold-start tolerance, memory needs, and compatible serverless inference limits. The relevant case is: A small CPU inference model receives only occasional requests, but a permanently provisioned endpoint stays idle for most hours. Operational setting: During an operational load and release rehearsal for SageMaker-hosted inference.
Question 106
At a production AWS ML data lake supporting customer-facing predictions, the following condition is observed: A daily S3 dataset consists of millions of tiny Parquet files, and training startup spends more time opening objects than scanning data. Engineers must avoid changing unrelated source partitions. Which TWO steps best establish the evidence and correct the identified root cause? Select TWO.
- Inspect feature scale distributions, model sensitivity, and whether fitted scaler statistics are shared across serving and training.
- Increase appropriate retention before long maintenance and persist a durable downstream historical copy for replay beyond retention.
- Measure file-size distribution, object request count, Spark task scheduling overhead, and read throughput.
- Compact the partitioned dataset into appropriately sized Parquet objects without changing the record keys or schema.
- Repartition or salt the skewed join appropriately, or use a justified broadcast join when the small side fits memory.
Explanation
Correct Answers
C:Measure file-size distribution, object request count, Spark task scheduling overhead, and read throughput. This directly tests columnar file compaction for small objects in the reported AWS ML workflow. Excessive small files create metadata and request overhead even with a columnar data format. It distinguishes the problem before choosing a change. The related targeted response is: Compact the partitioned dataset into appropriately sized Parquet objects without changing the record keys or schema. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A daily S3 dataset consists of millions of tiny Parquet files, and training startup spends more time opening objects than scanning data. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
D:Compact the partitioned dataset into appropriately sized Parquet objects without changing the record keys or schema. This is the appropriate controlled remedy for columnar file compaction for small objects. Excessive small files create metadata and request overhead even with a columnar data format. Verify the cause with: Measure file-size distribution, object request count, Spark task scheduling overhead, and read throughput. Then confirm: Training read time and object operations decline while the resulting row count and checksums match. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A daily S3 dataset consists of millions of tiny Parquet files, and training startup spends more time opening objects than scanning data. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Incorrect Answers
A:Inspect feature scale distributions, model sensitivity, and whether fitted scaler statistics are shared across serving and training. This selection concerns standardization of numerical features rather than the required diagnostic-and-remedy pair for columnar file compaction for small objects. Use evidence: Measure file-size distribution, object request count, Spark task scheduling overhead, and read throughput. and the matching correction: Compact the partitioned dataset into appropriately sized Parquet objects without changing the record keys or schema. Signal: A daily S3 dataset consists of millions of tiny Parquet files, and training startup spends more time opening objects than scanning data. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
B:Increase appropriate retention before long maintenance and persist a durable downstream historical copy for replay beyond retention. This selection concerns kinesis consumer retention gap rather than the required diagnostic-and-remedy pair for columnar file compaction for small objects. Use evidence: Measure file-size distribution, object request count, Spark task scheduling overhead, and read throughput. and the matching correction: Compact the partitioned dataset into appropriately sized Parquet objects without changing the record keys or schema. Signal: A daily S3 dataset consists of millions of tiny Parquet files, and training startup spends more time opening objects than scanning data. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
E:Repartition or salt the skewed join appropriately, or use a justified broadcast join when the small side fits memory. This selection concerns glue spark join skew on dominant identifier rather than the required diagnostic-and-remedy pair for columnar file compaction for small objects. Use evidence: Measure file-size distribution, object request count, Spark task scheduling overhead, and read throughput. and the matching correction: Compact the partitioned dataset into appropriately sized Parquet objects without changing the record keys or schema. Signal: A daily S3 dataset consists of millions of tiny Parquet files, and training startup spends more time opening objects than scanning data. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Question 107
While testing a second region for recoverable deployment of the same ML service, the following condition is observed: Two production regions deploy different endpoint instance sizes because engineers manually alter console settings after the shared CloudFormation stack is created. The deployment must be reproducible without manual console-only fixes. What targeted change should be implemented after verifying the stated cause?
- Require model evaluation and explicit approval gates before production deployment, blocking failed metric thresholds.
- Configure a justified endpoint variant auto scaling policy using a workload-relevant supported metric such as invocations per instance.
- Use a SageMaker Batch Transform or batch inference job for the scheduled bulk scoring workflow.
- Define the endpoint resources and parameters in versioned CloudFormation or CDK code, review a change set, and reconcile drift.
Explanation
Correct Answer
D:Define the endpoint resources and parameters in versioned CloudFormation or CDK code, review a change set, and reconcile drift. This is the appropriate controlled remedy for cloudformation stack parameter drift. Infrastructure as code only controls actual state when out-of-band changes are detected and reconciled. Verify the cause with: Compare template parameters, stack drift results, change sets, and deployed endpoint settings. Then confirm: Both regions match their approved parameterized templates and subsequent deployments remain repeatable. Reported signal: Two production regions deploy different endpoint instance sizes because engineers manually alter console settings after the shared CloudFormation stack is created. Operational setting: While testing a second region for recoverable deployment of the same ML service.
Incorrect Answers
A:Require model evaluation and explicit approval gates before production deployment, blocking failed metric thresholds. This addresses a different issue involving codepipeline approval gate on model metrics. It does not correct the stated cloudformation stack parameter drift mechanism. The targeted response is: Define the endpoint resources and parameters in versioned CloudFormation or CDK code, review a change set, and reconcile drift. Case evidence: Two production regions deploy different endpoint instance sizes because engineers manually alter console settings after the shared CloudFormation stack is created. Operational setting: While testing a second region for recoverable deployment of the same ML service.
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 cloudformation stack parameter drift mechanism. The targeted response is: Define the endpoint resources and parameters in versioned CloudFormation or CDK code, review a change set, and reconcile drift. Case evidence: Two production regions deploy different endpoint instance sizes because engineers manually alter console settings after the shared CloudFormation stack is created. Operational setting: While testing a second region for recoverable deployment of the same ML service.
C:Use a SageMaker Batch Transform or batch inference job for the scheduled bulk scoring workflow. This addresses a different issue involving real-time versus batch inference decision. It does not correct the stated cloudformation stack parameter drift mechanism. The targeted response is: Define the endpoint resources and parameters in versioned CloudFormation or CDK code, review a change set, and reconcile drift. Case evidence: Two production regions deploy different endpoint instance sizes because engineers manually alter console settings after the shared CloudFormation stack is created. Operational setting: While testing a second region for recoverable deployment of the same ML service.
Question 108
While testing a second region for recoverable deployment of the same ML service, the following condition is observed: A pipeline deploys a model directly from an unversioned notebook S3 path and cannot prove which evaluated candidate was approved. The deployment must be reproducible without manual console-only fixes. Which TWO steps best establish the evidence and correct the identified root cause? Select TWO.
- Define a parameterized SageMaker Pipeline that orders steps and records source data, code and model artifacts.
- Compare template parameters, stack drift results, change sets, and deployed endpoint settings.
- Promote an explicitly approved SageMaker Model Registry package version through repeatable pipeline stages.
- Set adequate minimum capacity, scaling limits, alarms, and request backpressure for the expected traffic ramp.
- Inspect model registry package versions, approval metadata, artifact hashes, and pipeline deployment inputs.
Explanation
Correct Answers
C:Promote an explicitly approved SageMaker Model Registry package version through repeatable pipeline stages. This is the appropriate controlled remedy for model registry version promotion provenance. Immutable package versions and approval metadata connect evaluation with deployment; mutable S3 nicknames do not. Verify the cause with: Inspect model registry package versions, approval metadata, artifact hashes, and pipeline deployment inputs. Then confirm: The production endpoint matches the reviewed model package and auditor can trace its evaluation and approver. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A pipeline deploys a model directly from an unversioned notebook S3 path and cannot prove which evaluated candidate was approved. Operational setting: While testing a second region for recoverable deployment of the same ML service.
E:Inspect model registry package versions, approval metadata, artifact hashes, and pipeline deployment inputs. This directly tests model registry version promotion provenance in the reported AWS ML workflow. Immutable package versions and approval metadata connect evaluation with deployment; mutable S3 nicknames do not. It distinguishes the problem before choosing a change. The related targeted response is: Promote an explicitly approved SageMaker Model Registry package version through repeatable pipeline stages. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A pipeline deploys a model directly from an unversioned notebook S3 path and cannot prove which evaluated candidate was approved. Operational setting: While testing a second region for recoverable deployment of the same ML service.
Incorrect Answers
A:Define a parameterized SageMaker Pipeline that orders steps and records source data, code and model artifacts. This selection concerns sagemaker pipelines parameterized retraining rather than the required diagnostic-and-remedy pair for model registry version promotion provenance. Use evidence: Inspect model registry package versions, approval metadata, artifact hashes, and pipeline deployment inputs. and the matching correction: Promote an explicitly approved SageMaker Model Registry package version through repeatable pipeline stages. Signal: A pipeline deploys a model directly from an unversioned notebook S3 path and cannot prove which evaluated candidate was approved. Operational setting: While testing a second region for recoverable deployment of the same ML service.
B:Compare template parameters, stack drift results, change sets, and deployed endpoint settings. This selection concerns cloudformation stack parameter drift rather than the required diagnostic-and-remedy pair for model registry version promotion provenance. Use evidence: Inspect model registry package versions, approval metadata, artifact hashes, and pipeline deployment inputs. and the matching correction: Promote an explicitly approved SageMaker Model Registry package version through repeatable pipeline stages. Signal: A pipeline deploys a model directly from an unversioned notebook S3 path and cannot prove which evaluated candidate was approved. Operational setting: While testing a second region for recoverable deployment of the same ML service.
D:Set adequate minimum capacity, scaling limits, alarms, and request backpressure for the expected traffic ramp. This selection concerns endpoint scale-to-zero cold capacity planning rather than the required diagnostic-and-remedy pair for model registry version promotion provenance. Use evidence: Inspect model registry package versions, approval metadata, artifact hashes, and pipeline deployment inputs. and the matching correction: Promote an explicitly approved SageMaker Model Registry package version through repeatable pipeline stages. Signal: A pipeline deploys a model directly from an unversioned notebook S3 path and cannot prove which evaluated candidate was approved. Operational setting: While testing a second region for recoverable deployment of the same ML service.
Question 109
While reproducing the alert in an isolated AWS ML integration account, the following condition is observed: A monitor issues thousands of alerts after a small harmless shift in an optional numeric feature and on-call engineers begin ignoring all violations. The proposed change must preserve logging and least privilege. Which verification result BEST demonstrates that the issue is resolved?
- The new container operates with rotated credentials and image scans find no active embedded password.
- Auditors see reproducible subgroup statistics and alarms for material disparities rather than only global totals.
- Representative benign fluctuations stay quiet while a known severe schema or distribution break pages the owner.
- The job reads authorized training data while attempts to access unrelated buckets remain denied.
Explanation
Correct Answer
C:Representative benign fluctuations stay quiet while a known severe schema or distribution break pages the owner. This is the acceptance evidence for the described model monitor violation alert calibration scenario. Monitoring must balance sensitivity and alert fatigue; treating every small statistical change as an incident is unreliable. The corrective implementation it validates is: Tune justified thresholds and group actionable monitor violations by severity without suppressing critical data-integrity checks. Reported signal: A monitor issues thousands of alerts after a small harmless shift in an optional numeric feature and on-call engineers begin ignoring all violations. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
Incorrect Answers
A:The new container operates with rotated credentials and image scans find no active embedded password. This could validate secret rotation for inference container credentials, but it does not establish that model monitor violation alert calibration is resolved for this workload. Confirm instead: Representative benign fluctuations stay quiet while a known severe schema or distribution break pages the owner. Observed problem: A monitor issues thousands of alerts after a small harmless shift in an optional numeric feature and on-call engineers begin ignoring all violations. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
B:Auditors see reproducible subgroup statistics and alarms for material disparities rather than only global totals. This could validate clarify postdeployment group bias monitoring, but it does not establish that model monitor violation alert calibration is resolved for this workload. Confirm instead: Representative benign fluctuations stay quiet while a known severe schema or distribution break pages the owner. Observed problem: A monitor issues thousands of alerts after a small harmless shift in an optional numeric feature and on-call engineers begin ignoring all violations. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
D:The job reads authorized training data while attempts to access unrelated buckets remain denied. This could validate sagemaker execution role s3 least privilege, but it does not establish that model monitor violation alert calibration is resolved for this workload. Confirm instead: Representative benign fluctuations stay quiet while a known severe schema or distribution break pages the owner. Observed problem: A monitor issues thousands of alerts after a small harmless shift in an optional numeric feature and on-call engineers begin ignoring all violations. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
Question 110
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 postchange observation provides the strongest acceptance evidence?
- Measured training cost drops while interrupted test jobs resume from checkpoints within the delivery window.
- A test forecast crossing the approved spend threshold alerts before uncontrolled parallel job growth.
- Auditors see reproducible subgroup statistics and alarms for material disparities rather than only global totals.
- The pipeline deploys with the authorized role and cannot assign an unrelated privileged role.
Explanation
Correct Answer
A:Measured training cost drops while interrupted test jobs resume from checkpoints within the delivery window. This is the acceptance evidence for the described spot training cost versus interruption recovery scenario. Spot savings depend on interruption handling; a lower hourly rate is not a guarantee of lower total job cost. The corrective implementation it validates is: Use supported managed Spot Training with durable checkpoints for interruptible phases and retain On-Demand capacity when deadlines demand it. 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
B:A test forecast crossing the approved spend threshold alerts before uncontrolled parallel job growth. This could validate aws budgets spend threshold for experimentation, but it does not establish that spot training cost versus interruption recovery is resolved for this workload. Confirm instead: Measured training cost drops while interrupted test jobs resume from checkpoints within the delivery window. Observed problem: 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.
C:Auditors see reproducible subgroup statistics and alarms for material disparities rather than only global totals. This could validate clarify postdeployment group bias monitoring, but it does not establish that spot training cost versus interruption recovery is resolved for this workload. Confirm instead: Measured training cost drops while interrupted test jobs resume from checkpoints within the delivery window. Observed problem: 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:The pipeline deploys with the authorized role and cannot assign an unrelated privileged role. This could validate iam passrole constraint for model deployment, but it does not establish that spot training cost versus interruption recovery is resolved for this workload. Confirm instead: Measured training cost drops while interrupted test jobs resume from checkpoints within the delivery window. Observed problem: 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.
