Amazon AWS Certified Machine Learning Engineer - Associate MLA-C01 Practice Test 23
Question 220
During a controlled replay of the data-preparation workflow in a separate ML staging account, the following condition is observed: A training job is dominated by extreme transaction-amount magnitudes, although most records have modest nonnegative amounts. The dataset contract must remain reproducible for a backfill. Which TWO steps best establish the evidence and correct the identified root cause? Select TWO.
- Apply a justified log1p transformation with documented treatment of invalid negatives and package it for serving.
- Inspect quantiles, heavy-tail plots, outliers, and the target model's response to feature scale.
- Inspect group sample sizes, label-positive fractions, and the appropriate SageMaker Clarify pretraining bias report.
- Remove or pseudonymize unnecessary PII before annotation and restrict access to the token mapping under separate controls.
- Revise instructions, route ambiguous samples for expert adjudication, and monitor label quality per class.
Explanation
Correct Answers
A:Apply a justified log1p transformation with documented treatment of invalid negatives and package it for serving. This is the appropriate controlled remedy for skewed numerical feature log transformation. A monotonic log transform can reduce heavy-tail influence, but invalid domains and train-serving parity must be handled. Verify the cause with: Inspect quantiles, heavy-tail plots, outliers, and the target model's response to feature scale. Then confirm: The transformed distribution is less skewed and inference uses identical guarded transformations. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A training job is dominated by extreme transaction-amount magnitudes, although most records have modest nonnegative amounts. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
B: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. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A training job is dominated by extreme transaction-amount magnitudes, although most records have modest nonnegative amounts. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
Incorrect Answers
C:Inspect group sample sizes, label-positive fractions, and the appropriate SageMaker Clarify pretraining bias report. This selection concerns clarify label imbalance measurement rather than the required diagnostic-and-remedy pair for skewed numerical feature log transformation. Use evidence: Inspect quantiles, heavy-tail plots, outliers, and the target model's response to feature scale. and the matching correction: Apply a justified log1p transformation with documented treatment of invalid negatives and package it for serving. Signal: A training job is dominated by extreme transaction-amount magnitudes, although most records have modest nonnegative amounts. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
D:Remove or pseudonymize unnecessary PII before annotation and restrict access to the token mapping under separate controls. This selection concerns pii masking before model labeling rather than the required diagnostic-and-remedy pair for skewed numerical feature log transformation. Use evidence: Inspect quantiles, heavy-tail plots, outliers, and the target model's response to feature scale. and the matching correction: Apply a justified log1p transformation with documented treatment of invalid negatives and package it for serving. Signal: A training job is dominated by extreme transaction-amount magnitudes, although most records have modest nonnegative amounts. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
E: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 skewed numerical feature log transformation. Use evidence: Inspect quantiles, heavy-tail plots, outliers, and the target model's response to feature scale. and the matching correction: Apply a justified log1p transformation with documented treatment of invalid negatives and package it for serving. Signal: A training job is dominated by extreme transaction-amount magnitudes, although most records have modest nonnegative amounts. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
Question 221
While reproducing a training incident in an independent SageMaker experiment, the following condition is observed: A custom SageMaker training image exits before fitting because its entrypoint does not understand the job's expected channel paths and arguments. The experiment must be compared fairly against its prior baseline. Which investigation provides the most decisive FIRST evidence for the underlying problem?
- Inspect Docker entrypoint, training logs, image architecture, input mount points, and required container contract.
- Inspect target types, business loss function, sample labels, and whether the required output is continuous or categorical.
- Compare training and validation curves, coefficient magnitudes, and error by split.
- Measure class prevalence, precision-recall curve, ROC curve, and alert volume at the intended operating threshold.
Explanation
Correct Answer
A:Inspect Docker entrypoint, training logs, image architecture, input mount points, and required container contract. This directly tests training image compatibility for byoc in the reported AWS ML workflow. Custom containers must implement the hosting or training contract; building a valid Docker image alone is insufficient. It distinguishes the problem before choosing a change. The related targeted response is: Adapt the custom training container to SageMaker training job conventions and verify locally against a representative channel layout. Reported signal: A custom SageMaker training image exits before fitting because its entrypoint does not understand the job's expected channel paths and arguments. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
Incorrect Answers
B:Inspect target types, business loss function, sample labels, and whether the required output is continuous or categorical. This is a possible diagnostic for regression versus classification problem formulation, but it does not directly test the training image compatibility for byoc failure reported here. Prioritize: Inspect Docker entrypoint, training logs, image architecture, input mount points, and required container contract. The relevant case is: A custom SageMaker training image exits before fitting because its entrypoint does not understand the job's expected channel paths and arguments. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
C:Compare training and validation curves, coefficient magnitudes, and error by split. This is a possible diagnostic for overfitting reduced with l2 regularization, but it does not directly test the training image compatibility for byoc failure reported here. Prioritize: Inspect Docker entrypoint, training logs, image architecture, input mount points, and required container contract. The relevant case is: A custom SageMaker training image exits before fitting because its entrypoint does not understand the job's expected channel paths and arguments. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
D:Measure class prevalence, precision-recall curve, ROC curve, and alert volume at the intended operating threshold. This is a possible diagnostic for roc-auc versus pr-auc for rare events, but it does not directly test the training image compatibility for byoc failure reported here. Prioritize: Inspect Docker entrypoint, training logs, image architecture, input mount points, and required container contract. The relevant case is: A custom SageMaker training image exits before fitting because its entrypoint does not understand the job's expected channel paths and arguments. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
Question 222
During a controlled replay of the data-preparation workflow in a separate ML staging account, the following condition is observed: A raw CSV field contains quoted commas and a header is sometimes missing, so naïve delimiter splitting shifts the label into the wrong column. The dataset contract must remain reproducible for a backfill. Which verification result BEST demonstrates that the issue is resolved?
- Historical evaluation reproduces features available at prediction time, and production freshness is monitored against the same contract.
- Validated records preserve the expected label and field order while malformed inputs are reported separately.
- Every held-out forecast uses only data that existed at prediction time and its backtest follows real ordering.
- Replaying the same notification leaves exactly one corresponding logical training record and records deduplication.
Explanation
Correct Answer
B:Validated records preserve the expected label and field order while malformed inputs are reported separately. This is the acceptance evidence for the described csv schema ambiguity during ingestion scenario. CSV parsing must respect quoting and schema; splitting on every comma corrupts valid quoted fields. The corrective implementation it validates is: Use a CSV parser with explicit quoting and schema validation; quarantine malformed records before training. Reported signal: A raw CSV field contains quoted commas and a header is sometimes missing, so naïve delimiter splitting shifts the label into the wrong column. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
Incorrect Answers
A:Historical evaluation reproduces features available at prediction time, and production freshness is monitored against the same contract. This could validate sagemaker feature store online offline consistency, but it does not establish that csv schema ambiguity during ingestion is resolved for this workload. Confirm instead: Validated records preserve the expected label and field order while malformed inputs are reported separately. Observed problem: A raw CSV field contains quoted commas and a header is sometimes missing, so naïve delimiter splitting shifts the label into the wrong column. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
C:Every held-out forecast uses only data that existed at prediction time and its backtest follows real ordering. This could validate chronological split for forecast evaluation, but it does not establish that csv schema ambiguity during ingestion is resolved for this workload. Confirm instead: Validated records preserve the expected label and field order while malformed inputs are reported separately. Observed problem: A raw CSV field contains quoted commas and a header is sometimes missing, so naïve delimiter splitting shifts the label into the wrong column. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
D:Replaying the same notification leaves exactly one corresponding logical training record and records deduplication. This could validate s3 object event at-least-once processing, but it does not establish that csv schema ambiguity during ingestion is resolved for this workload. Confirm instead: Validated records preserve the expected label and field order while malformed inputs are reported separately. Observed problem: A raw CSV field contains quoted commas and a header is sometimes missing, so naïve delimiter splitting shifts the label into the wrong column. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
Question 223
During an approved rollout of an AWS ML inference workload, the following condition is observed: A spiky serverless ML endpoint exceeds customer latency targets on cold starts even though average compute utilization is low. The team must meet its documented availability and rollback constraints. What postchange observation provides the strongest acceptance evidence?
- Identical deployment revisions resolve to the same container image digest and environment dependencies.
- Repeated runs with the same inputs reproduce the ordered processing and generate auditable versioned outputs.
- Cold-start p95 improves during planned demand windows while capacity cost remains measured.
- Distinct requested model IDs return correct predictions with acceptable cold load latency and lower aggregate idle cost.
Explanation
Correct Answer
C:Cold-start p95 improves during planned demand windows while capacity cost remains measured. This is the acceptance evidence for the described sagemaker endpoint provisioned concurrency tradeoff scenario. Provisioned concurrency can reduce cold-start effects but adds reserved cost and must match supported endpoint mode. The corrective implementation it validates is: Configure supported provisioned concurrency for the latency-critical interval or switch to provisioned real-time hosting if justified. Reported signal: A spiky serverless ML endpoint exceeds customer latency targets on cold starts even though average compute utilization is low. Operational setting: During an approved rollout of an AWS ML inference workload.
Incorrect Answers
A:Identical deployment revisions resolve to the same container image digest and environment dependencies. This could validate ecr immutable image digest for inference, but it does not establish that sagemaker endpoint provisioned concurrency tradeoff is resolved for this workload. Confirm instead: Cold-start p95 improves during planned demand windows while capacity cost remains measured. Observed problem: A spiky serverless ML endpoint exceeds customer latency targets on cold starts even though average compute utilization is low. Operational setting: During an approved rollout of an AWS ML inference workload.
B:Repeated runs with the same inputs reproduce the ordered processing and generate auditable versioned outputs. This could validate sagemaker pipelines parameterized retraining, but it does not establish that sagemaker endpoint provisioned concurrency tradeoff is resolved for this workload. Confirm instead: Cold-start p95 improves during planned demand windows while capacity cost remains measured. Observed problem: A spiky serverless ML endpoint exceeds customer latency targets on cold starts even though average compute utilization is low. Operational setting: During an approved rollout of an AWS ML inference workload.
D:Distinct requested model IDs return correct predictions with acceptable cold load latency and lower aggregate idle cost. This could validate multi-model endpoint hosting tradeoff, but it does not establish that sagemaker endpoint provisioned concurrency tradeoff is resolved for this workload. Confirm instead: Cold-start p95 improves during planned demand windows while capacity cost remains measured. Observed problem: A spiky serverless ML endpoint exceeds customer latency targets on cold starts even though average compute utilization is low. Operational setting: During an approved rollout of an AWS ML inference workload.
Question 224
At a production AWS ML data lake supporting customer-facing predictions, the following condition is observed: A real-time feature consumer lags only for one tenant because producers use the same partition key for most high-volume events. Engineers must avoid changing unrelated source partitions. Which investigation provides the most decisive FIRST evidence for the underlying problem?
- Inspect source fields, sensitivity classification, Ground Truth task input, and downstream annotation permissions.
- Inspect the stream retention setting, consumer checkpoint timestamps, and the oldest available sequence or time position.
- Compare per-shard incoming records, write throttles, partition-key cardinality, and enhanced fan-out consumer lag.
- Inspect feature availability timestamps and whether each label time follows the last feature observation.
Explanation
Correct Answer
C:Compare per-shard incoming records, write throttles, partition-key cardinality, and enhanced fan-out consumer lag. This directly tests kinesis data streams shard hot key in the reported AWS ML workflow. Kinesis shard throughput is bounded; a hot partition key can overload one shard despite spare total stream capacity. It distinguishes the problem before choosing a change. The related targeted response is: Choose partition keys with enough cardinality or reshard the stream for the intended throughput, while preserving required per-key ordering. Reported signal: A real-time feature consumer lags only for one tenant because producers use the same partition key for most high-volume events. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Incorrect Answers
A:Inspect source fields, sensitivity classification, Ground Truth task input, and downstream annotation permissions. This is a possible diagnostic for pii masking before model labeling, but it does not directly test the kinesis data streams shard hot key failure reported here. Prioritize: Compare per-shard incoming records, write throttles, partition-key cardinality, and enhanced fan-out consumer lag. The relevant case is: A real-time feature consumer lags only for one tenant because producers use the same partition key for most high-volume events. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
B: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 kinesis data streams shard hot key failure reported here. Prioritize: Compare per-shard incoming records, write throttles, partition-key cardinality, and enhanced fan-out consumer lag. The relevant case is: A real-time feature consumer lags only for one tenant because producers use the same partition key for most high-volume events. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
D:Inspect feature availability timestamps and whether each label time follows the last feature observation. This is a possible diagnostic for chronological split for forecast evaluation, but it does not directly test the kinesis data streams shard hot key failure reported here. Prioritize: Compare per-shard incoming records, write throttles, partition-key cardinality, and enhanced fan-out consumer lag. The relevant case is: A real-time feature consumer lags only for one tenant because producers use the same partition key for most high-volume events. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Question 225
At a production AWS ML data lake supporting customer-facing predictions, 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. Engineers must avoid changing unrelated source partitions. What postchange observation provides the strongest acceptance evidence?
- Historical evaluation reproduces features available at prediction time, and production freshness is monitored against the same contract.
- Concurrent workers observe consistent supported shared-file operations and training artifact delivery remains durable.
- Two ordinary runs ingest only newly eligible objects while a separately invoked backfill handles prior intervals.
- Neighboring times across midnight map close in cyclic feature space and validation improves without leakage.
Explanation
Correct Answer
A:Historical evaluation reproduces features available at prediction time, and production freshness is monitored against the same contract. This is the acceptance evidence for the described sagemaker feature store online offline consistency scenario. Online current state and offline historical records answer different temporal queries; blind latest-value joins cause leakage. The corrective implementation it validates is: Define point-in-time offline joins, freshness SLAs, and aligned feature version logic for online retrieval. 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: At a production AWS ML data lake supporting customer-facing predictions.
Incorrect Answers
B:Concurrent workers observe consistent supported shared-file operations and training artifact delivery remains durable. This could validate s3 versus efs training storage choice, but it does not establish that sagemaker feature store online offline consistency is resolved for this workload. Confirm instead: Historical evaluation reproduces features available at prediction time, and production freshness is monitored against the same contract. Observed problem: A fraud endpoint fetches the latest online customer feature while offline training joins an earlier feature snapshot, creating training-serving skew. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
C:Two ordinary runs ingest only newly eligible objects while a separately invoked backfill handles prior intervals. This could validate glue job bookmark state error, but it does not establish that sagemaker feature store online offline consistency is resolved for this workload. Confirm instead: Historical evaluation reproduces features available at prediction time, and production freshness is monitored against the same contract. Observed problem: A fraud endpoint fetches the latest online customer feature while offline training joins an earlier feature snapshot, creating training-serving skew. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
D:Neighboring times across midnight map close in cyclic feature space and validation improves without leakage. This could validate time-series cyclical calendar feature, but it does not establish that sagemaker feature store online offline consistency is resolved for this workload. Confirm instead: Historical evaluation reproduces features available at prediction time, and production freshness is monitored against the same contract. Observed problem: A fraud endpoint fetches the latest online customer feature while offline training joins an earlier feature snapshot, creating training-serving skew. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Question 226
While reproducing the alert in an isolated AWS ML integration account, the following condition is observed: A model's overall approval rate remains stable while approval differences between customer groups increase after a data-source migration. The proposed change must preserve logging and least privilege. What postchange observation provides the strongest acceptance evidence?
- Peak-window 5xx rate and p95 latency fall below service objectives with verified alarms.
- 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.
- Cost reports attribute the majority of in-scope ML spend to the correct team without altering access controls.
Explanation
Correct Answer
B:Auditors see reproducible subgroup statistics and alarms for material disparities rather than only global totals. This is the acceptance evidence for the described clarify postdeployment group bias monitoring scenario. Aggregate performance can mask group-specific shifts; fairness monitoring must examine appropriate group metrics. The corrective implementation it validates is: Use appropriate SageMaker Clarify model-bias monitoring and investigate supported group differences before deployment changes. Reported signal: A model's overall approval rate remains stable while approval differences between customer groups increase after a data-source migration. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
Incorrect Answers
A:Peak-window 5xx rate and p95 latency fall below service objectives with verified alarms. This could validate cloudwatch 5xxerrors endpoint saturation, but it does not establish that clarify postdeployment group bias monitoring is resolved for this workload. Confirm instead: Auditors see reproducible subgroup statistics and alarms for material disparities rather than only global totals. Observed problem: A model's overall approval rate remains stable while approval differences between customer groups increase after a data-source migration. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
C:Representative benign fluctuations stay quiet while a known severe schema or distribution break pages the owner. This could validate model monitor violation alert calibration, but it does not establish that clarify postdeployment group bias monitoring is resolved for this workload. Confirm instead: Auditors see reproducible subgroup statistics and alarms for material disparities rather than only global totals. Observed problem: A model's overall approval rate remains stable while approval differences between customer groups increase after a data-source migration. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
D:Cost reports attribute the majority of in-scope ML spend to the correct team without altering access controls. This could validate tag-based ml training cost attribution, but it does not establish that clarify postdeployment group bias monitoring is resolved for this workload. Confirm instead: Auditors see reproducible subgroup statistics and alarms for material disparities rather than only global totals. Observed problem: A model's overall approval rate remains stable while approval differences between customer groups increase after a data-source migration. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
Question 227
At a production AWS ML data lake supporting customer-facing predictions, 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. Engineers must avoid changing unrelated source partitions. What should the ML engineer inspect FIRST to distinguish this failure from unrelated issues?
- Compare online Feature Store record values and offline event-time keyed feature history for the same entity and model snapshot.
- Inspect prediction errors around midnight and feature representations for the periodic hour variable.
- Inspect AWS Glue Data Quality rule evaluations, missing-key rates, source partitions, and prior accepted baselines.
- Inspect workload read/write semantics, shared access requirements, filesystem operations, and object-store request patterns.
Explanation
Correct Answer
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. 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: At a production AWS ML data lake supporting customer-facing predictions.
Incorrect Answers
B:Inspect prediction errors around midnight and feature representations for the periodic hour variable. This is a possible diagnostic for time-series cyclical calendar feature, but it does not directly test the sagemaker feature store online offline consistency failure reported here. Prioritize: Compare online Feature Store record values and offline event-time keyed feature history for the same entity and model snapshot. The relevant case is: A fraud endpoint fetches the latest online customer feature while offline training joins an earlier feature snapshot, creating training-serving skew. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
C: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 sagemaker feature store online offline consistency failure reported here. Prioritize: Compare online Feature Store record values and offline event-time keyed feature history for the same entity and model snapshot. The relevant case is: A fraud endpoint fetches the latest online customer feature while offline training joins an earlier feature snapshot, creating training-serving skew. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
D:Inspect workload read/write semantics, shared access requirements, filesystem operations, and object-store request patterns. This is a possible diagnostic for s3 versus efs training storage choice, but it does not directly test the sagemaker feature store online offline consistency failure reported here. Prioritize: Compare online Feature Store record values and offline event-time keyed feature history for the same entity and model snapshot. The relevant case is: A fraud endpoint fetches the latest online customer feature while offline training joins an earlier feature snapshot, creating training-serving skew. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Question 228
During an approved rollout of an AWS ML inference workload, the following condition is observed: A spiky serverless ML endpoint exceeds customer latency targets on cold starts even though average compute utilization is low. The team must meet its documented availability and rollback constraints. Which TWO steps best establish the evidence and correct the identified root cause? Select TWO.
- Inspect model registry package versions, approval metadata, artifact hashes, and pipeline deployment inputs.
- Add schema, integration and sample endpoint inference tests to the ML build and deployment pipeline.
- Configure supported provisioned concurrency for the latency-critical interval or switch to provisioned real-time hosting if justified.
- Set adequate minimum capacity, scaling limits, alarms, and request backpressure for the expected traffic ramp.
- Inspect serverless invocation frequency, cold-start latency, supported provisioned concurrency settings, and cost.
Explanation
Correct Answers
C:Configure supported provisioned concurrency for the latency-critical interval or switch to provisioned real-time hosting if justified. This is the appropriate controlled remedy for sagemaker endpoint provisioned concurrency tradeoff. Provisioned concurrency can reduce cold-start effects but adds reserved cost and must match supported endpoint mode. Verify the cause with: Inspect serverless invocation frequency, cold-start latency, supported provisioned concurrency settings, and cost. Then confirm: Cold-start p95 improves during planned demand windows while capacity cost remains measured. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A spiky serverless ML endpoint exceeds customer latency targets on cold starts even though average compute utilization is low. Operational setting: During an approved rollout of an AWS ML inference workload.
E:Inspect serverless invocation frequency, cold-start latency, supported provisioned concurrency settings, and cost. This directly tests sagemaker endpoint provisioned concurrency tradeoff in the reported AWS ML workflow. Provisioned concurrency can reduce cold-start effects but adds reserved cost and must match supported endpoint mode. It distinguishes the problem before choosing a change. The related targeted response is: Configure supported provisioned concurrency for the latency-critical interval or switch to provisioned real-time hosting if justified. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A spiky serverless ML endpoint exceeds customer latency targets on cold starts even though average compute utilization is low. Operational setting: During an approved rollout of an AWS ML inference workload.
Incorrect Answers
A:Inspect model registry package versions, approval metadata, artifact hashes, and pipeline deployment inputs. This selection concerns model registry version promotion provenance rather than the required diagnostic-and-remedy pair for sagemaker endpoint provisioned concurrency tradeoff. Use evidence: Inspect serverless invocation frequency, cold-start latency, supported provisioned concurrency settings, and cost. and the matching correction: Configure supported provisioned concurrency for the latency-critical interval or switch to provisioned real-time hosting if justified. Signal: A spiky serverless ML endpoint exceeds customer latency targets on cold starts even though average compute utilization is low. Operational setting: During an approved rollout of an AWS ML inference workload.
B:Add schema, integration and sample endpoint inference tests to the ML build and deployment pipeline. This selection concerns continuous integration unit and endpoint contract tests rather than the required diagnostic-and-remedy pair for sagemaker endpoint provisioned concurrency tradeoff. Use evidence: Inspect serverless invocation frequency, cold-start latency, supported provisioned concurrency settings, and cost. and the matching correction: Configure supported provisioned concurrency for the latency-critical interval or switch to provisioned real-time hosting if justified. Signal: A spiky serverless ML endpoint exceeds customer latency targets on cold starts even though average compute utilization is low. Operational setting: During an approved rollout of an AWS ML inference workload.
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 sagemaker endpoint provisioned concurrency tradeoff. Use evidence: Inspect serverless invocation frequency, cold-start latency, supported provisioned concurrency settings, and cost. and the matching correction: Configure supported provisioned concurrency for the latency-critical interval or switch to provisioned real-time hosting if justified. Signal: A spiky serverless ML endpoint exceeds customer latency targets on cold starts even though average compute utilization is low. Operational setting: During an approved rollout of an AWS ML inference workload.
Question 229
While reproducing the alert in an isolated AWS ML integration account, the following condition is observed: An endpoint's client-side response time rises, but the team only examines SageMaker ModelLatency and concludes the model is slow. The proposed change must preserve logging and least privilege. Which engineering response resolves the actual issue without masking the symptoms?
- Enable appropriate CloudTrail S3 data events for the sensitive dataset with protected storage and retention.
- Use approved SageMaker interface endpoints and relevant private DNS and policy settings for required API access.
- Set budget alerts and bounded tuning job quotas before launching experiments, and enforce an owner-approved stopping policy.
- Identify the component contributing to total request time and optimize network or overhead separately from actual model inference.
Explanation
Correct Answer
D:Identify the component contributing to total request time and optimize network or overhead separately from actual model inference. This is the appropriate controlled remedy for cloudwatch invocations and modellatency distinction. Model compute time and overall endpoint latency are distinct signals; blaming model code without tracing overhead is premature. Verify the cause with: Compare ModelLatency with OverheadLatency, client network timing, invocation counts, and endpoint 5XX responses. Then confirm: The chosen fix reduces measured end-to-end p95 latency and component metrics explain the change. Reported signal: An endpoint's client-side response time rises, but the team only examines SageMaker ModelLatency and concludes the model is slow. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
Incorrect Answers
A: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 cloudwatch invocations and modellatency distinction mechanism. The targeted response is: Identify the component contributing to total request time and optimize network or overhead separately from actual model inference. Case evidence: An endpoint's client-side response time rises, but the team only examines SageMaker ModelLatency and concludes the model is slow. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
B:Use approved SageMaker interface endpoints and relevant private DNS and policy settings for required API access. This addresses a different issue involving private vpc endpoint policy for sagemaker apis. It does not correct the stated cloudwatch invocations and modellatency distinction mechanism. The targeted response is: Identify the component contributing to total request time and optimize network or overhead separately from actual model inference. Case evidence: An endpoint's client-side response time rises, but the team only examines SageMaker ModelLatency and concludes the model is slow. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
C: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 cloudwatch invocations and modellatency distinction mechanism. The targeted response is: Identify the component contributing to total request time and optimize network or overhead separately from actual model inference. Case evidence: An endpoint's client-side response time rises, but the team only examines SageMaker ModelLatency and concludes the model is slow. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
Question 230
As a regulated ML workload prepares for an automated regional rollout, 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 team must demonstrate a repeatable monitoring or security control. Which TWO complementary choices provide the decisive diagnosis AND corrective response? Select TWO.
- Separate prompt input-distribution monitoring from delayed model-quality evaluation and join only observed labels to predictions.
- Inspect identity and destination bucket policies, resource ARNs, organization restrictions, and CloudTrail data-event evidence.
- Restrict writes to approved bucket prefixes and apply appropriate resource and organizational guardrails.
- Inspect online versus offline feature computation code, event-time cutoffs, freshness, and sampled entity values.
- Use approved SageMaker interface endpoints and relevant private DNS and policy settings for required API access.
Explanation
Correct Answers
B: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. This is one part of a select-two decision; the matching evidence and corrective action are both required. 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: As a regulated ML workload prepares for an automated regional rollout.
C:Restrict writes to approved bucket prefixes and apply appropriate resource and organizational guardrails. This is the appropriate controlled remedy for cross-account s3 bucket exfiltration protection. Least privilege must constrain the destination resource; an unrestricted object-write policy can enable data exfiltration. Verify the cause with: Inspect identity and destination bucket policies, resource ARNs, organization restrictions, and CloudTrail data-event evidence. Then confirm: A test write to an approved prefix succeeds, while writes to unapproved external destinations are denied. This is one part of a select-two decision; the matching evidence and corrective action are both required. 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: As a regulated ML workload prepares for an automated regional rollout.
Incorrect Answers
A:Separate prompt input-distribution monitoring from delayed model-quality evaluation and join only observed labels to predictions. This selection concerns target-label delay in production quality monitoring rather than the required diagnostic-and-remedy pair for cross-account s3 bucket exfiltration protection. Use evidence: Inspect identity and destination bucket policies, resource ARNs, organization restrictions, and CloudTrail data-event evidence. and the matching correction: Restrict writes to approved bucket prefixes and apply appropriate resource and organizational guardrails. 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: As a regulated ML workload prepares for an automated regional rollout.
D:Inspect online versus offline feature computation code, event-time cutoffs, freshness, and sampled entity values. This selection concerns training-serving feature skew incident rather than the required diagnostic-and-remedy pair for cross-account s3 bucket exfiltration protection. Use evidence: Inspect identity and destination bucket policies, resource ARNs, organization restrictions, and CloudTrail data-event evidence. and the matching correction: Restrict writes to approved bucket prefixes and apply appropriate resource and organizational guardrails. 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: As a regulated ML workload prepares for an automated regional rollout.
E:Use approved SageMaker interface endpoints and relevant private DNS and policy settings for required API access. This selection concerns private vpc endpoint policy for sagemaker apis rather than the required diagnostic-and-remedy pair for cross-account s3 bucket exfiltration protection. Use evidence: Inspect identity and destination bucket policies, resource ARNs, organization restrictions, and CloudTrail data-event evidence. and the matching correction: Restrict writes to approved bucket prefixes and apply appropriate resource and organizational guardrails. 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: As a regulated ML workload prepares for an automated regional rollout.
