Amazon AWS Certified Machine Learning Engineer - Associate MLA-C01 Practice Test 74
Question 730
During an approved rollout of an AWS ML inference workload, the following condition is observed: A real-time ML service is configured for very low baseline capacity but traffic can surge tenfold within seconds, exhausting available instance capacity. The team must meet its documented availability and rollback constraints. Which investigation provides the most decisive FIRST evidence for the underlying problem?
- Inspect ECR image digest, tag history, pipeline manifests, and built-image package inventory.
- Inspect service quotas, minimum and maximum endpoint capacity, scaling cooldowns, startup time, and burst size.
- Inspect production variant weights, guardrail metrics, rollback procedure, and representativeness of staged exposure.
- Inspect model sizes, framework compatibility, per-model request rates, load latency, and isolation requirements.
Explanation
Correct Answer
B:Inspect service quotas, minimum and maximum endpoint capacity, scaling cooldowns, startup time, and burst size. This directly tests endpoint scale-to-zero cold capacity planning in the reported AWS ML workflow. Reactive scaling takes time and is bounded by quotas; zero spare capacity may not absorb an instantaneous burst. It distinguishes the problem before choosing a change. The related targeted response is: Set adequate minimum capacity, scaling limits, alarms, and request backpressure for the expected traffic ramp. Reported signal: A real-time ML service is configured for very low baseline capacity but traffic can surge tenfold within seconds, exhausting available instance capacity. Operational setting: During an approved rollout of an AWS ML inference workload.
Incorrect Answers
A:Inspect ECR image digest, tag history, pipeline manifests, and built-image package inventory. This is a possible diagnostic for ecr immutable image digest for inference, but it does not directly test the endpoint scale-to-zero cold capacity planning failure reported here. Prioritize: Inspect service quotas, minimum and maximum endpoint capacity, scaling cooldowns, startup time, and burst size. The relevant case is: A real-time ML service is configured for very low baseline capacity but traffic can surge tenfold within seconds, exhausting available instance capacity. Operational setting: During an approved rollout of an AWS ML inference workload.
C:Inspect production variant weights, guardrail metrics, rollback procedure, and representativeness of staged exposure. This is a possible diagnostic for canary variant rollout for model changes, but it does not directly test the endpoint scale-to-zero cold capacity planning failure reported here. Prioritize: Inspect service quotas, minimum and maximum endpoint capacity, scaling cooldowns, startup time, and burst size. The relevant case is: A real-time ML service is configured for very low baseline capacity but traffic can surge tenfold within seconds, exhausting available instance capacity. Operational setting: During an approved rollout of an AWS ML inference workload.
D:Inspect model sizes, framework compatibility, per-model request rates, load latency, and isolation requirements. This is a possible diagnostic for multi-model endpoint hosting tradeoff, but it does not directly test the endpoint scale-to-zero cold capacity planning failure reported here. Prioritize: Inspect service quotas, minimum and maximum endpoint capacity, scaling cooldowns, startup time, and burst size. The relevant case is: A real-time ML service is configured for very low baseline capacity but traffic can surge tenfold within seconds, exhausting available instance capacity. Operational setting: During an approved rollout of an AWS ML inference workload.
Question 731
As a regulated ML workload prepares for an automated regional rollout, the following condition is observed: An organization cannot allocate expensive SageMaker training jobs among departments because jobs and dependent storage lack common cost allocation tags. The team must demonstrate a repeatable monitoring or security control. What postchange observation provides the strongest acceptance evidence?
- An authorized training read emits attributable S3 object-access events while alerting detects unexpected access.
- Model quality updates only for labeled cohorts while latency and feature drift are monitored promptly.
- Backlog age stabilizes and requests complete within the agreed service-level window.
- Cost reports attribute the majority of in-scope ML spend to the correct team without altering access controls.
Explanation
Correct Answer
D:Cost reports attribute the majority of in-scope ML spend to the correct team without altering access controls. This is the acceptance evidence for the described tag-based ml training cost attribution scenario. Cost Explorer grouping requires supported activated resource tags; naming resources does not automatically allocate charges. The corrective implementation it validates is: Apply consistent approved cost allocation tags to ML jobs and supported linked resources and activate relevant reporting. Reported signal: An organization cannot allocate expensive SageMaker training jobs among departments because jobs and dependent storage lack common cost allocation tags. Operational setting: As a regulated ML workload prepares for an automated regional rollout.
Incorrect Answers
A:An authorized training read emits attributable S3 object-access events while alerting detects unexpected access. This could validate cloudtrail training data access audit coverage, but it does not establish that tag-based ml training cost attribution is resolved for this workload. Confirm instead: Cost reports attribute the majority of in-scope ML spend to the correct team without altering access controls. Observed problem: An organization cannot allocate expensive SageMaker training jobs among departments because jobs and dependent storage lack common cost allocation tags. Operational setting: As a regulated ML workload prepares for an automated regional rollout.
B:Model quality updates only for labeled cohorts while latency and feature drift are monitored promptly. This could validate target-label delay in production quality monitoring, but it does not establish that tag-based ml training cost attribution is resolved for this workload. Confirm instead: Cost reports attribute the majority of in-scope ML spend to the correct team without altering access controls. Observed problem: An organization cannot allocate expensive SageMaker training jobs among departments because jobs and dependent storage lack common cost allocation tags. Operational setting: As a regulated ML workload prepares for an automated regional rollout.
C:Backlog age stabilizes and requests complete within the agreed service-level window. This could validate asynchronous inference queue backlog alert, but it does not establish that tag-based ml training cost attribution is resolved for this workload. Confirm instead: Cost reports attribute the majority of in-scope ML spend to the correct team without altering access controls. Observed problem: An organization cannot allocate expensive SageMaker training jobs among departments because jobs and dependent storage lack common cost allocation tags. Operational setting: As a regulated ML workload prepares for an automated regional rollout.
Question 732
While reproducing the alert in an isolated AWS ML integration account, the following condition is observed: A real-time recommender fetches cached customer history updated hourly while offline training used a daily snapshot with different aggregation boundaries. The proposed change must preserve logging and least privilege. Select TWO: the appropriate diagnostic check and the targeted implementation change.
- Use a consistent versioned feature definition and test offline/online values at the same observation cutoff.
- Inspect online versus offline feature computation code, event-time cutoffs, freshness, and sampled entity values.
- Identify the component contributing to total request time and optimize network or overhead separately from actual model inference.
- Capture stable prediction identifiers and ingest verified labels with matching keys for model-quality analysis.
- Inspect training interruption tolerance, persistent checkpoints, allowed instance families, and deadline requirements.
Explanation
Correct Answers
A:Use a consistent versioned feature definition and test offline/online values at the same observation cutoff. This is the appropriate controlled remedy for training-serving feature skew incident. Mismatched feature windows and update cadence can degrade production predictions even if training accuracy is stable. Verify the cause with: Inspect online versus offline feature computation code, event-time cutoffs, freshness, and sampled entity values. Then confirm: Parity checks match transformations within documented freshness tolerances for test entities. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A real-time recommender fetches cached customer history updated hourly while offline training used a daily snapshot with different aggregation boundaries. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
B:Inspect online versus offline feature computation code, event-time cutoffs, freshness, and sampled entity values. This directly tests training-serving feature skew incident in the reported AWS ML workflow. Mismatched feature windows and update cadence can degrade production predictions even if training accuracy is stable. It distinguishes the problem before choosing a change. The related targeted response is: Use a consistent versioned feature definition and test offline/online values at the same observation cutoff. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A real-time recommender fetches cached customer history updated hourly while offline training used a daily snapshot with different aggregation boundaries. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
Incorrect Answers
C:Identify the component contributing to total request time and optimize network or overhead separately from actual model inference. This selection concerns cloudwatch invocations and modellatency distinction rather than the required diagnostic-and-remedy pair for training-serving feature skew incident. Use evidence: Inspect online versus offline feature computation code, event-time cutoffs, freshness, and sampled entity values. and the matching correction: Use a consistent versioned feature definition and test offline/online values at the same observation cutoff. Signal: A real-time recommender fetches cached customer history updated hourly while offline training used a daily snapshot with different aggregation boundaries. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
D:Capture stable prediction identifiers and ingest verified labels with matching keys for model-quality analysis. This selection concerns model monitor ground truth association rather than the required diagnostic-and-remedy pair for training-serving feature skew incident. Use evidence: Inspect online versus offline feature computation code, event-time cutoffs, freshness, and sampled entity values. and the matching correction: Use a consistent versioned feature definition and test offline/online values at the same observation cutoff. Signal: A real-time recommender fetches cached customer history updated hourly while offline training used a daily snapshot with different aggregation boundaries. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
E:Inspect training interruption tolerance, persistent checkpoints, allowed instance families, and deadline requirements. This selection concerns spot training cost versus interruption recovery rather than the required diagnostic-and-remedy pair for training-serving feature skew incident. Use evidence: Inspect online versus offline feature computation code, event-time cutoffs, freshness, and sampled entity values. and the matching correction: Use a consistent versioned feature definition and test offline/online values at the same observation cutoff. Signal: A real-time recommender fetches cached customer history updated hourly while offline training used a daily snapshot with different aggregation boundaries. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
Question 733
At the model-governance gate before promotion of a new ML artifact, the following condition is observed: Two notebooks deploy different model artifacts under the same informal name, so the team cannot reproduce which code and data generated production predictions. The team needs a defensible technical decision rather than an untracked workaround. Which investigation provides the most decisive FIRST evidence for the underlying problem?
- Inspect model package version, artifact URI, training metadata, data snapshot, and registry approval state.
- Inspect tensor summaries, gradient magnitudes, learning-rate schedule, and training loss anomalies with supported debugging hooks.
- Inspect confusion matrix, class prevalence, false-negative cost, and precision-recall curves across thresholds.
- Inspect interruption events, checkpoint directory mapping, restoration logic, and persistent S3 checkpoint artifacts.
Explanation
Correct Answer
A:Inspect model package version, artifact URI, training metadata, data snapshot, and registry approval state. This directly tests sagemaker model registry candidate approval in the reported AWS ML workflow. A human-readable model nickname is not a reproducible deployment identifier; versioned registry artifacts support controlled promotion. It distinguishes the problem before choosing a change. The related targeted response is: Register immutable model package versions with lineage and require explicit approval before promotion to production. Reported signal: Two notebooks deploy different model artifacts under the same informal name, so the team cannot reproduce which code and data generated production predictions. Operational setting: At the model-governance gate before promotion of a new ML artifact.
Incorrect Answers
B:Inspect tensor summaries, gradient magnitudes, learning-rate schedule, and training loss anomalies with supported debugging hooks. This is a possible diagnostic for model debugger exploding gradient diagnosis, but it does not directly test the sagemaker model registry candidate approval failure reported here. Prioritize: Inspect model package version, artifact URI, training metadata, data snapshot, and registry approval state. The relevant case is: Two notebooks deploy different model artifacts under the same informal name, so the team cannot reproduce which code and data generated production predictions. Operational setting: At the model-governance gate before promotion of a new ML artifact.
C:Inspect confusion matrix, class prevalence, false-negative cost, and precision-recall curves across thresholds. This is a possible diagnostic for precision versus recall fraud threshold, but it does not directly test the sagemaker model registry candidate approval failure reported here. Prioritize: Inspect model package version, artifact URI, training metadata, data snapshot, and registry approval state. The relevant case is: Two notebooks deploy different model artifacts under the same informal name, so the team cannot reproduce which code and data generated production predictions. Operational setting: At the model-governance gate before promotion of a new ML artifact.
D:Inspect interruption events, checkpoint directory mapping, restoration logic, and persistent S3 checkpoint artifacts. This is a possible diagnostic for managed spot training checkpoint recovery, but it does not directly test the sagemaker model registry candidate approval failure reported here. Prioritize: Inspect model package version, artifact URI, training metadata, data snapshot, and registry approval state. The relevant case is: Two notebooks deploy different model artifacts under the same informal name, so the team cannot reproduce which code and data generated production predictions. Operational setting: At the model-governance gate before promotion of a new ML artifact.
Question 734
At a production AWS ML data lake supporting customer-facing predictions, the following condition is observed: An offline churn dataset uses the latest cancellation-status feature for all historical rows, exposing outcomes not known when earlier predictions would run. Engineers must avoid changing unrelated source partitions. Which measured outcome would confirm that the correction worked as intended?
- Memory use stays within limits and previously unseen categories are processed without dimension mismatch.
- For every training row the joined feature timestamp is at or before its prediction timestamp and no label-derived value leaks.
- Replaying the same notification leaves exactly one corresponding logical training record and records deduplication.
- Every held-out forecast uses only data that existed at prediction time and its backtest follows real ordering.
Explanation
Correct Answer
B:For every training row the joined feature timestamp is at or before its prediction timestamp and no label-derived value leaks. This is the acceptance evidence for the described point-in-time feature data leakage scenario. A feature created after the prediction event cannot legitimately inform that earlier prediction. The corrective implementation it validates is: Perform point-in-time correct joins using the most recent eligible pre-prediction feature record, not the globally latest value. Reported signal: An offline churn dataset uses the latest cancellation-status feature for all historical rows, exposing outcomes not known when earlier predictions would run. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Incorrect Answers
A:Memory use stays within limits and previously unseen categories are processed without dimension mismatch. This could validate high-cardinality categorical encoding, but it does not establish that point-in-time feature data leakage is resolved for this workload. Confirm instead: For every training row the joined feature timestamp is at or before its prediction timestamp and no label-derived value leaks. Observed problem: An offline churn dataset uses the latest cancellation-status feature for all historical rows, exposing outcomes not known when earlier predictions would run. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
C: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 point-in-time feature data leakage is resolved for this workload. Confirm instead: For every training row the joined feature timestamp is at or before its prediction timestamp and no label-derived value leaks. Observed problem: An offline churn dataset uses the latest cancellation-status feature for all historical rows, exposing outcomes not known when earlier predictions would run. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
D: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 point-in-time feature data leakage is resolved for this workload. Confirm instead: For every training row the joined feature timestamp is at or before its prediction timestamp and no label-derived value leaks. Observed problem: An offline churn dataset uses the latest cancellation-status feature for all historical rows, exposing outcomes not known when earlier predictions would run. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Question 735
During an operational load and release rehearsal for SageMaker-hosted inference, the following condition is observed: A private-subnet inference deployment cannot retrieve its S3 model artifact because the network has no authorized path to the bucket. Change control requires measurable endpoint or pipeline acceptance evidence. What postchange observation provides the strongest acceptance evidence?
- Identical deployment revisions resolve to the same container image digest and environment dependencies.
- Distinct requested model IDs return correct predictions with acceptable cold load latency and lower aggregate idle cost.
- The deployment loads artifacts privately and blocked public routes remain unused.
- The compiled artifact runs on the target device with validated outputs and acceptable memory and latency.
Explanation
Correct Answer
C:The deployment loads artifacts privately and blocked public routes remain unused. This is the acceptance evidence for the described sagemaker vpc endpoint access for private artifacts scenario. VPC isolation requires explicit supported network paths to AWS services; IAM permission alone does not create connectivity. The corrective implementation it validates is: Provide an approved S3 gateway or interface endpoint path and least-privilege bucket/role permissions without unnecessarily opening internet egress. Reported signal: A private-subnet inference deployment cannot retrieve its S3 model artifact because the network has no authorized path to the bucket. Operational setting: During an operational load and release rehearsal for SageMaker-hosted inference.
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 vpc endpoint access for private artifacts is resolved for this workload. Confirm instead: The deployment loads artifacts privately and blocked public routes remain unused. Observed problem: A private-subnet inference deployment cannot retrieve its S3 model artifact because the network has no authorized path to the bucket. Operational setting: During an operational load and release rehearsal for SageMaker-hosted inference.
B: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 vpc endpoint access for private artifacts is resolved for this workload. Confirm instead: The deployment loads artifacts privately and blocked public routes remain unused. Observed problem: A private-subnet inference deployment cannot retrieve its S3 model artifact because the network has no authorized path to the bucket. Operational setting: During an operational load and release rehearsal for SageMaker-hosted inference.
D:The compiled artifact runs on the target device with validated outputs and acceptable memory and latency. This could validate sagemaker neo edge compiler compatibility, but it does not establish that sagemaker vpc endpoint access for private artifacts is resolved for this workload. Confirm instead: The deployment loads artifacts privately and blocked public routes remain unused. Observed problem: A private-subnet inference deployment cannot retrieve its S3 model artifact because the network has no authorized path to the bucket. Operational setting: During an operational load and release rehearsal for SageMaker-hosted inference.
Question 736
As a regional data-pipeline rollout is validated against approved AWS storage and streams, the following condition is observed: New event producers change a numerical source field to strings while legacy objects remain numeric, causing the shared Glue transformation to fail. The data team must preserve event-time meaning and lineage. Which measured outcome would confirm that the correction worked as intended?
- Application p95 latency returns to target while exported records meet a measured completeness and staleness objective.
- Annotation workers can label required content without viewing direct identifiers and authorized audits can trace approved mappings.
- A held-out split uses the training-fitted median and independent evaluation no longer incorporates held-out distribution values.
- Both supported source versions produce the agreed typed feature, and invalid values enter an auditable reject path.
Explanation
Correct Answer
D:Both supported source versions produce the agreed typed feature, and invalid values enter an auditable reject path. This is the acceptance evidence for the described glue schema evolution with mixed input types scenario. A catalog entry does not automatically reconcile incompatible field types across all source partitions. The corrective implementation it validates is: Define explicit compatible conversions and schema contracts, isolate incompatible versions, and validate before publishing features. Reported signal: New event producers change a numerical source field to strings while legacy objects remain numeric, causing the shared Glue transformation to fail. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
Incorrect Answers
A:Application p95 latency returns to target while exported records meet a measured completeness and staleness objective. This could validate rds extraction overload into ml, but it does not establish that glue schema evolution with mixed input types is resolved for this workload. Confirm instead: Both supported source versions produce the agreed typed feature, and invalid values enter an auditable reject path. Observed problem: New event producers change a numerical source field to strings while legacy objects remain numeric, causing the shared Glue transformation to fail. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
B:Annotation workers can label required content without viewing direct identifiers and authorized audits can trace approved mappings. This could validate pii masking before model labeling, but it does not establish that glue schema evolution with mixed input types is resolved for this workload. Confirm instead: Both supported source versions produce the agreed typed feature, and invalid values enter an auditable reject path. Observed problem: New event producers change a numerical source field to strings while legacy objects remain numeric, causing the shared Glue transformation to fail. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
C:A held-out split uses the training-fitted median and independent evaluation no longer incorporates held-out distribution values. This could validate median imputation fitted on training split, but it does not establish that glue schema evolution with mixed input types is resolved for this workload. Confirm instead: Both supported source versions produce the agreed typed feature, and invalid values enter an auditable reject path. Observed problem: New event producers change a numerical source field to strings while legacy objects remain numeric, causing the shared Glue transformation to fail. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
Question 737
At the model-governance gate before promotion of a new ML artifact, the following condition is observed: A fraud-screening model has high global accuracy because most purchases are legitimate, while expensive fraud cases are often missed. The team needs a defensible technical decision rather than an untracked workaround. Which measured outcome would confirm that the correction worked as intended?
- The production response remains on the primary while shadow metrics reveal candidate performance and resource needs.
- Fraud recall meets the agreed target with documented review-volume and false-positive tradeoffs.
- Multiple workers reduce epoch time with acceptable convergence and measured communication overhead.
- The managed training job starts the intended script and produces a usable model artifact in the output location.
Explanation
Correct Answer
B:Fraud recall meets the agreed target with documented review-volume and false-positive tradeoffs. This is the acceptance evidence for the described precision versus recall fraud threshold scenario. Accuracy can conceal poor minority-class detection; threshold selection trades recall against precision. The corrective implementation it validates is: Select a decision threshold using recall or cost-aware precision/recall targets and measure on the untouched holdout. Reported signal: A fraud-screening model has high global accuracy because most purchases are legitimate, while expensive fraud cases are often missed. Operational setting: At the model-governance gate before promotion of a new ML artifact.
Incorrect Answers
A:The production response remains on the primary while shadow metrics reveal candidate performance and resource needs. This could validate shadow deployment model evaluation without response change, but it does not establish that precision versus recall fraud threshold is resolved for this workload. Confirm instead: Fraud recall meets the agreed target with documented review-volume and false-positive tradeoffs. Observed problem: A fraud-screening model has high global accuracy because most purchases are legitimate, while expensive fraud cases are often missed. Operational setting: At the model-governance gate before promotion of a new ML artifact.
C:Multiple workers reduce epoch time with acceptable convergence and measured communication overhead. This could validate data parallelism versus model parallelism, but it does not establish that precision versus recall fraud threshold is resolved for this workload. Confirm instead: Fraud recall meets the agreed target with documented review-volume and false-positive tradeoffs. Observed problem: A fraud-screening model has high global accuracy because most purchases are legitimate, while expensive fraud cases are often missed. Operational setting: At the model-governance gate before promotion of a new ML artifact.
D:The managed training job starts the intended script and produces a usable model artifact in the output location. This could validate training image compatibility for byoc, but it does not establish that precision versus recall fraud threshold is resolved for this workload. Confirm instead: Fraud recall meets the agreed target with documented review-volume and false-positive tradeoffs. Observed problem: A fraud-screening model has high global accuracy because most purchases are legitimate, while expensive fraud cases are often missed. Operational setting: At the model-governance gate before promotion of a new ML artifact.
Question 738
During a controlled replay of the data-preparation workflow in a separate ML staging account, the following condition is observed: An S3 event-triggered feature ingestion function processes the same uploaded object twice, creating duplicate labeled samples. The dataset contract must remain reproducible for a backfill. Which verification result BEST demonstrates that the issue is resolved?
- Replaying the same notification leaves exactly one corresponding logical training record and records deduplication.
- Annotation workers can label required content without viewing direct identifiers and authorized audits can trace approved mappings.
- Two ordinary runs ingest only newly eligible objects while a separately invoked backfill handles prior intervals.
- Training converges more consistently, and the same raw request gives aligned standardized values online.
Explanation
Correct Answer
A:Replaying the same notification leaves exactly one corresponding logical training record and records deduplication. This is the acceptance evidence for the described s3 object event at-least-once processing scenario. Event notifications can be delivered more than once; consuming an event is not proof of exactly-once effects. The corrective implementation it validates is: Make the ingestion handler idempotent using a stable object/version key and conditional ledger writes. Reported signal: An S3 event-triggered feature ingestion function processes the same uploaded object twice, creating duplicate labeled samples. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
Incorrect Answers
B:Annotation workers can label required content without viewing direct identifiers and authorized audits can trace approved mappings. This could validate pii masking before model labeling, but it does not establish that s3 object event at-least-once processing is resolved for this workload. Confirm instead: Replaying the same notification leaves exactly one corresponding logical training record and records deduplication. Observed problem: An S3 event-triggered feature ingestion function processes the same uploaded object twice, creating duplicate labeled samples. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
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 s3 object event at-least-once processing is resolved for this workload. Confirm instead: Replaying the same notification leaves exactly one corresponding logical training record and records deduplication. Observed problem: An S3 event-triggered feature ingestion function processes the same uploaded object twice, creating duplicate labeled samples. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
D:Training converges more consistently, and the same raw request gives aligned standardized values online. This could validate standardization of numerical features, but it does not establish that s3 object event at-least-once processing is resolved for this workload. Confirm instead: Replaying the same notification leaves exactly one corresponding logical training record and records deduplication. Observed problem: An S3 event-triggered feature ingestion function processes the same uploaded object twice, creating duplicate labeled samples. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
Question 739
While reproducing the alert in an isolated AWS ML integration account, the following condition is observed: A recommendation endpoint drops peak requests with HTTP 5xx as invocations per instance rise, while the dashboard averages latency across an entire day. The proposed change must preserve logging and least privilege. Which investigation provides the most decisive FIRST evidence for the underlying problem?
- Inspect inference timestamps, eventual ground-truth arrival, join keys, and the provenance of displayed quality metrics.
- Inspect routing hashes, experiment assignments, variant weights, and request identity consistency.
- Inspect endpoint 5XXErrors, InvocationsPerInstance, instance availability, and per-minute latency distribution.
- Measure memory, throughput, p95 latency and cost across supported candidate endpoint configurations.
Explanation
Correct Answer
C:Inspect endpoint 5XXErrors, InvocationsPerInstance, instance availability, and per-minute latency distribution. This directly tests cloudwatch 5xxerrors endpoint saturation in the reported AWS ML workflow. Daily averages can conceal short capacity failures; monitoring must use an incident-relevant interval. It distinguishes the problem before choosing a change. The related targeted response is: Configure actionable alarms, capacity, and supported scaling or backpressure based on the burst window. Reported signal: A recommendation endpoint drops peak requests with HTTP 5xx as invocations per instance rise, while the dashboard averages latency across an entire day. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
Incorrect Answers
A:Inspect inference timestamps, eventual ground-truth arrival, join keys, and the provenance of displayed quality metrics. This is a possible diagnostic for target-label delay in production quality monitoring, but it does not directly test the cloudwatch 5xxerrors endpoint saturation failure reported here. Prioritize: Inspect endpoint 5XXErrors, InvocationsPerInstance, instance availability, and per-minute latency distribution. The relevant case is: A recommendation endpoint drops peak requests with HTTP 5xx as invocations per instance rise, while the dashboard averages latency across an entire day. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
B:Inspect routing hashes, experiment assignments, variant weights, and request identity consistency. This is a possible diagnostic for champion challenger a/b assignment consistency, but it does not directly test the cloudwatch 5xxerrors endpoint saturation failure reported here. Prioritize: Inspect endpoint 5XXErrors, InvocationsPerInstance, instance availability, and per-minute latency distribution. The relevant case is: A recommendation endpoint drops peak requests with HTTP 5xx as invocations per instance rise, while the dashboard averages latency across an entire day. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
D:Measure memory, throughput, p95 latency and cost across supported candidate endpoint configurations. This is a possible diagnostic for inference recommender benchmarking instance family, but it does not directly test the cloudwatch 5xxerrors endpoint saturation failure reported here. Prioritize: Inspect endpoint 5XXErrors, InvocationsPerInstance, instance availability, and per-minute latency distribution. The relevant case is: A recommendation endpoint drops peak requests with HTTP 5xx as invocations per instance rise, while the dashboard averages latency across an entire day. Operational setting: While reproducing the alert in an isolated AWS ML integration account.
Question 740
At a production AWS ML data lake supporting customer-facing predictions, the following condition is observed: A source feed silently begins emitting missing customer IDs, and the downstream pipeline trains on rows with null entity keys instead of failing. Engineers must avoid changing unrelated source partitions. Which verification result BEST demonstrates that the issue is resolved?
- Invalid batches are rejected with attributable rule failures and accepted training data meets the entity-key contract.
- Identical inputs create identical token IDs in train and serving paths, including punctuation and unknown words.
- Two ordinary runs ingest only newly eligible objects while a separately invoked backfill handles prior intervals.
- The report quantifies class or group imbalance with reproducible group counts and documents remediation assumptions.
Explanation
Correct Answer
A:Invalid batches are rejected with attributable rule failures and accepted training data meets the entity-key contract. This is the acceptance evidence for the described glue data quality non-null contract scenario. Schema existence alone cannot establish semantic completeness or identifier uniqueness in ML inputs. The corrective implementation it validates is: Enforce a required non-null or uniqueness contract with thresholds; quarantine violating partitions before feature publication. Reported signal: A source feed silently begins emitting missing customer IDs, and the downstream pipeline trains on rows with null entity keys instead of failing. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Incorrect Answers
B:Identical inputs create identical token IDs in train and serving paths, including punctuation and unknown words. This could validate text tokenization mismatch in serving, but it does not establish that glue data quality non-null contract is resolved for this workload. Confirm instead: Invalid batches are rejected with attributable rule failures and accepted training data meets the entity-key contract. Observed problem: A source feed silently begins emitting missing customer IDs, and the downstream pipeline trains on rows with null entity keys instead of failing. 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 glue data quality non-null contract is resolved for this workload. Confirm instead: Invalid batches are rejected with attributable rule failures and accepted training data meets the entity-key contract. Observed problem: A source feed silently begins emitting missing customer IDs, and the downstream pipeline trains on rows with null entity keys instead of failing. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
D:The report quantifies class or group imbalance with reproducible group counts and documents remediation assumptions. This could validate clarify label imbalance measurement, but it does not establish that glue data quality non-null contract is resolved for this workload. Confirm instead: Invalid batches are rejected with attributable rule failures and accepted training data meets the entity-key contract. Observed problem: A source feed silently begins emitting missing customer IDs, and the downstream pipeline trains on rows with null entity keys instead of failing. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
