Microsoft AI-300 Practice Test 39
Question 380
During a production Azure Machine Learning model lifecycle incident, the following condition occurs: A model is rolled back to old weights but its scoring container still uses the new incompatible environment, so production errors continue after rollback. The operator must safeguard training reproducibility and serving continuity. Select TWO: the strongest initial evidence and the aligned controlled correction.
- Compare candidate and previous deployment model version, environment digest, scoring code, endpoint configuration and historical release manifest.
- Deploy the approved model through a suitable managed online endpoint and provision capacity for synchronous requests.
- Capture a governed immutable dataset or versioned data asset and reference that exact version in notebook and scheduled jobs.
- Inspect online endpoint traffic weights, deployment versions, request tracing, rollout approval and available prior stable deployment.
- Restore the complete previously approved deployment package, including model, environment and scoring code, rather than only changing weights.
Explanation
Correct Answers
A:Compare candidate and previous deployment model version, environment digest, scoring code, endpoint configuration and historical release manifest. This is the most relevant evidence for rollback cannot recover a previous environment version. A deployment's behavior depends on the whole inference package; changing model weights alone cannot undo incompatible runtime changes. The matching controlled remediation is: Restore the complete previously approved deployment package, including model, environment and scoring code, rather than only changing weights.
E:Restore the complete previously approved deployment package, including model, environment and scoring code, rather than only changing weights. This targets rollback cannot recover a previous environment version. A deployment's behavior depends on the whole inference package; changing model weights alone cannot undo incompatible runtime changes. Verify the failure through: Compare candidate and previous deployment model version, environment digest, scoring code, endpoint configuration and historical release manifest. Validate success with: A representative rollback test restores known-good predictions and endpoint health within the recovery target.
Incorrect Answers
B:Deploy the approved model through a suitable managed online endpoint and provision capacity for synchronous requests. This is associated with online scoring requirement deployed as batch endpoint, not the essential evidence-and-fix pair for rollback cannot recover a previous environment version. First examine: Compare candidate and previous deployment model version, environment digest, scoring code, endpoint configuration and historical release manifest. Then correct: Restore the complete previously approved deployment package, including model, environment and scoring code, rather than only changing weights.
C:Capture a governed immutable dataset or versioned data asset and reference that exact version in notebook and scheduled jobs. This is associated with exploration notebook reads production data without reproducible snapshot, not the essential evidence-and-fix pair for rollback cannot recover a previous environment version. First examine: Compare candidate and previous deployment model version, environment digest, scoring code, endpoint configuration and historical release manifest. Then correct: Restore the complete previously approved deployment package, including model, environment and scoring code, rather than only changing weights.
D:Inspect online endpoint traffic weights, deployment versions, request tracing, rollout approval and available prior stable deployment. This is associated with canary rollout sends all traffic to unvalidated candidate, not the essential evidence-and-fix pair for rollback cannot recover a previous environment version. First examine: Compare candidate and previous deployment model version, environment digest, scoring code, endpoint configuration and historical release manifest. Then correct: Restore the complete previously approved deployment package, including model, environment and scoring code, rather than only changing weights.
Question 381
During a Microsoft Foundry generative AI deployment review, the following condition occurs: An agent receives a newly released model under full production traffic before anyone verifies tool calling and safety regressions. The deployment must retain governed identities, endpoint contracts and auditable release history. Which acceptance test shows the AI operations defect is corrected?
- A production-sized document completes evaluation without unintended context loss or unsupported token-limit errors.
- A test prompt or model deployment change affects only the development environment until an approved promotion occurs.
- Failed canary metrics restore 100% of traffic to the last approved deployment without changing unrelated agent code.
- The managed identity invokes the approved endpoint and a revoked developer key is no longer needed for production inference.
Explanation
Correct Answer
C:Failed canary metrics restore 100% of traffic to the last approved deployment without changing unrelated agent code. Gradual rollout with a known-good model mitigates production blast radius when foundation model behavior changes. a direct corrective action is to route a controlled percentage of requests to the candidate and promote only after automated quality, safety and latency gates pass. This matters here because an agent receives a newly released model under full production traffic before anyone verifies tool calling and….
Incorrect Answers
A:A production-sized document completes evaluation without unintended context loss or unsupported token-limit errors. That addresses a different configuration issue rather than the main requirement. Verify instead: Failed canary metrics restore 100% of traffic to the last approved deployment without changing unrelated agent code.
B:A test prompt or model deployment change affects only the development environment until an approved promotion occurs. That addresses a different configuration issue rather than the main requirement. Verify instead: Failed canary metrics restore 100% of traffic to the last approved deployment without changing unrelated agent code.
D:The managed identity invokes the approved endpoint and a revoked developer key is no longer needed for production inference. That addresses a different configuration issue rather than the main requirement. Verify instead: Failed canary metrics restore 100% of traffic to the last approved deployment without changing unrelated agent code.
Question 382
During a retrieval or model-customization performance review, the following condition occurs: A team tunes its retrieval parameters against the same 300 hand-labeled questions later used as the formal deployment acceptance test, overstating generalization. The change must keep citations, privacy and previously approved behavior intact. Which diagnostic evidence most directly distinguishes the described AI operations failure?
- Compare query and index embedding model versions, vector dimensions, index schema, retrieval score distributions and migration logs.
- Inspect top-k search scores, threshold, query normalization, relevant passages and test question recall.
- Examine fine-tuning run outputs, model registry versions, promotion approvals, held-out evaluations and rollout routing settings.
- Examine question-set provenance, tuning history, held-out set separation, document version and evaluator configuration.
Explanation
Correct Answer
D:Examine question-set provenance, tuning history, held-out set separation, document version and evaluator configuration. This is the most relevant evidence for offline rag evaluator leaks training questions into test set. Repeated optimization against the final test set leaks evaluation information into model selection, making reported quality overly optimistic. The matching controlled remediation is: Split labeled queries into development and untouched final evaluation sets, with repeatable leakage-resistant A/B comparisons.
Incorrect Answers
A:Compare query and index embedding model versions, vector dimensions, index schema, retrieval score distributions and migration logs. This would investigate embedding model replaced without reindexing stored vectors rather than the documented offline rag evaluator leaks training questions into test set issue. The relevant investigation is: Examine question-set provenance, tuning history, held-out set separation, document version and evaluator configuration.
B:Inspect top-k search scores, threshold, query normalization, relevant passages and test question recall. This would investigate rag similarity threshold filters all relevant evidence rather than the documented offline rag evaluator leaks training questions into test set issue. The relevant investigation is: Examine question-set provenance, tuning history, held-out set separation, document version and evaluator configuration.
C:Examine fine-tuning run outputs, model registry versions, promotion approvals, held-out evaluations and rollout routing settings. This would investigate fine-tuned candidate bypasses promotion review after training rather than the documented offline rag evaluator leaks training questions into test set issue. The relevant investigation is: Examine question-set provenance, tuning history, held-out set separation, document version and evaluator configuration.
Question 383
While preparing a secure Foundry model or agent release, the following condition occurs: An engineer commits a Foundry service API key directly in a Bicep parameter file, making the key readable in repository history despite later deletion. The remediation must meet enterprise access and reliability requirements without a broad bypass. Select TWO: the strongest initial evidence and the aligned controlled correction.
- Inspect IaC parameter history, deployment logs, credential usage, Key Vault references and access audit events.
- Revoke the exposed key, use managed identity or approved secure Key Vault secret references, and scrub sensitive release configuration from ongoing builds.
- Inspect Foundry project and resource regions, model deployment regional availability, enterprise residency constraints and API error codes.
- Design and test balanced representative prompt examples under fairness and privacy constraints, retaining a versioned evaluation baseline.
- Permit only the required documented tool and service destinations via supported private or managed-network egress controls.
Explanation
Correct Answers
A:Inspect IaC parameter history, deployment logs, credential usage, Key Vault references and access audit events. This is the most relevant evidence for foundry infrastructure parameters embed a secret in source repository. Declarative infrastructure should manage references and identities rather than persist credentials as version-controlled cleartext values. The matching controlled remediation is: Revoke the exposed key, use managed identity or approved secure Key Vault secret references, and scrub sensitive release configuration from ongoing builds.
B:Revoke the exposed key, use managed identity or approved secure Key Vault secret references, and scrub sensitive release configuration from ongoing builds. This targets foundry infrastructure parameters embed a secret in source repository. Declarative infrastructure should manage references and identities rather than persist credentials as version-controlled cleartext values. Verify the failure through: Inspect IaC parameter history, deployment logs, credential usage, Key Vault references and access audit events. Validate success with: The reviewed deployment succeeds without reusable secrets in committed files and old credentials no longer authorize calls.
Incorrect Answers
C:Inspect Foundry project and resource regions, model deployment regional availability, enterprise residency constraints and API error codes. This is associated with foundry project created in mismatched resource location, not the essential evidence-and-fix pair for foundry infrastructure parameters embed a secret in source repository. First examine: Inspect IaC parameter history, deployment logs, credential usage, Key Vault references and access audit events. Then correct: Revoke the exposed key, use managed identity or approved secure Key Vault secret references, and scrub sensitive release configuration from ongoing builds.
D:Design and test balanced representative prompt examples under fairness and privacy constraints, retaining a versioned evaluation baseline. This is associated with few-shot prompt examples bias outputs toward one subgroup, not the essential evidence-and-fix pair for foundry infrastructure parameters embed a secret in source repository. First examine: Inspect IaC parameter history, deployment logs, credential usage, Key Vault references and access audit events. Then correct: Revoke the exposed key, use managed identity or approved secure Key Vault secret references, and scrub sensitive release configuration from ongoing builds.
E:Permit only the required documented tool and service destinations via supported private or managed-network egress controls. This is associated with agent tool calls fail behind overbroad outbound block, not the essential evidence-and-fix pair for foundry infrastructure parameters embed a secret in source repository. First examine: Inspect IaC parameter history, deployment logs, credential usage, Key Vault references and access audit events. Then correct: Revoke the exposed key, use managed identity or approved secure Key Vault secret references, and scrub sensitive release configuration from ongoing builds.
Question 384
During a Microsoft Foundry generative AI deployment review, the following condition occurs: A Foundry agent invokes its LLM successfully but receives access denied when querying an Azure AI Search index because the assigned managed identity lacks search data reader rights. The deployment must retain governed identities, endpoint contracts and auditable release history. What measured outcome validates the solution while preserving the stated constraints?
- Repeatable IaC releases create the expected project and model deployment before starting inference smoke tests.
- The agent completes retrieval and tool operations without unrestricted public network access or accidental calls to unknown domains.
- Representative order queries produce parseable permitted fields and reject or safely handle malformed outputs.
- Authorized agent queries retrieve permitted indexed chunks while direct access to unrelated indexes remains denied.
Explanation
Correct Answer
D:Authorized agent queries retrieve permitted indexed chunks while direct access to unrelated indexes remains denied. Foundry model access and Azure AI Search data-plane authorization are separate permission boundaries for a retrieval-augmented agent. a direct corrective action is to grant the narrowly scoped search data access role to the correct agent identity and use identity-based connection configuration.
Incorrect Answers
A:Repeatable IaC releases create the expected project and model deployment before starting inference smoke tests. That addresses a different configuration issue rather than the main requirement. Verify instead: Authorized agent queries retrieve permitted indexed chunks while direct access to unrelated indexes remains denied.
B:The agent completes retrieval and tool operations without unrestricted public network access or accidental calls to unknown domains. That addresses a different configuration issue rather than the main requirement. Verify instead: Authorized agent queries retrieve permitted indexed chunks while direct access to unrelated indexes remains denied.
C:Representative order queries produce parseable permitted fields and reject or safely handle malformed outputs. That addresses a different configuration issue rather than the main requirement. Verify instead: Authorized agent queries retrieve permitted indexed chunks while direct access to unrelated indexes remains denied.
Question 385
During a Foundry generative AI quality and observability incident, the following condition occurs: A user-facing agent becomes slow after an LLM deployment change, but the team only records total response time and cannot tell retrieval delay from model generation delay. The fix must preserve evaluation validity and avoid unnecessary user-data collection. Which investigation best verifies the cause before changing production configuration?
- Examine end-to-end request traces, retrieval timing, model time to first token, generation duration, request concurrency and percentiles.
- Inspect evaluation dataset field names, input-output column mapping, evaluator requirements, row samples and failed scoring logs.
- Examine continuous monitoring configuration, trace deployment IDs, sampling window, evaluator assignment and per-version quality results.
- Inspect tokens per request, input/output token ratios, concurrency, throttling status and model deployment quota.
Explanation
Correct Answer
A:Examine end-to-end request traces, retrieval timing, model time to first token, generation duration, request concurrency and percentiles. This is the most relevant evidence for generation latency spike blamed on retrieval without breakdown. Latency budgets should separate agent orchestration, retrieval and generation; total duration alone cannot locate the bottleneck. The matching controlled remediation is: Instrument stage-level latency and compare generation and retrieval spans before applying a targeted capacity or retrieval change. This matters here because a user-facing agent becomes slow after an LLM deployment change, but the team only records total response time and….
Incorrect Answers
B:Inspect evaluation dataset field names, input-output column mapping, evaluator requirements, row samples and failed scoring logs. This would investigate genai evaluation dataset lacks expected answer column mapping rather than the documented generation latency spike blamed on retrieval without breakdown issue. The relevant investigation is: Examine end-to-end request traces, retrieval timing, model time to first token, generation duration, request concurrency and percentiles.
C:Examine continuous monitoring configuration, trace deployment IDs, sampling window, evaluator assignment and per-version quality results. This would investigate foundry continuous evaluation ignores production deployment version rather than the documented generation latency spike blamed on retrieval without breakdown issue. The relevant investigation is: Examine end-to-end request traces, retrieval timing, model time to first token, generation duration, request concurrency and percentiles.
D:Inspect tokens per request, input/output token ratios, concurrency, throttling status and model deployment quota. This would investigate genai throughput report mixes responses and tokens rather than the documented generation latency spike blamed on retrieval without breakdown issue. The relevant investigation is: Examine end-to-end request traces, retrieval timing, model time to first token, generation duration, request concurrency and percentiles.
Question 386
During a production Azure Machine Learning model lifecycle incident, the following condition occurs: A model is trained across four GPU workers but all workers ingest the same unsharded dataset in each epoch, wasting resources and biasing effective training statistics. The operator must safeguard training reproducibility and serving continuity. Which investigation best verifies the cause before changing production configuration?
- Inspect MLflow tracking URI and experiment name, run parameter history, metric logs, artifact URIs and the input data version of each job.
- Review distributed sampler configuration, global and local ranks, dataset partition checks, epoch seeding and examples processed per worker.
- Inspect request payload schema, scoring script init/run implementation, model input signature, container logs and test invocation.
- Check mlflow.autolog arguments, run artifact listing, model flavor metadata, configured output path and registration job logs.
Explanation
Correct Answer
B:Review distributed sampler configuration, global and local ranks, dataset partition checks, epoch seeding and examples processed per worker. This is the most relevant evidence for distributed training uses one shard for every worker. Data-parallel training must coordinate sampling and gradient aggregation; launching multiple processes does not automatically shard training data. The matching controlled remediation is: Configure the framework's appropriate distributed data sampler and synchronize epoch handling and metrics aggregation.
Incorrect Answers
A:Inspect MLflow tracking URI and experiment name, run parameter history, metric logs, artifact URIs and the input data version of each job. This would investigate mlflow experiment logs omit run parameters rather than the documented distributed training uses one shard for every worker issue. The relevant investigation is: Review distributed sampler configuration, global and local ranks, dataset partition checks, epoch seeding and examples processed per worker.
C:Inspect request payload schema, scoring script init/run implementation, model input signature, container logs and test invocation. This would investigate endpoint scoring script expects wrong json schema rather than the documented distributed training uses one shard for every worker issue. The relevant investigation is: Review distributed sampler configuration, global and local ranks, dataset partition checks, epoch seeding and examples processed per worker.
D:Check mlflow.autolog arguments, run artifact listing, model flavor metadata, configured output path and registration job logs. This would investigate autologging excludes required model artifact rather than the documented distributed training uses one shard for every worker issue. The relevant investigation is: Review distributed sampler configuration, global and local ranks, dataset partition checks, epoch seeding and examples processed per worker.
Question 387
As an MLOps engineer reviews a model training or inference pipeline, the following condition occurs: A real-time predictor appears healthy by mean request latency while high-value users face multi-second waits at the 99th percentile after a traffic spike. The resolution must preserve quality gates, documented artifacts and rollback options. Which production-safe remediation should the engineer apply to this AI workflow?
- Configure tail-latency and error-rate dashboards with capacity remedies appropriate to the measured bottleneck.
- Configure a suitable seasonal baseline or context-aware drift monitor while retaining meaningful anomaly thresholds.
- Configure a relevant supported primary metric and validation scheme for rare positive events and compare operating thresholds after training.
- Monitor available proxy signals immediately and compute labeled performance only after verified outcomes arrive through a governed join.
Explanation
Correct Answer
A:Configure tail-latency and error-rate dashboards with capacity remedies appropriate to the measured bottleneck. This targets endpoint latency average conceals tail degradation. Average endpoint latency can hide slow outliers; tail distributions and cohort behavior are needed for user-facing service assurance. Verify the failure through: Examine Azure Monitor request latency percentiles, response codes, deployment load, compute utilization and per-client cohorts. Validate success with: P95/P99 and error-rate service objectives recover under representative peak-load traffic without unrelated scale changes. This matters here because as an MLOps engineer reviews a model training or inference pipeline, the following condition occurs: A real-time….
Incorrect Answers
B:Configure a suitable seasonal baseline or context-aware drift monitor while retaining meaningful anomaly thresholds. This could address data drift monitor uses wrong reference season, but it does not solve the specified endpoint latency average conceals tail degradation failure. The matching remediation is: Configure tail-latency and error-rate dashboards with capacity remedies appropriate to the measured bottleneck.
C:Configure a relevant supported primary metric and validation scheme for rare positive events and compare operating thresholds after training. This could address automl primary metric does not match rare-event goal, but it does not solve the specified endpoint latency average conceals tail degradation failure. The matching remediation is: Configure tail-latency and error-rate dashboards with capacity remedies appropriate to the measured bottleneck.
D:Monitor available proxy signals immediately and compute labeled performance only after verified outcomes arrive through a governed join. This could address accuracy monitoring lacks delayed outcome labels, but it does not solve the specified endpoint latency average conceals tail degradation failure. The matching remediation is: Configure tail-latency and error-rate dashboards with capacity remedies appropriate to the measured bottleneck.
Question 388
During a production Azure Machine Learning model lifecycle incident, the following condition occurs: An engineer promotes the model with the lowest training loss even though the held-out validation loss worsens, indicating overfitting. The operator must safeguard training reproducibility and serving continuity. What measured outcome validates the solution while preserving the stated constraints?
- An injected null-rate increase triggers a targeted alert while valid fluctuations remain within control limits.
- The selected version meets validation criteria without depending on apparent fit to the training examples alone.
- Repeat runs show explainable variance with recorded seeds and identical inputs where deterministic execution is supported.
- The finished job produces a discoverable artifact that can be loaded and registered by the next pipeline stage.
Explanation
Correct Answer
B:The selected version meets validation criteria without depending on apparent fit to the training examples alone. A training objective measures fit on observed examples; generalization claims require independent validation and selection rules. a direct corrective action is to use a documented holdout or cross-validation-based model selection procedure and require a production-relevant promotion threshold.
Incorrect Answers
A:An injected null-rate increase triggers a targeted alert while valid fluctuations remain within control limits. That addresses a different configuration issue rather than the main requirement. Verify instead: The selected version meets validation criteria without depending on apparent fit to the training examples alone.
C:Repeat runs show explainable variance with recorded seeds and identical inputs where deterministic execution is supported. That addresses a different configuration issue rather than the main requirement. Verify instead: The selected version meets validation criteria without depending on apparent fit to the training examples alone.
D:The finished job produces a discoverable artifact that can be loaded and registered by the next pipeline stage. That addresses a different configuration issue rather than the main requirement. Verify instead: The selected version meets validation criteria without depending on apparent fit to the training examples alone.
Question 389
During a Microsoft Foundry generative AI deployment review, the following condition occurs: A planned model upgrade cannot allocate the required extra provisioned throughput units in its compliance region, but a release pipeline attempts the new deployment before checking quota. The deployment must retain governed identities, endpoint contracts and auditable release history. Which observable postchange result proves the intended remediation succeeded?
- Test queries reach the intended deployed model with attributable project usage and without cross-project credential exposure.
- The model version upgrade completes without an outage or unauthorized cross-region processing.
- Replaying approved test prompts yields compatible output contracts until the next reviewed version is promoted.
- Peak requests meet approved latency and throttling targets under a representative load test while capacity costs remain documented.
Explanation
Correct Answer
B:The model version upgrade completes without an outage or unauthorized cross-region processing. Model capacity is region and deployment-type dependent; version upgrades cannot assume unused PTUs are available on demand. a direct corrective action is to preflight capacity and schedule an approved upgrade with quota adjustment or temporary side-by-side resources that fit the region.
Incorrect Answers
A:Test queries reach the intended deployed model with attributable project usage and without cross-project credential exposure. That addresses a different configuration issue rather than the main requirement. Verify instead: The model version upgrade completes without an outage or unauthorized cross-region processing.
C:Replaying approved test prompts yields compatible output contracts until the next reviewed version is promoted. That addresses a different configuration issue rather than the main requirement. Verify instead: The model version upgrade completes without an outage or unauthorized cross-region processing.
D:Peak requests meet approved latency and throttling targets under a representative load test while capacity costs remain documented. That addresses a different configuration issue rather than the main requirement. Verify instead: The model version upgrade completes without an outage or unauthorized cross-region processing.
Question 390
During a production Azure Machine Learning model lifecycle incident, the following condition occurs: A real-time predictor appears healthy by mean request latency while high-value users face multi-second waits at the 99th percentile after a traffic spike. The operator must safeguard training reproducibility and serving continuity. Which diagnostic evidence most directly distinguishes the described AI operations failure?
- Inspect request payload schema, scoring script init/run implementation, model input signature, container logs and test invocation.
- Inspect reference data time window, drift feature metrics, alert threshold, seasonal segmentation and actual model outcomes.
- Examine Azure Monitor request latency percentiles, response codes, deployment load, compute utilization and per-client cohorts.
- Inspect model retraining pipeline graph, evaluation outputs, baseline comparison, deployment step conditions and release records.
Explanation
Correct Answer
C:Examine Azure Monitor request latency percentiles, response codes, deployment load, compute utilization and per-client cohorts. This is the most relevant evidence for endpoint latency average conceals tail degradation. Average endpoint latency can hide slow outliers; tail distributions and cohort behavior are needed for user-facing service assurance. The matching controlled remediation is: Configure tail-latency and error-rate dashboards with capacity remedies appropriate to the measured bottleneck.
Incorrect Answers
A:Inspect request payload schema, scoring script init/run implementation, model input signature, container logs and test invocation. This would investigate endpoint scoring script expects wrong json schema rather than the documented endpoint latency average conceals tail degradation issue. The relevant investigation is: Examine Azure Monitor request latency percentiles, response codes, deployment load, compute utilization and per-client cohorts.
B:Inspect reference data time window, drift feature metrics, alert threshold, seasonal segmentation and actual model outcomes. This would investigate data drift monitor uses wrong reference season rather than the documented endpoint latency average conceals tail degradation issue. The relevant investigation is: Examine Azure Monitor request latency percentiles, response codes, deployment load, compute utilization and per-client cohorts.
D:Inspect model retraining pipeline graph, evaluation outputs, baseline comparison, deployment step conditions and release records. This would investigate retraining pipeline deploys without model quality gate rather than the documented endpoint latency average conceals tail degradation issue. The relevant investigation is: Examine Azure Monitor request latency percentiles, response codes, deployment load, compute utilization and per-client cohorts.