McAfee-Secured Website

Exam Code: AI-300

Exam Name: Operationalizing Machine Learning and Generative AI Solutions

Certification Provider: Microsoft

Corresponding Certification: Microsoft Certified: Machine Learning Operations Engineer Associate

100% Microsoft AI-300 Guaranteed Success With Testking Training Materials

Get Real AI-300 Exam Questions & Expert Verified Answers as Experienced on the Real Exam!

213 Questions & Answers with Testing Engine

"Operationalizing Machine Learning and Generative AI Solutions Exam", also known as AI-300 exam, is a Microsoft certification exam.

Pass your tests with the always up-to-date AI-300 Exam Engine. Your AI-300 training materials keep you at the head of the pack!

Satisfaction Guaranteed

Money Back Guarantee

Testking provides hassle-free money back guarantee with our products. That is because we have 100% trust in the abilities of our professional and experience product team, and our record is a proof of that.

99.6% PASS RATE
Was: $137.49
Now: $124.99

Microsoft AI-300 Practice Test 34

Question 330

A scoring service works but its application logs do not show failures for malformed inputs. What observability improvement is relevant?

  1. Run representative end-to-end semantic inference tests against known cases and expected outputs.
  2. Validate the laptop's approved network route and DNS access to the private endpoint.
  3. Capture structured request errors and scoring diagnostics in the supported endpoint logs.
  4. Use the exact active scoring URI issued for the newly created endpoint.

Explanation

Correct Answer

C:

Serving logs identify request validation failures and model exceptions that aggregate uptime metrics may hide.

Incorrect Answers

A:

Run representative end-to-end semantic inference tests against known cases and expected outputs. That operation would be relevant if: A deployed model's health probe succeeds but its predictions are nonsensical after a feature schema change. Its corresponding reason or verification is: Healthy infrastructure cannot guarantee correct feature interpretation or model behavior; application-level checks remain necessary. However the problem in this question is: A scoring service works but its application logs do not show failures for malformed inputs. The requirement instead calls for Capture structured request errors and scoring diagnostics in the supported endpoint logs; the question-specific check is: Serving logs identify request validation failures and model exceptions that aggregate uptime metrics may hide.

B:

Validate the laptop's approved network route and DNS access to the private endpoint. A separate source case supports that approach when: An Azure ML online endpoint works from a VNet-hosted test runner but fails from a laptop after private access was enabled. The supporting rationale or observable result is: Private network enforcement can intentionally block clients outside the authorized path despite healthy model serving. This question describes a different issue: A scoring service works but its application logs do not show failures for malformed inputs. Address it through Capture structured request errors and scoring diagnostics in the supported endpoint logs, with the required justification or evidence: Serving logs identify request validation failures and model exceptions that aggregate uptime metrics may hide.

D:

Use the exact active scoring URI issued for the newly created endpoint. This is useful for the distinct issue: A newly provisioned online endpoint returns 404 because the client calls an old endpoint hostname. The relevant technical rationale or evidence for that issue is: Endpoint names and addresses are specific to deployed resources; using a stale hostname cannot reach the intended service. Here the reported issue is: A scoring service works but its application logs do not show failures for malformed inputs. The relevant response is: Capture structured request errors and scoring diagnostics in the supported endpoint logs. The first result cannot verify this different correction: Serving logs identify request validation failures and model exceptions that aggregate uptime metrics may hide.

Question 331

A batch scoring job fails on one corrupt input file but the endpoint metadata appears healthy. What should the team inspect?

  1. Check deployed capacity, rate limits, and throttling metrics against traffic volume.
  2. Review the failed batch job logs and input-file error details.
  3. Run representative end-to-end semantic inference tests against known cases and expected outputs.
  4. Capture structured request errors and scoring diagnostics in the supported endpoint logs.

Explanation

Correct Answer

B:

Batch failure diagnosis belongs to the specific job execution, not solely the endpoint's static configuration.

Incorrect Answers

A:

The alternative, Check deployed capacity, rate limits, and throttling metrics against traffic volume, belongs to a different technical condition: Clients report intermittent 429 responses after a sudden increase in inference rate. There the source-specific mechanism or proof is: Throttling signals resource or service limit pressure rather than necessarily a malformed model artifact. The condition asked about here is: A batch scoring job fails on one corrupt input file but the endpoint metadata appears healthy. A suitable correction is Review the failed batch job logs and input-file error details; the targeted justification or result is: Batch failure diagnosis belongs to the specific job execution, not solely the endpoint's static configuration.

C:

Run representative end-to-end semantic inference tests against known cases and expected outputs. That operation would be relevant if: A deployed model's health probe succeeds but its predictions are nonsensical after a feature schema change. Its corresponding reason or verification is: Healthy infrastructure cannot guarantee correct feature interpretation or model behavior; application-level checks remain necessary. However the problem in this question is: A batch scoring job fails on one corrupt input file but the endpoint metadata appears healthy. The requirement instead calls for Review the failed batch job logs and input-file error details; the question-specific check is: Batch failure diagnosis belongs to the specific job execution, not solely the endpoint's static configuration.

D:

Capture structured request errors and scoring diagnostics in the supported endpoint logs. A separate source case supports that approach when: A scoring service works but its application logs do not show failures for malformed inputs. The supporting rationale or observable result is: Serving logs identify request validation failures and model exceptions that aggregate uptime metrics may hide. This question describes a different issue: A batch scoring job fails on one corrupt input file but the endpoint metadata appears healthy. Address it through Review the failed batch job logs and input-file error details, with the required justification or evidence: Batch failure diagnosis belongs to the specific job execution, not solely the endpoint's static configuration.

Question 332

A deployed model's health probe succeeds but its predictions are nonsensical after a feature schema change. What additional validation is needed?

  1. Run representative end-to-end semantic inference tests against known cases and expected outputs.
  2. Capture structured request errors and scoring diagnostics in the supported endpoint logs.
  3. Inspect the endpoint's traffic allocation and routing configuration.
  4. Use the exact active scoring URI issued for the newly created endpoint.

Explanation

Correct Answer

A:

Healthy infrastructure cannot guarantee correct feature interpretation or model behavior; application-level checks remain necessary.

Incorrect Answers

B:

Capture structured request errors and scoring diagnostics in the supported endpoint logs. It addresses another case: A scoring service works but its application logs do not show failures for malformed inputs. For that case, the distinguishing justification or acceptance check is: Serving logs identify request validation failures and model exceptions that aggregate uptime metrics may hide. The present case instead states: A deployed model's health probe succeeds but its predictions are nonsensical after a feature schema change. It calls for Run representative end-to-end semantic inference tests against known cases and expected outputs. Showing success for the other case would not establish: Healthy infrastructure cannot guarantee correct feature interpretation or model behavior; application-level checks remain necessary.

C:

The alternative, Inspect the endpoint's traffic allocation and routing configuration, belongs to a different technical condition: A deployment exists under an endpoint but receives no production calls; another deployment receives all traffic. There the source-specific mechanism or proof is: Endpoint routing percentages determine which active deployment receives client requests, independent of the presence of model artifacts. The condition asked about here is: A deployed model's health probe succeeds but its predictions are nonsensical after a feature schema change. A suitable correction is Run representative end-to-end semantic inference tests against known cases and expected outputs; the targeted justification or result is: Healthy infrastructure cannot guarantee correct feature interpretation or model behavior; application-level checks remain necessary.

D:

Use the exact active scoring URI issued for the newly created endpoint. That operation would be relevant if: A newly provisioned online endpoint returns 404 because the client calls an old endpoint hostname. Its corresponding reason or verification is: Endpoint names and addresses are specific to deployed resources; using a stale hostname cannot reach the intended service. However the problem in this question is: A deployed model's health probe succeeds but its predictions are nonsensical after a feature schema change. The requirement instead calls for Run representative end-to-end semantic inference tests against known cases and expected outputs; the question-specific check is: Healthy infrastructure cannot guarantee correct feature interpretation or model behavior; application-level checks remain necessary.

Question 333

An endpoint deployment repeatedly restarts because its model requires more RAM than the selected instance offers. What evidence identifies the cause?

  1. Use the exact active scoring URI issued for the newly created endpoint.
  2. Review the failed batch job logs and input-file error details.
  3. Compare request payload schema with the model's logged input signature.
  4. Inspect container restart and out-of-memory signals, then select sufficient compute.

Explanation

Correct Answer

D:

Memory exhaustion appears in deployment events and resource metrics, not merely in client authentication settings.

Incorrect Answers

A:

Use the exact active scoring URI issued for the newly created endpoint. This is useful for the distinct issue: A newly provisioned online endpoint returns 404 because the client calls an old endpoint hostname. The relevant technical rationale or evidence for that issue is: Endpoint names and addresses are specific to deployed resources; using a stale hostname cannot reach the intended service. Here the reported issue is: An endpoint deployment repeatedly restarts because its model requires more RAM than the selected instance offers. The relevant response is: Inspect container restart and out-of-memory signals, then select sufficient compute. The first result cannot verify this different correction: Memory exhaustion appears in deployment events and resource metrics, not merely in client authentication settings.

B:

Review the failed batch job logs and input-file error details. It addresses another case: A batch scoring job fails on one corrupt input file but the endpoint metadata appears healthy. For that case, the distinguishing justification or acceptance check is: Batch failure diagnosis belongs to the specific job execution, not solely the endpoint's static configuration. The present case instead states: An endpoint deployment repeatedly restarts because its model requires more RAM than the selected instance offers. It calls for Inspect container restart and out-of-memory signals, then select sufficient compute. Showing success for the other case would not establish: Memory exhaustion appears in deployment events and resource metrics, not merely in client authentication settings.

C:

The alternative, Compare request payload schema with the model's logged input signature, belongs to a different technical condition: An online endpoint returns HTTP 400 after a client changed the casing and types of feature names. There the source-specific mechanism or proof is: Malformed inputs can be rejected even when the deployment and network are healthy. The condition asked about here is: An endpoint deployment repeatedly restarts because its model requires more RAM than the selected instance offers. A suitable correction is Inspect container restart and out-of-memory signals, then select sufficient compute; the targeted justification or result is: Memory exhaustion appears in deployment events and resource metrics, not merely in client authentication settings.

Question 334

A newly provisioned online endpoint returns 404 because the client calls an old endpoint hostname. What must be corrected?

  1. Check deployed capacity, rate limits, and throttling metrics against traffic volume.
  2. Review the failed batch job logs and input-file error details.
  3. Use the exact active scoring URI issued for the newly created endpoint.
  4. Supply a valid token or key matching the endpoint's configured authentication mode.

Explanation

Correct Answer

C:

Endpoint names and addresses are specific to deployed resources; using a stale hostname cannot reach the intended service.

Incorrect Answers

A:

Check deployed capacity, rate limits, and throttling metrics against traffic volume. A separate source case supports that approach when: Clients report intermittent 429 responses after a sudden increase in inference rate. The supporting rationale or observable result is: Throttling signals resource or service limit pressure rather than necessarily a malformed model artifact. This question describes a different issue: A newly provisioned online endpoint returns 404 because the client calls an old endpoint hostname. Address it through Use the exact active scoring URI issued for the newly created endpoint, with the required justification or evidence: Endpoint names and addresses are specific to deployed resources; using a stale hostname cannot reach the intended service.

B:

Review the failed batch job logs and input-file error details. This is useful for the distinct issue: A batch scoring job fails on one corrupt input file but the endpoint metadata appears healthy. The relevant technical rationale or evidence for that issue is: Batch failure diagnosis belongs to the specific job execution, not solely the endpoint's static configuration. Here the reported issue is: A newly provisioned online endpoint returns 404 because the client calls an old endpoint hostname. The relevant response is: Use the exact active scoring URI issued for the newly created endpoint. The first result cannot verify this different correction: Endpoint names and addresses are specific to deployed resources; using a stale hostname cannot reach the intended service.

D:

Supply a valid token or key matching the endpoint's configured authentication mode. It addresses another case: A managed online endpoint returns HTTP 401 from a client that has the correct URL but no authorization header. For that case, the distinguishing justification or acceptance check is: Endpoint authentication is enforced independently of URL reachability; an unsigned request is rejected before scoring. The present case instead states: A newly provisioned online endpoint returns 404 because the client calls an old endpoint hostname. It calls for Use the exact active scoring URI issued for the newly created endpoint. Showing success for the other case would not establish: Endpoint names and addresses are specific to deployed resources; using a stale hostname cannot reach the intended service.

Question 335

A new fraud model must receive only five percent of live requests while the existing model handles the rest. Which rollout mechanism fits?

  1. Retain the blue deployment until green passes validation and traffic ramp checks.
  2. Configure a canary traffic split between the new and existing endpoint deployments.
  3. Route traffic back to the known-good deployment using endpoint traffic weights.
  4. Record the applied traffic allocation and deployment versions in version-controlled release metadata.

Explanation

Correct Answer

B:

Weighted traffic routing limits exposure of an unproven model while collecting representative production behavior.

Incorrect Answers

A:

Retain the blue deployment until green passes validation and traffic ramp checks. That operation would be relevant if: A release pipeline creates a green deployment and immediately removes blue before any tests. Its corresponding reason or verification is: Keeping the previous version available allows swift rollback if new production traffic reveals model defects. However the problem in this question is: A new fraud model must receive only five percent of live requests while the existing model handles the rest. The requirement instead calls for Configure a canary traffic split between the new and existing endpoint deployments; the question-specific check is: Weighted traffic routing limits exposure of an unproven model while collecting representative production behavior.

C:

Route traffic back to the known-good deployment using endpoint traffic weights. A separate source case supports that approach when: A new model's production error rate rises sharply during the initial 10-percent rollout. The supporting rationale or observable result is: Traffic rollback can restore the previous serving model without deleting evidence from the failed candidate deployment. This question describes a different issue: A new fraud model must receive only five percent of live requests while the existing model handles the rest. Address it through Configure a canary traffic split between the new and existing endpoint deployments, with the required justification or evidence: Weighted traffic routing limits exposure of an unproven model while collecting representative production behavior.

D:

Record the applied traffic allocation and deployment versions in version-controlled release metadata. This is useful for the distinct issue: An engineer increases traffic to a candidate model using a local YAML edit that is not committed. The relevant technical rationale or evidence for that issue is: Rollout history should reveal which model served each interval and why a distribution change was approved. Here the reported issue is: A new fraud model must receive only five percent of live requests while the existing model handles the rest. The relevant response is: Configure a canary traffic split between the new and existing endpoint deployments. The first result cannot verify this different correction: Weighted traffic routing limits exposure of an unproven model while collecting representative production behavior.

Question 336

A new model's production error rate rises sharply during the initial 10-percent rollout. What is the immediate safer recovery?

  1. Route traffic back to the known-good deployment using endpoint traffic weights.
  2. Require an approved release gate tied to measured canary health and the exact model version.
  3. Conduct representative load and latency tests before ramping serving traffic.
  4. Gradually increase candidate traffic only after validating dependent service connectivity.

Explanation

Correct Answer

A:

Traffic rollback can restore the previous serving model without deleting evidence from the failed candidate deployment.

Incorrect Answers

B:

The alternative, Require an approved release gate tied to measured canary health and the exact model version, belongs to a different technical condition: A new deployment was validated but receives 100 percent of requests without an approval record. There the source-specific mechanism or proof is: Controlled promotion separates technical availability from authorized activation of a production model revision. The condition asked about here is: A new model's production error rate rises sharply during the initial 10-percent rollout. A suitable correction is Route traffic back to the known-good deployment using endpoint traffic weights; the targeted justification or result is: Traffic rollback can restore the previous serving model without deleting evidence from the failed candidate deployment.

C:

Conduct representative load and latency tests before ramping serving traffic. That operation would be relevant if: A team promotes a new model based only on successful smoke tests at one request per second. Production handles 1,000 requests per second. Its corresponding reason or verification is: Passing low-load checks does not demonstrate sufficient throughput or acceptable tail latency under real demand. However the problem in this question is: A new model's production error rate rises sharply during the initial 10-percent rollout. The requirement instead calls for Route traffic back to the known-good deployment using endpoint traffic weights; the question-specific check is: Traffic rollback can restore the previous serving model without deleting evidence from the failed candidate deployment.

D:

Gradually increase candidate traffic only after validating dependent service connectivity. A separate source case supports that approach when: A blue-green deployment's new version is correct but its feature service connection fails under real traffic. The supporting rationale or observable result is: Traffic ramps expose integration faults before they affect all customers and preserve capacity for rapid rollback. This question describes a different issue: A new model's production error rate rises sharply during the initial 10-percent rollout. Address it through Route traffic back to the known-good deployment using endpoint traffic weights, with the required justification or evidence: Traffic rollback can restore the previous serving model without deleting evidence from the failed candidate deployment.

Question 337

Business reviewers want to compare a new model's predictions against real traffic without using those predictions for customers. Which pattern should be assessed?

  1. Preserve the immutable prior model artifact, environment, and deployment manifest.
  2. Conduct representative load and latency tests before ramping serving traffic.
  3. Gradually increase candidate traffic only after validating dependent service connectivity.
  4. Mirror production requests to the new deployment where supported, without serving its output to clients.

Explanation

Correct Answer

D:

Shadow inference tests behavior on representative requests while customers continue receiving the stable deployment's predictions.

Incorrect Answers

A:

Preserve the immutable prior model artifact, environment, and deployment manifest. It addresses another case: The operator needs a reproducible way to return to an exact prior model if rollback occurs weeks after the current release. For that case, the distinguishing justification or acceptance check is: Rollback must restore known code and dependencies, not rely on an ambiguous latest registry reference. The present case instead states: Business reviewers want to compare a new model's predictions against real traffic without using those predictions for customers. It calls for Mirror production requests to the new deployment where supported, without serving its output to clients. Showing success for the other case would not establish: Shadow inference tests behavior on representative requests while customers continue receiving the stable deployment's predictions.

B:

The alternative, Conduct representative load and latency tests before ramping serving traffic, belongs to a different technical condition: A team promotes a new model based only on successful smoke tests at one request per second. Production handles 1,000 requests per second. There the source-specific mechanism or proof is: Passing low-load checks does not demonstrate sufficient throughput or acceptable tail latency under real demand. The condition asked about here is: Business reviewers want to compare a new model's predictions against real traffic without using those predictions for customers. A suitable correction is Mirror production requests to the new deployment where supported, without serving its output to clients; the targeted justification or result is: Shadow inference tests behavior on representative requests while customers continue receiving the stable deployment's predictions.

C:

Gradually increase candidate traffic only after validating dependent service connectivity. That operation would be relevant if: A blue-green deployment's new version is correct but its feature service connection fails under real traffic. Its corresponding reason or verification is: Traffic ramps expose integration faults before they affect all customers and preserve capacity for rapid rollback. However the problem in this question is: Business reviewers want to compare a new model's predictions against real traffic without using those predictions for customers. The requirement instead calls for Mirror production requests to the new deployment where supported, without serving its output to clients; the question-specific check is: Shadow inference tests behavior on representative requests while customers continue receiving the stable deployment's predictions.

Question 338

A release pipeline creates a green deployment and immediately removes blue before any tests. How should a safe rollout be staged?

  1. Keep the existing deployment until all production traffic has safely moved away from it.
  2. Compare relevant output-quality metrics as well as infrastructure health during rollout.
  3. Retain the blue deployment until green passes validation and traffic ramp checks.
  4. Run contract-compatibility tests against the new deployment before shifting production traffic.

Explanation

Correct Answer

C:

Keeping the previous version available allows swift rollback if new production traffic reveals model defects.

Incorrect Answers

A:

Keep the existing deployment until all production traffic has safely moved away from it. This is useful for the distinct issue: A data scientist deletes the old blue deployment after switching green to 20 percent of traffic, leaving the other 80 percent unroutable. The relevant technical rationale or evidence for that issue is: A split requires both deployments to remain serving; deleting one while it receives traffic can break requests. Here the reported issue is: A release pipeline creates a green deployment and immediately removes blue before any tests. The relevant response is: Retain the blue deployment until green passes validation and traffic ramp checks. The first result cannot verify this different correction: Keeping the previous version available allows swift rollback if new production traffic reveals model defects.

B:

Compare relevant output-quality metrics as well as infrastructure health during rollout. It addresses another case: An automated canary analysis uses only request success rate, but the new model produces incorrect classifications. For that case, the distinguishing justification or acceptance check is: Error-free HTTP responses do not prove model predictions are accurate or safe enough to promote. The present case instead states: A release pipeline creates a green deployment and immediately removes blue before any tests. It calls for Retain the blue deployment until green passes validation and traffic ramp checks. Showing success for the other case would not establish: Keeping the previous version available allows swift rollback if new production traffic reveals model defects.

D:

The alternative, Run contract-compatibility tests against the new deployment before shifting production traffic, belongs to a different technical condition: A model's API contract changed between versions and causes 400 responses for old clients. There the source-specific mechanism or proof is: Safe model rollout includes interface compatibility, not merely readiness of the new container. The condition asked about here is: A release pipeline creates a green deployment and immediately removes blue before any tests. A suitable correction is Retain the blue deployment until green passes validation and traffic ramp checks; the targeted justification or result is: Keeping the previous version available allows swift rollback if new production traffic reveals model defects.

Question 339

A model's API contract changed between versions and causes 400 responses for old clients. What release practice prevents sudden breakage?

  1. Configure a canary traffic split between the new and existing endpoint deployments.
  2. Run contract-compatibility tests against the new deployment before shifting production traffic.
  3. Retain the blue deployment until green passes validation and traffic ramp checks.
  4. Route traffic back to the known-good deployment using endpoint traffic weights.

Explanation

Correct Answer

B:

Safe model rollout includes interface compatibility, not merely readiness of the new container.

Incorrect Answers

A:

Configure a canary traffic split between the new and existing endpoint deployments. A separate source case supports that approach when: A new fraud model must receive only five percent of live requests while the existing model handles the rest. The supporting rationale or observable result is: Weighted traffic routing limits exposure of an unproven model while collecting representative production behavior. This question describes a different issue: A model's API contract changed between versions and causes 400 responses for old clients. Address it through Run contract-compatibility tests against the new deployment before shifting production traffic, with the required justification or evidence: Safe model rollout includes interface compatibility, not merely readiness of the new container.

C:

Retain the blue deployment until green passes validation and traffic ramp checks. This is useful for the distinct issue: A release pipeline creates a green deployment and immediately removes blue before any tests. The relevant technical rationale or evidence for that issue is: Keeping the previous version available allows swift rollback if new production traffic reveals model defects. Here the reported issue is: A model's API contract changed between versions and causes 400 responses for old clients. The relevant response is: Run contract-compatibility tests against the new deployment before shifting production traffic. The first result cannot verify this different correction: Safe model rollout includes interface compatibility, not merely readiness of the new container.

D:

Route traffic back to the known-good deployment using endpoint traffic weights. It addresses another case: A new model's production error rate rises sharply during the initial 10-percent rollout. For that case, the distinguishing justification or acceptance check is: Traffic rollback can restore the previous serving model without deleting evidence from the failed candidate deployment. The present case instead states: A model's API contract changed between versions and causes 400 responses for old clients. It calls for Run contract-compatibility tests against the new deployment before shifting production traffic. Showing success for the other case would not establish: Safe model rollout includes interface compatibility, not merely readiness of the new container.

Question 340

A team promotes a new model based only on successful smoke tests at one request per second. Production handles 1,000 requests per second. What validation is missing?

  1. Conduct representative load and latency tests before ramping serving traffic.
  2. Require an approved release gate tied to measured canary health and the exact model version.
  3. Route traffic back to the known-good deployment using endpoint traffic weights.
  4. Preserve the immutable prior model artifact, environment, and deployment manifest.

Explanation

Correct Answer

A:

Passing low-load checks does not demonstrate sufficient throughput or acceptable tail latency under real demand.

Incorrect Answers

B:

Require an approved release gate tied to measured canary health and the exact model version. A separate source case supports that approach when: A new deployment was validated but receives 100 percent of requests without an approval record. The supporting rationale or observable result is: Controlled promotion separates technical availability from authorized activation of a production model revision. This question describes a different issue: A team promotes a new model based only on successful smoke tests at one request per second. Production handles 1,000 requests per second. Address it through Conduct representative load and latency tests before ramping serving traffic, with the required justification or evidence: Passing low-load checks does not demonstrate sufficient throughput or acceptable tail latency under real demand.

C:

Route traffic back to the known-good deployment using endpoint traffic weights. This is useful for the distinct issue: A new model's production error rate rises sharply during the initial 10-percent rollout. The relevant technical rationale or evidence for that issue is: Traffic rollback can restore the previous serving model without deleting evidence from the failed candidate deployment. Here the reported issue is: A team promotes a new model based only on successful smoke tests at one request per second. Production handles 1,000 requests per second. The relevant response is: Conduct representative load and latency tests before ramping serving traffic. The first result cannot verify this different correction: Passing low-load checks does not demonstrate sufficient throughput or acceptable tail latency under real demand.

D:

Preserve the immutable prior model artifact, environment, and deployment manifest. It addresses another case: The operator needs a reproducible way to return to an exact prior model if rollback occurs weeks after the current release. For that case, the distinguishing justification or acceptance check is: Rollback must restore known code and dependencies, not rely on an ambiguous latest registry reference. The present case instead states: A team promotes a new model based only on successful smoke tests at one request per second. Production handles 1,000 requests per second. It calls for Conduct representative load and latency tests before ramping serving traffic. Showing success for the other case would not establish: Passing low-load checks does not demonstrate sufficient throughput or acceptable tail latency under real demand.

Operationalizing Machine Learning and Generative AI Solutions (AI-300) Practice Tests