Microsoft AI-300 (Operationalizing Machine Learning and Generative AI Solutions) Practice Test 26

Question 250

Operations must troubleshoot calls to a protected Azure Machine Learning online endpoint. The immediate design goal is to distinguish permission to invoke a model from permission to update its deployment configuration. Which approach addresses that goal?

  1. HTTP status investigation
  2. Data plane versus control plane
  3. Key-based invocation
  4. Token-based invocation

Explanation

Correct Answer

B:

Inference authorization and management privileges protect different operations. A user who can read endpoint metadata is not necessarily authorized to score requests.

Incorrect Answers

A:

Status and request logs narrow whether failure occurs in authentication, request parsing, or serving; retrying every error without classification obscures root causes; the outstanding need is to distinguish permission to invoke a model from permission to update its deployment configuration.

C:

Key authentication verifies possession of a configured secret for invocation; it is distinct from Azure RBAC authorization for managing the endpoint resource; the needed capability is to distinguish permission to invoke a model from permission to update its deployment configuration.

D:

Token invocation relies on supported token issuance and correct request headers; sending a storage account key does not authenticate an ML endpoint; the requirement, by contrast, is to distinguish permission to invoke a model from permission to update its deployment configuration.

Question 251

Operations must troubleshoot calls to a protected Azure Machine Learning online endpoint. The team is required to replace an exposed inference key and update clients through an approved secret distribution process. Which option is the best fit?

  1. HTTP status investigation
  2. Credential rotation
  3. Endpoint scoring URI
  4. Token-based invocation

Explanation

Correct Answer

B:

Key rotation limits continued use of a compromised credential. Changing traffic weights alone does not revoke knowledge of the old key.

Incorrect Answers

A:

Status and request logs narrow whether failure occurs in authentication, request parsing, or serving; retrying every error without classification obscures root causes; the requirement, by contrast, is to replace an exposed inference key and update clients through an approved secret distribution process.

C:

The scoring URI is the service data-plane interface for inference requests; resource-management endpoints control configuration and are not interchangeable prediction targets; the specified outcome instead requires the team to replace an exposed inference key and update clients through an approved secret distribution process.

D:

Token invocation relies on supported token issuance and correct request headers; sending a storage account key does not authenticate an ML endpoint; the implementation under review needs to replace an exposed inference key and update clients through an approved secret distribution process.

Question 252

Operations must troubleshoot calls to a protected Azure Machine Learning online endpoint. An operator is tasked to use the specific response code to distinguish unauthorized calls, invalid payloads, and unhealthy inference workers. Which mechanism provides that capability?

  1. Data plane versus control plane
  2. Endpoint scoring URI
  3. Token-based invocation
  4. HTTP status investigation
  5. Key-based invocation

Explanation

Correct Answer

D:

Status and request logs narrow whether failure occurs in authentication, request parsing, or serving. Retrying every error without classification obscures root causes.

Incorrect Answers

A:

Inference authorization and management privileges protect different operations; a user who can read endpoint metadata is not necessarily authorized to score requests; the implementation under review needs to use the specific response code to distinguish unauthorized calls, invalid payloads, and unhealthy inference workers.

B:

The scoring URI is the service data-plane interface for inference requests; resource-management endpoints control configuration and are not interchangeable prediction targets; the requested task remains to use the specific response code to distinguish unauthorized calls, invalid payloads, and unhealthy inference workers.

C:

Token invocation relies on supported token issuance and correct request headers; sending a storage account key does not authenticate an ML endpoint; this alternative does not achieve the requirement to use the specific response code to distinguish unauthorized calls, invalid payloads, and unhealthy inference workers.

E:

Key authentication verifies possession of a configured secret for invocation; it is distinct from Azure RBAC authorization for managing the endpoint resource; the outstanding need is to use the specific response code to distinguish unauthorized calls, invalid payloads, and unhealthy inference workers.

Question 253

A production team plans a new model release without interrupting the existing service. The immediate design goal is to create a separate new deployment while preserving the previously serving deployment as fallback. Which approach addresses that goal?

  1. Release metric gate
  2. Blue-green deployment
  3. Immediate rollback
  4. Traffic mirroring

Explanation

Correct Answer

B:

Blue-green release keeps the old model ready as a rollback option. Deploying over the only running version makes recovery harder.

Incorrect Answers

A:

A metric-based gate ties promotion to observed production behavior; offline accuracy alone cannot establish a safe operational rollout; the actual design criterion is to create a separate new deployment while preserving the previously serving deployment as fallback.

C:

Traffic rollback restores client responses from a known-good deployment without waiting for retraining; preserving the prior deployment is the key prerequisite; this approach leaves the team unable to create a separate new deployment while preserving the previously serving deployment as fallback.

D:

Shadow traffic tests a candidate with realistic input without allowing its output to affect the user; it should not be confused with a live canary that serves some responses; the needed capability is to create a separate new deployment while preserving the previously serving deployment as fallback.

Question 254

A production team plans a new model release without interrupting the existing service. The team is required to send a small controlled fraction of live requests to a candidate and inspect its reliability before expansion. Which option is the best fit?

  1. Direct deployment test
  2. Traffic mirroring
  3. Canary traffic split
  4. Blue-green deployment

Explanation

Correct Answer

C:

Canary allocation limits exposure to a new model. The traffic percentages must be deliberately managed and sum to the intended routing allocation.

Incorrect Answers

A:

A deployment-specific invocation can validate a new version before general traffic is routed to it; such direct calls are not equivalent to changing default routing; the needed capability is to send a small controlled fraction of live requests to a candidate and inspect its reliability before expansion.

B:

Shadow traffic tests a candidate with realistic input without allowing its output to affect the user; it should not be confused with a live canary that serves some responses; the requirement, by contrast, is to send a small controlled fraction of live requests to a candidate and inspect its reliability before expansion.

D:

Blue-green release keeps the old model ready as a rollback option; deploying over the only running version makes recovery harder; the specified outcome instead requires the team to send a small controlled fraction of live requests to a candidate and inspect its reliability before expansion.

Question 255

A production team plans a new model release without interrupting the existing service. An operator is tasked to copy a fraction of real requests to a candidate while keeping client-visible responses from the existing deployment. Which mechanism provides that capability?

  1. Blue-green deployment
  2. Release metric gate
  3. Canary traffic split
  4. Traffic mirroring

Explanation

Correct Answer

D:

Shadow traffic tests a candidate with realistic input without allowing its output to affect the user. It should not be confused with a live canary that serves some responses.

Incorrect Answers

A:

Blue-green release keeps the old model ready as a rollback option; deploying over the only running version makes recovery harder; the specified outcome instead requires the team to copy a fraction of real requests to a candidate while keeping client-visible responses from the existing deployment.

B:

A metric-based gate ties promotion to observed production behavior; offline accuracy alone cannot establish a safe operational rollout; the implementation under review needs to copy a fraction of real requests to a candidate while keeping client-visible responses from the existing deployment.

C:

Canary allocation limits exposure to a new model; the traffic percentages must be deliberately managed and sum to the intended routing allocation; the requested task remains to copy a fraction of real requests to a candidate while keeping client-visible responses from the existing deployment.

Question 256

A production team plans a new model release without interrupting the existing service. The immediate design goal is to invoke a named candidate deployment intentionally while its general endpoint traffic weight remains zero. Which approach addresses that goal?

  1. Direct deployment test
  2. Release metric gate
  3. Traffic mirroring
  4. Blue-green deployment

Explanation

Correct Answer

A:

A deployment-specific invocation can validate a new version before general traffic is routed to it. Such direct calls are not equivalent to changing default routing.

Incorrect Answers

B:

A metric-based gate ties promotion to observed production behavior; offline accuracy alone cannot establish a safe operational rollout; the actual design criterion is to invoke a named candidate deployment intentionally while its general endpoint traffic weight remains zero.

C:

Shadow traffic tests a candidate with realistic input without allowing its output to affect the user; it should not be confused with a live canary that serves some responses; the outstanding need is to invoke a named candidate deployment intentionally while its general endpoint traffic weight remains zero.

D:

Blue-green release keeps the old model ready as a rollback option; deploying over the only running version makes recovery harder; this approach leaves the team unable to invoke a named candidate deployment intentionally while its general endpoint traffic weight remains zero.

Question 257

A production team plans a new model release without interrupting the existing service. The team is required to shift production traffic back to the previous healthy deployment after candidate errors breach a release threshold. Which option is the best fit?

  1. Release metric gate
  2. Canary traffic split
  3. Immediate rollback
  4. Direct deployment test

Explanation

Correct Answer

C:

Traffic rollback restores client responses from a known-good deployment without waiting for retraining. Preserving the prior deployment is the key prerequisite.

Incorrect Answers

A:

A metric-based gate ties promotion to observed production behavior; offline accuracy alone cannot establish a safe operational rollout; this approach leaves the team unable to shift production traffic back to the previous healthy deployment after candidate errors breach a release threshold.

B:

Canary allocation limits exposure to a new model; the traffic percentages must be deliberately managed and sum to the intended routing allocation; the needed capability is to shift production traffic back to the previous healthy deployment after candidate errors breach a release threshold.

D:

A deployment-specific invocation can validate a new version before general traffic is routed to it; such direct calls are not equivalent to changing default routing; the team would still need to shift production traffic back to the previous healthy deployment after candidate errors breach a release threshold.

Question 258

A new Azure ML model needs to be released beside the current serving version without losing the old deployment as fallback. The team must test the named candidate directly while its general endpoint traffic weight remains zero; after an independently configured canary period, promotion is allowed only if latency and error results meet approved limits. Which THREE release mechanisms support this workflow? Choose THREE.

  1. Direct deployment test
  2. Immediate rollback
  3. Release metric gate
  4. Canary traffic split
  5. Traffic mirroring
  6. Blue-green deployment

Explanation

Correct Answers

A:

A deployment-specific invocation can validate a new version before general traffic is routed to it; such direct calls are not equivalent to changing default routing. Here, the named candidate is exercised without routing ordinary endpoint traffic to it.

C:

A metric-based gate ties promotion to observed production behavior; offline accuracy alone cannot establish a safe operational rollout. Here, promotion remains conditional on measured errors and latency during the controlled period.

F:

Blue-green release keeps the old model ready as a rollback option; deploying over the only running version makes recovery harder. Here, the known-good serving deployment stays available throughout the change.

Incorrect Answers

B:

Traffic rollback restores client responses from a known-good deployment without waiting for retraining; preserving the prior deployment is the key prerequisite; this mechanism leaves unresolved a preserved blue-green fallback, zero-weight candidate test, and promotion gate.

D:

Canary allocation limits exposure to a new model; the traffic percentages must be deliberately managed and sum to the intended routing allocation; the observed fault still involves a preserved blue-green fallback, zero-weight candidate test, and promotion gate.

E:

Shadow traffic tests a candidate with realistic input without allowing its output to affect the user; it should not be confused with a live canary that serves some responses; the immediate requirement instead concerns a preserved blue-green fallback, zero-weight candidate test, and promotion gate.

Question 259

An inference endpoint experiences errors after a serving image or model update. The immediate design goal is to inspect worker initialization and scoring exceptions for a deployment that returns HTTP 500 responses. Which approach addresses that goal?

  1. Deployment provisioning events
  2. Request payload replay
  3. Deployment container logs
  4. Health status monitoring
  5. Deployment version comparison

Explanation

Correct Answer

C:

Container logs reveal application faults and model loading errors. An endpoint resource showing succeeded provisioning does not rule out runtime exceptions.

Incorrect Answers

A:

Provisioning diagnostics identify resource creation and image build errors; evaluation plots do not explain insufficient compute quota; the requested task remains to inspect worker initialization and scoring exceptions for a deployment that returns HTTP 500 responses.

B:

Replaying a captured safe payload can reveal malformed fields or serialization bugs; sensitive user data should be removed or protected during diagnosis; this alternative does not achieve the requirement to inspect worker initialization and scoring exceptions for a deployment that returns HTTP 500 responses.

D:

Health signals reveal whether replicas are ready to serve real requests; the control-plane deployment state alone may not establish runtime availability; the outstanding need is to inspect worker initialization and scoring exceptions for a deployment that returns HTTP 500 responses.

E:

Version comparison narrows the change responsible for a regression; changing unrelated network policy before examining differences can delay recovery; this approach leaves the team unable to inspect worker initialization and scoring exceptions for a deployment that returns HTTP 500 responses.

Question 260

An inference endpoint experiences errors after a serving image or model update. The team is required to test compatible scoring code and runtime configuration locally before using cloud deployment cycles. Which option is the best fit?

  1. Health status monitoring
  2. Deployment container logs
  3. Local endpoint debugging
  4. Request payload replay

Explanation

Correct Answer

C:

Local execution can surface dependency and request-format mistakes early. It does not prove cloud network permissions or VM quota availability.

Incorrect Answers

A:

Health signals reveal whether replicas are ready to serve real requests; the control-plane deployment state alone may not establish runtime availability; the outstanding need is to test compatible scoring code and runtime configuration locally before using cloud deployment cycles.

B:

Container logs reveal application faults and model loading errors; an endpoint resource showing succeeded provisioning does not rule out runtime exceptions; this approach leaves the team unable to test compatible scoring code and runtime configuration locally before using cloud deployment cycles.

D:

Replaying a captured safe payload can reveal malformed fields or serialization bugs; sensitive user data should be removed or protected during diagnosis; the requirement, by contrast, is to test compatible scoring code and runtime configuration locally before using cloud deployment cycles.

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