Question 100
A managed-identity request receives 403 even though an access token was issued. Where should the engineer look?
- Restrict public network access according to the resource's supported networking controls
- Use a role that grants the required AI service data actions, not only management read access
- Redact secret values from application and deployment logs
- Check that the identity has the required data-plane role on the AI resource
Explanation
Correct Answer
D:Obtaining a token proves authentication; authorization can still fail when the principal lacks a suitable role or scope. In the regulated Azure AI application, confirm the fix by retesting the affected behavior after checking that the identity has the required data-plane role on the AI resource.
Incorrect Answers
A:Although useful elsewhere in the regulated Azure AI application, creating a private endpoint does not necessarily disable the public access path; both must be governed deliberately; the present case is different and calls for the team to check that the identity has the required data-plane role on the AI resource.
B:Management-plane visibility and data-plane inference permissions are separate authorization concerns in its own workflow, but that is a different correction from the present requirement to check that the identity has the required data-plane role on the AI resource. The reported issue would remain.
C:That change addresses another failure mode: secure storage is undermined if retrieved credentials are written into accessible telemetry; it does not address this case, for which the relevant intervention is to check that the identity has the required data-plane role on the AI resource.
Question 101
An agent occasionally calls a tool twice, causing duplicate billing operations. Which evidence can confirm the sequence?
- Use aggregate metrics for trends and separate traces for individual request paths
- Inspect correlated tool-call spans and identifiers for the affected conversation
- Standardize supported diagnostic destinations and retention policies by environment
- Include model-specific input/output token use and applicable pricing meters
Explanation
Correct Answer
B:Request traces can reveal duplicate invocations and their order; aggregated request counts cannot prove a single workflow's actions. In the customer-facing AI service under an SLO, confirm the fix by retesting the affected behavior after inspecting correlated tool-call spans and identifiers for the affected conversation.
Incorrect Answers
A:That change addresses another failure mode: resource metrics summarize behavior across traffic; they do not reconstruct each retrieval, tool, and model operation; it does not address this case, for which the relevant intervention is to inspect correlated tool-call spans and identifiers for the affected conversation.
C:Although useful elsewhere in the customer-facing AI service under an SLO, consistent routing makes correlated incident analysis possible without assuming a resource automatically retains all logs forever; the present case is different and calls for the team to inspect correlated tool-call spans and identifiers for the affected conversation.
D:Inference billing depends on usage dimensions and model/deployment terms, not merely HTTP request volume in its own workflow, but that is a different correction from the present requirement to inspect correlated tool-call spans and identifiers for the affected conversation. The reported issue would remain.
Question 102
A trace includes sensitive user messages and private documents. What should be configured before enabling broad telemetry access?
- Review actual quota allocation by relevant model, region, subscription, and deployment
- Treat cost alerts as notifications and implement separate authorized enforcement if needed
- Redact or minimize sensitive prompt content and restrict access to trace data
- Verify the selected diagnostic categories and their destination configuration
Explanation
Correct Answer
C:Tracing can expose the very data an AI application processes; observability needs privacy and access controls. In the customer-facing AI service under an SLO, confirm the fix by retesting the affected behavior after redacting or minimize sensitive prompt content and restrict access to trace data.
Incorrect Answers
A:That change addresses another failure mode: quota may be shared or constrained by specific dimensions; a single displayed number is not necessarily per-team capacity; it does not address this case, for which the relevant intervention is to redact or minimize sensitive prompt content and restrict access to trace data.
B:A standard budget threshold does not itself guarantee resource suspension or hard cost containment in its own workflow, but that is a different correction from the present requirement to redact or minimize sensitive prompt content and restrict access to trace data. The reported issue would remain.
D:Although useful elsewhere in the customer-facing AI service under an SLO, creating a settings object without enabling relevant categories or a valid destination does not populate useful records; the present case is different and calls for the team to redact or minimize sensitive prompt content and restrict access to trace data.
Question 103
In a customer-facing AI service under an SLO, engineers must follow a request through retrieval, prompts, tools, and inference steps to isolate slow operations. The expected evidence is that a slow user turn is attributable to a specific component span. Which implementation is the best match?
- Configure Azure Cost Management budgets and alerts
- Enable Azure diagnostic settings
- Enable application tracing
- Inspect Azure Monitor service metrics
Explanation
Correct Answer
C:Distributed traces record spans for the application workflow, revealing where time and errors occur when instrumentation is configured. For this customer-facing AI service under an SLO, the intended evidence is that a slow user turn is attributable to a specific component span. A trace is not an Azure billing invoice or an image classification score.
Incorrect Answers
A:Cost Management budgets compare scoped expenses with configured limits and can notify designated parties; it may be used to notify finance when forecast or actual consumption passes a spending threshold in a customer-facing AI service under an SLO, but the current requirement is instead to follow a request through retrieval, prompts, tools, and inference steps to isolate slow operations. That separate capability does not prove that a slow user turn is attributable to a specific component span; A budget alert is not a request-rate throttle or a content safety filter, so selecting that alternative would not meet the stated acceptance check.
B:Diagnostic settings export available resource telemetry to targets such as Log Analytics or storage; it may be used to route supported resource logs and metrics to a monitoring destination for investigation in a customer-facing AI service under an SLO, but the current requirement is instead to follow a request through retrieval, prompts, tools, and inference steps to isolate slow operations. That separate capability does not prove that a slow user turn is attributable to a specific component span; They do not automatically enforce a monthly spending ceiling, so selecting that alternative would not meet the stated acceptance check.
D:Platform metrics aggregate measurable service behavior and help detect capacity or health trends; it may be used to measure resource throughput, latency, and request failures over time in a customer-facing AI service under an SLO, but the current requirement is instead to follow a request through retrieval, prompts, tools, and inference steps to isolate slow operations. That separate capability does not prove that a slow user turn is attributable to a specific component span; They do not preserve detailed prompt content or serve as financial budget limits, so selecting that alternative would not meet the stated acceptance check.
Question 104
The SRE team needs to retain Azure AI resource diagnostic events for post-incident queries in Log Analytics. Which TWO setup actions belong in the logging configuration? (Choose TWO.)
- Compare time-series request latency percentiles before and after the deployment
- Set the diagnostic destination to the intended Log Analytics workspace
- Handle 429 responses with bounded exponential backoff and service-aware retries
- Enable the relevant supported resource diagnostic log categories
- Create a cost budget with actual or forecast threshold notifications
Explanation
Correct Answers
B:Routing the enabled categories to a retained workspace makes investigation and query correlation possible after the incident has ended.
D:Selecting diagnostic categories controls which service events are exported; merely creating a monitoring workspace does not turn on those categories.
Incorrect Answers
A:Latency distributions reveal tail delays that a simple success count or monthly bill cannot diagnose in its own workflow, but that is a different correction from the present requirement to enable the relevant supported resource diagnostic log categories and set the diagnostic destination to the intended Log Analytics workspace. The reported issue would remain.
C:Although useful elsewhere in the customer-facing AI service under an SLO, respectful retries recover from transient throttling without amplifying overload; unbounded immediate retries are counterproductive; the present case is different and calls for the team to enable the relevant supported resource diagnostic log categories and set the diagnostic destination to the intended Log Analytics workspace.
E:That change addresses another failure mode: cost Management budgets can alert on usage thresholds; they are monitoring controls, not an automatic hard spending cap; it does not address this case, for which the relevant intervention is to enable the relevant supported resource diagnostic log categories and set the diagnostic destination to the intended Log Analytics workspace.
Question 105
A developer increases client concurrency after seeing 429 failures. What should be done first?
- Standardize supported diagnostic destinations and retention policies by environment
- Include model-specific input/output token use and applicable pricing meters
- Measure utilization against request and token rate limits before increasing load
- Inspect correlated tool-call spans and identifiers for the affected conversation
Explanation
Correct Answer
C:More concurrent calls can aggravate throttling; capacity diagnosis is necessary before scaling producers. In the customer-facing AI service under an SLO, confirm the fix by retesting the affected behavior after measuring utilization against request and token rate limits before increasing load.
Incorrect Answers
A:Although useful elsewhere in the customer-facing AI service under an SLO, consistent routing makes correlated incident analysis possible without assuming a resource automatically retains all logs forever; the present case is different and calls for the team to measure utilization against request and token rate limits before increasing load.
B:That change addresses another failure mode: inference billing depends on usage dimensions and model/deployment terms, not merely HTTP request volume; it does not address this case, for which the relevant intervention is to measure utilization against request and token rate limits before increasing load.
D:Request traces can reveal duplicate invocations and their order; aggregated request counts cannot prove a single workflow's actions in its own workflow, but that is a different correction from the present requirement to measure utilization against request and token rate limits before increasing load. The reported issue would remain.
Question 106
A text generation deployment accepts the first burst but rejects later requests with rate-limit errors. What should clients implement?
- Send supported diagnostic logs to a retained Log Analytics workspace
- Create a cost budget with actual or forecast threshold notifications
- Handle 429 responses with bounded exponential backoff and service-aware retries
- Compare time-series request latency percentiles before and after the deployment
Explanation
Correct Answer
C:Respectful retries recover from transient throttling without amplifying overload; unbounded immediate retries are counterproductive. In the customer-facing AI service under an SLO, confirm the fix by retesting the affected behavior after handling 429 responses with bounded exponential backoff and service-aware retries.
Incorrect Answers
A:Although useful elsewhere in the customer-facing AI service under an SLO, diagnostic settings export supported service logs for later querying; metrics alone may not capture the operational events; the present case is different and calls for the team to handle 429 responses with bounded exponential backoff and service-aware retries.
B:Cost Management budgets can alert on usage thresholds; they are monitoring controls, not an automatic hard spending cap in its own workflow, but that is a different correction from the present requirement to handle 429 responses with bounded exponential backoff and service-aware retries. The reported issue would remain.
D:That change addresses another failure mode: latency distributions reveal tail delays that a simple success count or monthly bill cannot diagnose; it does not address this case, for which the relevant intervention is to handle 429 responses with bounded exponential backoff and service-aware retries.
Question 107
An engineer claims that an Azure Monitor chart is a substitute for request-level prompt tracing. What distinction should be made?
- Measure utilization against request and token rate limits before increasing load
- Standardize supported diagnostic destinations and retention policies by environment
- Inspect correlated tool-call spans and identifiers for the affected conversation
- Use aggregate metrics for trends and separate traces for individual request paths
Explanation
Correct Answer
D:Resource metrics summarize behavior across traffic; they do not reconstruct each retrieval, tool, and model operation. In the customer-facing AI service under an SLO, confirm the fix by retesting the affected behavior after using aggregate metrics for trends and separate traces for individual request paths.
Incorrect Answers
A:Although useful elsewhere in the customer-facing AI service under an SLO, more concurrent calls can aggravate throttling; capacity diagnosis is necessary before scaling producers; the present case is different and calls for the team to use aggregate metrics for trends and separate traces for individual request paths.
B:That change addresses another failure mode: consistent routing makes correlated incident analysis possible without assuming a resource automatically retains all logs forever; it does not address this case, for which the relevant intervention is to use aggregate metrics for trends and separate traces for individual request paths.
C:Request traces can reveal duplicate invocations and their order; aggregated request counts cannot prove a single workflow's actions in its own workflow, but that is a different correction from the present requirement to use aggregate metrics for trends and separate traces for individual request paths. The reported issue would remain.
Question 108
The product manager expects a budget alert to shut off the inference endpoint at the limit. What is the key correction?
- Correlate throttled requests with token throughput and the deployed model quota
- Treat cost alerts as notifications and implement separate authorized enforcement if needed
- Review actual quota allocation by relevant model, region, subscription, and deployment
- Verify the selected diagnostic categories and their destination configuration
Explanation
Correct Answer
B:A standard budget threshold does not itself guarantee resource suspension or hard cost containment. In the customer-facing AI service under an SLO, confirm the fix by retesting the affected behavior after treating cost alerts as notifications and implement separate authorized enforcement if needed.
Incorrect Answers
A:That change addresses another failure mode: a request limit or token ceiling can produce throttling even if ordinary service health metrics appear normal; it does not address this case, for which the relevant intervention is to treat cost alerts as notifications and implement separate authorized enforcement if needed.
C:Although useful elsewhere in the customer-facing AI service under an SLO, quota may be shared or constrained by specific dimensions; a single displayed number is not necessarily per-team capacity; the present case is different and calls for the team to treat cost alerts as notifications and implement separate authorized enforcement if needed.
D:Creating a settings object without enabling relevant categories or a valid destination does not populate useful records in its own workflow, but that is a different correction from the present requirement to treat cost alerts as notifications and implement separate authorized enforcement if needed. The reported issue would remain.
Question 109
A release for a customer-facing AI service under an SLO is blocked until the application can understand why inference traffic is throttled at peaks despite a healthy model. The tester will verify that request throttling correlates with the available capacity or limit. What should be implemented?
- Configure Azure Cost Management budgets and alerts
- Check quota and rate-limit configuration
- Enable application tracing
- Enable Azure diagnostic settings
Explanation
Correct Answer
B:Quotas and rate limits constrain allowed request volume or token throughput according to the applicable service and deployment. For this customer-facing AI service under an SLO, the intended evidence is that request throttling correlates with the available capacity or limit. Raising a dashboard's display interval will not increase serving entitlement.
Incorrect Answers
A:Cost Management budgets compare scoped expenses with configured limits and can notify designated parties; it may be used to notify finance when forecast or actual consumption passes a spending threshold in a customer-facing AI service under an SLO, but the current requirement is instead to understand why inference traffic is throttled at peaks despite a healthy model. That separate capability does not prove that request throttling correlates with the available capacity or limit; A budget alert is not a request-rate throttle or a content safety filter, so selecting that alternative would not meet the stated acceptance check.
C:Distributed traces record spans for the application workflow, revealing where time and errors occur when instrumentation is configured; it may be used to follow a request through retrieval, prompts, tools, and inference steps to isolate slow operations in a customer-facing AI service under an SLO, but the current requirement is instead to understand why inference traffic is throttled at peaks despite a healthy model. That separate capability does not prove that request throttling correlates with the available capacity or limit; A trace is not an Azure billing invoice or an image classification score, so selecting that alternative would not meet the stated acceptance check.
D:Diagnostic settings export available resource telemetry to targets such as Log Analytics or storage; it may be used to route supported resource logs and metrics to a monitoring destination for investigation in a customer-facing AI service under an SLO, but the current requirement is instead to understand why inference traffic is throttled at peaks despite a healthy model. That separate capability does not prove that request throttling correlates with the available capacity or limit; They do not automatically enforce a monthly spending ceiling, so selecting that alternative would not meet the stated acceptance check.
Question 110
Operations sees a rise in HTTP 429 responses despite steady end-user traffic. What metric relationship is most useful?
- Correlate throttled requests with token throughput and the deployed model quota
- Treat cost alerts as notifications and implement separate authorized enforcement if needed
- Redact or minimize sensitive prompt content and restrict access to trace data
- Review actual quota allocation by relevant model, region, subscription, and deployment
Explanation
Correct Answer
A:A request limit or token ceiling can produce throttling even if ordinary service health metrics appear normal. In the customer-facing AI service under an SLO, confirm the fix by retesting the affected behavior after correlating throttled requests with token throughput and the deployed model quota.
Incorrect Answers
B:Although useful elsewhere in the customer-facing AI service under an SLO, a standard budget threshold does not itself guarantee resource suspension or hard cost containment; the present case is different and calls for the team to correlate throttled requests with token throughput and the deployed model quota.
C:Tracing can expose the very data an AI application processes; observability needs privacy and access controls in its own workflow, but that is a different correction from the present requirement to correlate throttled requests with token throughput and the deployed model quota. The reported issue would remain.
D:That change addresses another failure mode: quota may be shared or constrained by specific dimensions; a single displayed number is not necessarily per-team capacity; it does not address this case, for which the relevant intervention is to correlate throttled requests with token throughput and the deployed model quota.