Question 380
A dependent service is persistently failing, and every incoming request triggers more doomed calls. What is the purpose of a circuit breaker?
- To erase all error telemetry once a threshold is reached.
- To guarantee that every remote operation has succeeded.
- Immediately route all traffic back to the dependency after one successful probe without considering the recovery policy or remaining failure rate.
- To convert failed results into fabricated successful data.
- To temporarily stop repeated calls to the failing dependency and probe recovery under a defined policy.
Explanation
Correct Answer
E:Failing fast reduces wasted load while controlled probes can detect when the dependency becomes usable again. An open circuit protects both the caller and the struggling dependency from repeated wasted work. Recovery probes should be limited so testing availability does not recreate the original surge.
Incorrect Answers
A:Observability remains necessary to understand the outage and recovery behavior.
B:A circuit breaker manages calls under failure; it cannot guarantee remote operation success.
C:A single success may not show stable recovery under load. Controlled transition and monitoring help avoid recreating the same overload cycle.
D:A resilience mechanism must not manufacture substantive task outcomes.
Question 381
A workflow allows five retries, each with a 30-second timeout, but the user-facing request has a 45-second deadline. What should the retry controller enforce?
- Only the timeout of the last attempt.
- Five full retries regardless of the user-facing deadline.
- A remaining total deadline budget that constrains attempt duration and whether another retry can begin.
- Restart the 45-second deadline for each attempt because each retry is a new HTTP request.
Explanation
Correct Answer
C:Whole-request budgeting prevents individually bounded attempts from producing an unbounded or excessive total delay. Elapsed time includes prior calls, delays, and retries, not just the current attempt. The controller may need to decline another retry when its expected duration cannot fit the remaining window.
Incorrect Answers
A:Earlier attempts also consume the end-to-end budget.
B:Per-attempt limits can accumulate beyond the time available for the overall task.
D:The user-facing deadline covers the entire operation, including prior attempts and waits. Resetting it per attempt defeats that end-to-end constraint.
Question 382
A streaming response emits several text fragments and then terminates with a transport error before successful completion. Which TWO handling choices are appropriate? Choose TWO.
- Assume an opening JSON brace proves a complete object arrived.
- Record the response as incomplete rather than silently treating the fragments as a finished answer.
- Use a bounded recovery policy that accounts for any already executed side effects.
- Mark every downstream action complete because some text was received.
- Append a fabricated ending and present it as the original model response.
Explanation
Correct Answers
B:An interrupted stream may omit critical conclusions or leave its output structurally unfinished.
C:Recovery must distinguish incomplete text from external actions that may already have occurred.
Incorrect Answers
A:A prefix cannot establish that the rest of the object or terminal status was received.
D:Partial generation is not evidence that the task or its actions completed.
E:Inventing missing content conceals the transport failure and may change the intended result.
Question 383
A consumer writes each partial JSON fragment directly into a production record. What should it do instead?
- Assemble and validate a complete successful result before atomically applying the intended record update.
- Treat every closing brace as proof that all application checks passed.
- Allow each fragment to overwrite arbitrary fields without a schema.
- Commit the first fragment and discard the rest of the response.
Explanation
Correct Answer
A:A completion boundary prevents partial generation from becoming a partially corrupted production record. A partial object can omit required values or leave a record temporarily inconsistent. Validating before one deliberate commit keeps intermediate generation separate from authoritative application state.
Incorrect Answers
B:Syntax alone does not establish the correct final object or its business validity.
C:Uncontrolled partial writes can expose inconsistent intermediate state.
D:An initial fragment rarely constitutes the complete validated record.
Question 384
A user reports one failed API request among thousands of successful calls. Which diagnostic metadata is most useful to retain with the error?
- Only an hourly aggregate failure count with no link to individual calls.
- The complete credential used for the request in a public log.
- Only the model's final response text, omitting transport and trace metadata.
- The service request ID, timestamp, relevant status, and application trace correlation.
Explanation
Correct Answer
D:Correlated identifiers let operators connect the user-visible failure to a specific request and workflow event. A request ID narrows provider-side investigation, while an application trace links it to the user's workflow. Useful diagnostics can be retained without copying credentials into the log.
Incorrect Answers
A:Aggregates reveal a trend but cannot isolate the user's particular failed request.
B:A secret value is unnecessary for correlation and should not be exposed for diagnostics.
C:Answer text may not identify the specific failed API call or its operational cause.
Question 385
Average response time is acceptable, but some users wait far beyond the product limit. Which monitoring addition is most relevant?
- The overall request count without a response-time distribution.
- Only the fastest observed response.
- Only the mean latency of successful requests, excluding timed-out requests from observation.
- High-percentile latency and timeout rates segmented by relevant workload.
Explanation
Correct Answer
D:Tail measurements reveal slow-request behavior hidden by an acceptable mean. For example, an acceptable median can coexist with an unacceptable high-percentile delay. Workload segmentation helps distinguish large-document tasks from routine queries rather than averaging their experiences together.
Incorrect Answers
A:Volume alone does not reveal which fraction of users experiences excessive latency.
B:The minimum latency says nothing about the slow experiences causing complaints.
C:Filtering out timeouts can make the service look responsive while omitting the users with the worst experience. Tail latency and timeout rates expose that hidden group.
Question 386
A search cache keys only on the text 'holiday policy.' It serves results across countries whose employment policies differ. What should the cache key or partitioning include?
- The relevant retrieval context, including jurisdiction and other filters that change the valid result set.
- The final answer's length preference while omitting jurisdiction.
- A constant key so all policy queries can share one result.
- Only the number of characters in the query.
Explanation
Correct Answer
A:Cache identity must preserve the dimensions that determine which evidence answers the request. Jurisdiction changes which documents can answer the question even when query text is identical. Reusing a result across that boundary can produce a fluent answer based on the wrong policy.
Incorrect Answers
B:Response length does not distinguish the underlying country-specific evidence.
C:Broader cache collision would increase the chance of returning an inapplicable policy.
D:Equal-length queries do not identify the policy jurisdiction needed by the task.
Question 387
A workflow has retry and fallback code, but it has only been tested against healthy services. Which THREE tests provide meaningful resilience evidence? Choose THREE.
- Check that retry and fallback branches appear in source code without executing them under failure conditions.
- Inject a partial-success condition and check reconciliation of completed and incomplete work.
- Inject a persistent outage and verify the terminal fallback or failure report.
- Measure latency against a healthy dependency and use that measurement as evidence that an outage will terminate correctly.
- Inject a transient dependency failure and observe bounded recovery.
- Rerun only successful fixtures while checking that the fallback function exists.
Explanation
Correct Answers
B:Mixed outcomes test whether recovery preserves successful work and accurately accounts for the remainder.
C:A lasting fault checks whether the workflow stops appropriately rather than looping indefinitely.
E:A controlled temporary fault tests whether retries recover without violating their limits.
Incorrect Answers
A:Static presence does not demonstrate correct timing, bounds, or recovery behavior. These paths need observations from the conditions they are meant to handle.
D:Healthy-path timing does not exercise retry exhaustion or fallback. The terminal behavior must be tested while the dependency remains unavailable.
F:Code presence and healthy-path reruns do not demonstrate that resilience branches work under faults.
Question 388
A new prompt version passes offline checks but changes a production-critical extraction workflow. Which rollout provides additional controlled evidence?
- Change the model, prompt, corpus, and schema simultaneously with no version records.
- Replace the old version for every user without monitoring.
- Send only the easiest internal demonstration queries to the new version and treat that result as representative of the production-critical workload.
- Deploy to a limited representative traffic slice with monitored acceptance criteria and a rollback path.
- Route all traffic to the new version but retain a rollback button without monitoring acceptance criteria.
Explanation
Correct Answer
D:A canary reveals real-workload behavior while limiting the impact of a regression. The slice should include enough representative cases to reveal relevant regressions. Predefined stop criteria make rollback a measurable response to evidence rather than an improvised judgment.
Incorrect Answers
A:Untracked simultaneous changes make regression attribution and reversal difficult.
B:An unmonitored all-at-once change maximizes exposure to failures not found offline.
C:A deliberately easy slice can miss the conditions that matter in production. Canary evidence needs a relevant input mix and monitored acceptance criteria.
E:A rollback mechanism alone does not bound initial exposure or reveal when to use it. A monitored limited rollout supplies evidence before broad adoption.
Question 389
A release changes the prompt, output schema, and retrieval configuration. Rolling back only the prompt leaves the service broken. What should release management version?
- A free-text note saying 'latest settings.'
- Only the source document filenames without their versions.
- The compatible set of prompt, schema, model selection, retrieval settings, and dependent code.
- Only the human-readable release name, allowing each component to resolve its own latest configuration at startup.
Explanation
Correct Answer
C:A coherent release bundle supports restoring a known working combination rather than an incompatible subset. Prompt and schema changes can depend on one another, so reverting just one can leave an invalid interface. A release record should identify the exact compatible configuration to restore.
Incorrect Answers
A:An unversioned label cannot identify the exact configuration to restore.
B:Filenames alone cannot reconstruct the released configuration or content state.
D:A shared name does not identify a compatible set if its components drift independently. Recovery needs the specific versions that worked together.
Question 390
A fixed prompt and dataset produce different outcomes across repeated runs. What should a reliability estimate report?
- A guarantee that the next run will match the most recent one.
- Change the prompt between every run and attribute all score variation to repeated sampling.
- Performance over enough repeated trials to characterize variability on consequential cases.
- Only the best run because all runs used the same prompt.
Explanation
Correct Answer
C:Repeated observations reveal instability that a single favorable run can hide. Reporting the distribution or range of relevant outcomes is more informative than publishing only the best run. Consequential unstable cases may need a stronger control even if average accuracy is acceptable.
Incorrect Answers
A:Observed variation provides no basis for that deterministic prediction.
B:Changing the prompt introduces a second cause of variation. Repeated-run reliability needs the relevant configuration held fixed so its variability can be characterized.
D:Selecting the maximum produces an optimistic estimate of ordinary behavior.