Question 270
Many workers retry a transient failure after exactly the same delay. Why add jitter?
- To guarantee that every retried call completes before its deadline.
- To make malformed requests valid.
- To remove all need for retry limits.
- To spread retry attempts and reduce synchronized bursts.
Explanation
Correct Answer
D:Randomized delays spread retry traffic over time, reducing synchronized bursts that can overwhelm a recovering dependency.
Incorrect Answers
A:Jitter spreads load but cannot guarantee service recovery or eliminate the need for bounded retry and deadline policies.
B:Timing variation cannot repair invalid request content.
C:Jitter controls synchronization, not total attempts or deadlines.
Question 271
A service doubles completed requests per minute but individual requests take longer. What changed?
- Both metrics necessarily improved.
- Only the model's factual accuracy, because throughput is a quality score.
- Throughput improved while per-request latency worsened.
- Neither metric can be measured independently.
Explanation
Correct Answer
C:The metrics measure different properties and can move in opposite directions under batching or increased concurrency.
Incorrect Answers
A:Higher volume does not imply faster individual completion.
B:Completed requests per time measures capacity, while correctness requires a separate task-quality metric.
D:Throughput and latency are distinct observable quantities.
Question 272
A retry deadline is measured using the system wall clock. A clock synchronization moves that clock backward, extending the retry loop unexpectedly. What should elapsed-time budgeting use?
- A longer timeout expressed in wall-clock minutes.
- A monotonic clock for elapsed durations.
- Convert the wall clock to UTC before measuring the elapsed interval.
- Count completed retry attempts and assume each took a fixed duration.
Explanation
Correct Answer
B:A monotonic time source is designed not to move backward with wall-clock corrections. Calendar timestamps can still be recorded separately for human-readable logs.
Incorrect Answers
A:Increasing the allowance does not repair the underlying measurement instability when the chosen clock moves backward.
C:UTC removes timezone ambiguity but does not prevent wall-clock adjustments. A monotonic source is needed for durations unaffected by that correction.
D:Network calls and backoff durations can vary, so attempt counts do not reliably measure elapsed time or enforce the actual deadline.
Question 273
A Python object graph contains a reference cycle and cannot be encoded as ordinary JSON. What should the request builder do?
- Send the object’s memory address as its full data.
- Construct an explicit serializable data-transfer representation.
- Disable circular-reference checks without changing the object graph.
- Use str(object) and treat the diagnostic representation as JSON.
Explanation
Correct Answer
B:Create an explicit serializable request structure containing the needed values, rather than attempting to transmit process-local object relationships or addresses.
Incorrect Answers
A:An address does not transmit the referenced structure.
C:The cycle remains and recursive encoding can still fail or recurse indefinitely. Removing detection does not create a finite supported JSON representation.
D:An object's display representation is not a JSON serialization contract. Construct a finite data-transfer shape containing the fields intended for transmission.
Question 274
An internal API documents a read-only GET endpoint and a separate POST endpoint for creation. Which should a lookup tool use?
- A guessed DELETE endpoint.
- POST creation because any successful response is sufficient.
- The documented GET lookup endpoint.
- Both endpoints on every lookup.
Explanation
Correct Answer
C:Use the documented lookup method and preserve its read semantics; choosing a creation endpoint would introduce an unintended state change.
Incorrect Answers
A:An undocumented destructive operation does not implement the stated read requirement.
B:A response from the wrong operation can create unwanted state rather than answer the lookup.
D:Unnecessary creation introduces side effects absent from the request.
Question 275
Simple requests meet quality targets on a cheaper model, while difficult requests do not. What strategy merits evaluation?
- Send every difficult request to the cheapest model regardless of evidence.
- Route requests randomly without measuring category difficulty.
- Assume the router itself cannot make errors.
- Route by tested task difficulty with a fallback for uncertain routing.
Explanation
Correct Answer
D:A measured policy can use lower-cost capacity where adequate while preserving stronger handling for difficult or ambiguous cases.
Incorrect Answers
A:That knowingly fails the measured quality requirement for the difficult slice.
B:Random routing does not preserve the observed relationship between task difficulty and which model meets quality targets.
C:Routing quality must be evaluated because mistakes affect the downstream model choice.
Question 276
On the same workload, model A matches model B's quality, costs less, and has lower latency. No other requirement favors B. How should B be treated in this measured comparison?
- As equivalent to A because their quality scores match.
- As dominated by A on the stated criteria.
- As preferable because its higher price proves greater reliability.
- As automatically preferable whenever it is the newer release.
Explanation
Correct Answer
B:B provides no measured advantage and is worse on cost and latency. Removing it from the efficient candidate set follows from the supplied comparison, not from a claim that it is inferior on every possible workload.
Incorrect Answers
A:Equal quality does not erase the cost and latency differences. The product is comparing all three stated criteria.
C:The scenario supplies equal measured quality and no additional benefit. Price alone cannot establish an unstated reliability advantage.
D:Release age is not one of the demonstrated benefits. A newer label does not overturn the measured criteria in this scenario.
Question 277
A newer model rejects a previously used request parameter. What is required before migration?
- Ignore every rejected request during rollout.
- Adapt configuration using current migration guidance and rerun compatibility tests.
- Assume an SDK upgrade alone proves behavioral compatibility.
- Rename the parameter arbitrarily.
Explanation
Correct Answer
B:A model upgrade can change supported controls; keeping old parameters unchanged is not automatically safe. Adapt the request to the selected release's supported parameters and re-evaluate behavior before shifting production traffic.
Incorrect Answers
A:Dropping failures hides an unresolved migration defect.
C:The application’s request and output expectations still need validation.
D:Invented fields do not satisfy the new contract.
Question 278
A fallback model is available during outages but fails the required extraction accuracy. What should the design do?
- Treat any generated answer as an acceptable fallback.
- Use a constrained degraded mode or human path rather than claiming normal-quality equivalence.
- Hide which cases were degraded from operational records.
- Assume all model families are interchangeable.
Explanation
Correct Answer
B:A useful fallback must meet a defined minimum quality contract or route to a clearly stated degraded or human-assisted path.
Incorrect Answers
A:Output existence is not evidence of required accuracy.
C:This prevents measuring the impact of fallback behavior.
D:Capabilities and measured performance can differ materially.
Question 279
A task requires a specific supported input modality. What should model selection check before comparing benchmark scores?
- Only a low token price without checking input-type support.
- Only the model’s conversational tone.
- Only the highest published general benchmark.
- Required modality support on the selected access platform.
Explanation
Correct Answer
D:A high general score is irrelevant if the application cannot submit its necessary input through that configuration.
Incorrect Answers
A:An inexpensive model cannot serve a required modality that the chosen model-platform combination does not accept.
B:Tone does not provide missing input support.
C:General capability does not establish feature compatibility.
Question 280
A team upgrades models while keeping the same named effort setting. What should it avoid assuming?
- Identical behavior from the same setting across releases.
- That settings should be recorded.
- That supported combinations should be verified.
- That the task’s acceptance criteria still apply.
Explanation
Correct Answer
A:The same setting label can map to different behavior on another release, so retain configuration records and run workload-specific comparisons.
Incorrect Answers
B:Recording them remains useful for comparison.
C:Compatibility checks remain necessary during migration.
D:Business requirements continue to govern the upgraded system.