Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 18

Question 170

A gateway retries three times, a client retries three times, and each tool also retries. A single failure causes many requests. What should be coordinated?

  1. A separate full retry allowance for every nested library
  2. The aggregate retry count only after all layers finish retrying
  3. Equal per-layer retry counts without sharing attempt or deadline information
  4. An end-to-end retry budget across layers

Explanation

Correct Answer

D:

Independent retry policies can multiply attempts. A shared deadline and attempt budget prevent recovery logic from overwhelming the same failing dependency.

Incorrect Answers

A:

Independent allowances multiply across layers and can exceed the intended request budget. The design needs accounting across the full call path.

B:

Observing amplification afterward helps diagnosis but does not limit it. Retry allowances need coordination before nested layers multiply attempts.

C:

Equal local limits can still multiply across nested layers. A shared overall allowance is needed to bound work for the original request.

Question 171

Thousands of clients receive throttling responses simultaneously and retry after the same fixed delay. Load spikes repeat. What helps spread recovery traffic?

  1. Randomized jitter within a bounded backoff policy
  2. An identical shorter delay for every client
  3. A longer but identical delay for every client
  4. More simultaneous retries for each failed request

Explanation

Correct Answer

A:

Jitter reduces synchronization among clients that failed together. It should complement the service's retry guidance and overall request deadline.

Incorrect Answers

B:

A shorter common delay preserves synchronization and can intensify the repeated spike. The missing property is distributed retry timing.

C:

A longer fixed delay can shift the next spike without spreading the clients apart. Random variation reduces synchronized retry waves within the allowed backoff bounds.

D:

Parallel retries multiply demand during an overload condition. They are likely to worsen the contention the rate limit is signaling.

Question 172

One tenant's large batch consumes the entire shared model-call allowance, starving other tenants. Which control best supports fair service?

  1. A single first-come queue with no tenant accounting
  2. Give every tenant the full global quota independently
  3. Let the model choose which tenant deserves access
  4. Tenant-aware admission limits within the shared quota

Explanation

Correct Answer

D:

Per-tenant budgets prevent one workload from monopolizing all capacity. The scheduler can preserve agreed service shares while respecting the overall limit.

Incorrect Answers

A:

Arrival order can let a large burst occupy the whole allowance. It does not protect the agreed service of other tenants.

B:

Independent full quotas oversubscribe the actual shared resource. Local limits must fit inside the global capacity constraint.

C:

Capacity allocation should follow explicit service policy. A language-model preference is not a dependable or auditable fairness mechanism.

Question 173

A streaming response delivers several text fragments and then an error event. What should the application infer?

  1. The response may be incomplete despite earlier successful fragments
  2. The initial HTTP success proves the entire answer completed
  3. All earlier fragments should be marked as independently verified facts
  4. The last visible sentence is necessarily the final answer

Explanation

Correct Answer

A:

Streaming progress does not guarantee successful completion. The client must track terminal status and avoid treating partial output as a validated final artifact.

Incorrect Answers

B:

Errors can occur after streaming begins. A successful connection establishment does not establish the outcome of the full generation.

C:

Token delivery is not factual verification. Partial text retains the same evidence and validation requirements as any generated content.

D:

A stream can stop at an arbitrary boundary. A complete-looking sentence does not prove that all required content or validation finished.

Question 174

A streamed tool-call argument arrives in fragments. The first fragment looks like a valid account identifier, but the remaining object is incomplete. What should the executor do?

  1. Fill missing arguments from a previous unrelated request
  2. Execute as soon as any identifier appears
  3. Treat every fragment as a separate tool request
  4. Wait for a complete validated call before execution

Explanation

Correct Answer

D:

Partial arguments can change meaning as more fields arrive. Executing before the call is complete risks using missing or malformed parameters.

Incorrect Answers

A:

Cross-request defaults can target the wrong resource or user. Missing current arguments should not be replaced with unrelated historical data.

B:

An identifier alone may not define the operation or its scope. Early execution can commit a different action from the final call.

C:

Fragments belong to one structured call. Interpreting them as independent requests can create duplicate or invalid operations.

Question 175

A workload contains independent nightly classifications with no immediate user waiting. Which processing design is worth evaluating for cost efficiency?

  1. A single prompt containing every record without size limits
  2. A promise that batch mode guarantees every item succeeds
  3. A permanently open interactive stream for each record
  4. A retry loop that waits for all records to finish synchronously
  5. An asynchronous batch path supported by the chosen platform

Explanation

Correct Answer

E:

The workload tolerates delayed completion, so it can use a batch contract if supported. The application must handle job status and individual result outcomes.

Incorrect Answers

A:

An unbounded combined request can exceed context or output limits. Batch execution still needs supported per-request boundaries and result tracking.

B:

Batch submission does not remove validation, capacity, or content failures. The client still needs per-item outcome handling and recovery.

C:

Interactive connections add operational overhead without serving a real-time requirement. They may be a poor fit for independent overnight work.

D:

Synchronous waiting ties resource usage to the entire job. An asynchronous contract better matches the allowed completion window.

Question 176

A high-volume client creates a new secure connection for every short model request. Connection setup dominates latency. What should be evaluated?

  1. Supported connection pooling and reuse
  2. DNS caching while opening a fresh TLS connection for every request
  3. A larger response buffer while preserving one connection per call
  4. More client threads while retaining a new connection for every call

Explanation

Correct Answer

A:

Reusing connections can reduce repeated setup overhead. Pool limits, timeout behavior, and isolation still need to fit the client and service contracts.

Incorrect Answers

B:

DNS caching can reduce lookup work but leaves the stated repeated connection setup in place. Connection reuse targets that handshake overhead more directly.

C:

Buffer size affects response handling rather than repeated connection establishment. It does not remove the setup cost dominating these short requests.

D:

Concurrency may overlap work but continues creating the same per-call handshake overhead. It can also increase connection pressure without fixing the identified cause.

Question 177

A request has a five-second deadline but permits a five-second retrieval call followed by a five-second model call. What is wrong with this configuration?

  1. The deadline is invalid unless every component receives the full five seconds
  2. Component limits do not fit the overall deadline
  3. The user should receive a success response at five seconds regardless
  4. Every component is safe because each limit equals five seconds

Explanation

Correct Answer

B:

Sequential stages consume the same end-to-end time allowance. Their budgets must include other work and leave enough time for completion or controlled fallback.

Incorrect Answers

A:

A component can receive a smaller allocation within the total deadline. The problem is that sequential allowances exceed the overall budget, not that each component needs the full amount.

C:

A deadline does not establish that work completed. The application must report the actual outcome or a defined partial/failure state.

D:

A local limit does not reset the user's deadline. Two sequential five-second stages can exceed the total even if neither violates its own timeout.

Question 178

A monitoring system samples only successful requests to reduce log volume. Why is this unsuitable for diagnosing production failures?

  1. Successful requests always contain the same errors as failed ones
  2. Increasing success sample volume guarantees failure coverage
  3. Sampling any subset is inherently invalid
  4. The sample systematically excludes the cases of interest

Explanation

Correct Answer

D:

Success-only sampling can make the system appear healthy while omitting failure paths. A diagnostic strategy should retain representative or prioritized error evidence.

Incorrect Answers

A:

Some defects occur only on exceptional paths. Their absence from successful traces is precisely why the sample cannot substitute for failure data.

B:

More observations from the wrong population do not add omitted failures. The inclusion policy must change, not just the sample size.

C:

Sampling can be useful when its design matches the question. The problem here is selection bias against the target failure population.

Question 179

An engineer proposes a separate metric label for every full user prompt. What is the main observability concern?

  1. Metrics require raw prompts to measure response time
  2. Unbounded cardinality and unnecessary sensitive content in metrics
  3. Unique labels automatically improve aggregation accuracy
  4. Hashing the prompt values would remove the cardinality concern

Explanation

Correct Answer

B:

Prompt values can create enormous label sets and expose user data. Metrics should use controlled dimensions, with protected traces for appropriate detail.

Incorrect Answers

A:

Timing can be measured with request identifiers and bounded dimensions. Full prompt text is not necessary for a latency metric.

C:

Excessive uniqueness can make aggregation expensive and less useful. A metric needs dimensions that support the intended comparison.

D:

A distinct hash for every prompt still creates a large number of label values. Hashing can obscure readable content without making the metric dimensions bounded.

Question 180

A trace records tool start and end times but no parent-child relationships. Concurrent branches are difficult to interpret. What should be added?

  1. Separate timestamp lists without parent-child relationships
  2. Only a global average tool duration
  3. Causal links between the request and its nested operations
  4. A longer final-answer transcript

Explanation

Correct Answer

C:

Parentage identifies which operation triggered another and supports reconstruction of concurrent paths. Timestamps alone may not reveal the dependency structure.

Incorrect Answers

A:

Timestamps can suggest order but do not reliably establish causality among concurrent operations. Nested spans need explicit relationships to the originating request.

B:

An aggregate removes more transaction detail. It cannot reconstruct which nested call belongs to which branch of the request.

D:

Answer prose does not reliably encode the execution graph. Causality should be captured by instrumentation as operations occur.

Claude Certified Architect - Professional (CCAR-P) Practice Tests