5 Questions to Test Your Implementation and Integration Knowledge (AIP-C01)
Practice Implementation and Integration for the AWS Generative AI Developer – Professional (AIP-C01) exam with the 5 questions below. It carries about 26% of the exam score. Every option is explained, so a wrong pick still teaches something.
AWS Generative AI Developer – Professional (AIP-C01) · Implementation and Integration · 5 questions with answers and explanations
🧭 Reading Implementation and Integration questions
Implementation scenarios put a GenAI feature into production: agents calling tools, partners calling APIs, models that throttle or slow down, and new code that must ship safely. You choose the deployment, integration or resilience pattern. The best answers limit damage and cost together, so watch for options that retry forever, pay for idle capacity or roll changes out all at once.
Three things to look for
• When a dependency keeps failing, prefer bounded retries with backoff and a way to stop calling it, rather than longer or endless retries.
• Match the deployment to the traffic: steady, predictable load and intermittent load favor different options, and idle cost often decides between them.
• If a smaller specialized model meets both the accuracy and latency targets, serving it usually beats tuning the price of a much larger one.
📊 This set at a glance
Exam: AIP-C01 · 75 questions · 180 minutes · passing score 750 · the second largest of 5 domains
This set: 3 advanced, 2 applied · 1 comparison, 1 concept check, 1 scenario, 1 trade-off, 1 troubleshooting
Topics: Amazon API Gateway, AWS Step Functions, AWS Lambda, Amazon DynamoDB, Amazon CloudWatch, AWS CodeDeploy, Amazon Bedrock, Amazon Bedrock Model Distillation
Question 1 of 5
A company exposes a GenAI feature to partners and must limit each partner to 10,000 requests per day. Which feature fits?
Show answer
✅ Answer: D
Difficulty: applied · Topic: Amazon API Gateway
Per-client quotas with API Gateway usage plans
Amazon API Gateway REST APIs can meter clients with API keys and usage plans:
• Each partner gets its own API key, and methods are configured to require a key.
• A usage plan attached to the API stage sets throttling (steady rate and burst in requests per second) and a quota (maximum requests per day, week or month).
• Example: a plan with a quota of 10,000 requests per DAY; once a partner's key reaches it, further calls are rejected with HTTP 429 until the period resets.
• Usage per key can be viewed or exported for reporting.
Quotas and throttling are applied on a best-effort basis, and API keys identify callers rather than authenticate them, so combine them with an authorizer (IAM, Amazon Cognito or a Lambda authorizer).
Why each option works or fails
• ✗ A. Lambda counting in memory: Lambda execution environments are created and discarded independently, so in-memory counters are not shared across concurrent instances and are lost when environments end.
• ✗ B. S3 bucket policy by IP: bucket policies control access to S3 objects; they cannot count or limit calls to a GenAI API.
• ✗ C. CloudWatch dashboard: a dashboard only displays request volumes; it does not stop a partner once the limit is reached.
• ✓ D. Usage plans with per-partner API keys (correct): a usage plan enforces a daily request quota for each key, which is exactly a 10,000-requests-per-day limit per partner without custom code.
🎯 Key idea: Enforce per-partner request quotas with API Gateway usage plans and API keys; dashboards and in-memory counters cannot enforce limits.
Question 2 of 5
An agentic workflow orchestrated by AWS Step Functions repeatedly invokes a downstream FM endpoint that has started throttling. The state machine's naive Retry configuration keeps retrying indefinitely, driving up cost without resolving the underlying issue. What should be added to make this workflow safeguarded against this failure mode?
Show answer
✅ Answer: D
Difficulty: advanced · Topic: AWS Step Functions, AWS Lambda, Amazon DynamoDB, Amazon CloudWatch
Retries alone can make an outage worse
A Step Functions Retry block retries a failed state with IntervalSeconds, BackoffRate and MaxAttempts (plus optional MaxDelaySeconds and JitterStrategy). A Catch block routes the error to a fallback state once retries are exhausted. If retries are effectively unlimited, every execution keeps calling a throttled endpoint, adding load and cost while the endpoint cannot recover.
The circuit breaker pattern
A circuit breaker stops calling a dependency that keeps failing:
• Closed: calls go through; failures are counted, for example in an Amazon DynamoDB item keyed by endpoint.
• Open: after a threshold (say 5 failures in 60 seconds), the workflow skips the call and returns a fallback or fails fast.
• Half-open: after a cool-down, one trial call checks whether the endpoint has recovered.
In Step Functions, a first state reads the breaker status from DynamoDB, a Choice state skips the model call when the circuit is open, and bounded Retry plus Catch updates the counter.
Why each option works or fails
• ✗ A. Unlimited MaxAttempts: this is the current failure mode made explicit; endless retries keep adding load and cost and never address the throttling.
• ✗ B. Disabling CloudWatch alarms: hiding the alarms removes visibility into the problem while the runaway retries and costs continue unchanged.
• ✗ C. More Lambda memory: throttling is a quota limit on the model endpoint, so a faster Lambda function does not reduce rejections; it may even send retries sooner.
• ✓ D. Circuit breaker plus bounded Retry/Catch (correct): capping attempts with backoff and tripping a shared breaker after a failure threshold stops all executions from hammering the throttled endpoint, controlling cost until it recovers.
🎯 Key idea: Safeguard agent workflows with bounded retries and a circuit breaker that stops calling a failing dependency, not with longer or unlimited retries.
Question 3 of 5
A team wants to reduce the risk of deploying a new version of a Lambda function that a live production agent depends on for a critical tool call. Which deployment strategy best minimizes the blast radius of a bad release?
Show answer
✅ Answer: D
Difficulty: applied · Topic: AWS CodeDeploy, AWS Lambda, Amazon CloudWatch
Gradual deployment for Lambda
A Lambda alias (for example live) points to a published version. An alias can split traffic between two versions by weight.
AWS CodeDeploy automates that shift:
• Canary: move a small share first, then the rest, for example CodeDeployDefault.LambdaCanary10Percent5Minutes sends 10 percent for 5 minutes, then 100 percent.
• Linear: move in equal steps, for example CodeDeployDefault.LambdaLinear10PercentEvery1Minute.
• All-at-once: move 100 percent immediately.
CodeDeploy watches CloudWatch alarms, for example on the function's Errors metric or the tool's failure rate. If an alarm fires during the shift, it moves the alias back to the old version automatically. Optional pre-traffic and post-traffic hook functions can run tests.
Example
A bad release that fails 30 percent of calls affects only 10 percent of traffic for a few minutes before rollback, instead of every agent request.
Why each option works or fails
• ✗ A. All-at-once: every request hits the new version immediately, so a bad release affects all agent tool calls until someone notices and reverts.
• ✗ B. Console code edits: changes are untracked, go straight to all traffic, and reverting depends on someone remembering the previous code.
• ✗ C. Manual alias switch later: keeping versions separate is good, but the eventual switch is still all-or-nothing with no automatic rollback when errors rise.
• ✓ D. CodeDeploy canary or linear with alarm rollback (correct): only a small share of traffic reaches the new version at first, and a failing release is rolled back automatically, which minimizes the blast radius.
🎯 Key idea: For critical Lambda functions, shift traffic gradually with CodeDeploy and let CloudWatch alarms roll back automatically.
Question 4 of 5
A production chatbot's primary Bedrock model occasionally experiences elevated latency during regional capacity events. The team wants improved resilience but has a limited budget for additional Provisioned Throughput. What is the best design?
Show answer
✅ Answer: A
Difficulty: advanced · Topic: Amazon Bedrock
Fallback as a resilience pattern
A fallback is a second path used only when the primary fails. For model calls it can be:
• a secondary model, for example a different model that meets the quality bar;
• a secondary Region or wider capacity pool, for example calling the model through a cross-Region inference profile, which lets Bedrock serve requests from any Region in the profile's set to absorb traffic bursts.
How to implement it
• Set a client timeout on the primary call that matches the user experience, for example 8 seconds.
• On timeout, throttling or server error, call the fallback once.
• Log and emit a metric for each fallback, so the team sees how often it happens and can alarm on it.
• Test fallback output quality in advance, since answers may differ slightly.
Because the fallback runs on demand only when needed, it adds almost no cost on normal days, unlike reserving extra Provisioned Throughput.
Why each option works or fails
• ✓ A. Fallback on failure with logging (correct): requests that time out or fail are retried on a secondary model or Region, users still get answers during capacity events, and the team pays for the fallback only when it is used.
• ✗ B. Reduce alerting: fewer alerts hide the problem but do nothing for users whose requests are slow or failing.
• ✗ C. Fallback first on every call: it pays for the secondary model on every request, gives users the secondary model's quality most of the time, and still has no plan when that model fails.
• ✗ D. Friendly error message: users still fail to get answers during every capacity event; it improves wording, not resilience.
🎯 Key idea: Use a secondary model or Region only when the primary fails, and log each fallback, to gain resilience without paying for idle capacity.
Question 5 of 5
A marketplace classifies 5 million short product reviews per day into five sentiment classes. Each classification must return within 300 ms to support live moderation. A large general-purpose model reaches 94% accuracy. A smaller model fine-tuned (or distilled) on labeled reviews reaches 93% at about 1/15 of the cost. The business requires at least 92% accuracy. Which approach is BEST?
Show answer
✅ Answer: D
Difficulty: advanced · Topic: Amazon Bedrock, Amazon Bedrock Model Distillation
Choosing a model by requirements, not by size
Start from the hard requirements, then minimize cost among the options that meet them:
• Accuracy floor: at least 92%.
• Latency: under 300 ms per review, in real time.
• Volume: 5 million requests per day, so per-request price dominates.
A smaller model fine-tuned on labeled reviews, or distilled from a large teacher model (Amazon Bedrock Model Distillation does this), learns a narrow task well and is typically faster and far cheaper per request.
Example: at 1/15 of the cost, a daily bill of $1,500 on the large model would be about $100 on the smaller one, and 93% accuracy still clears the 92% bar.
Why each option works or fails
• ✗ A. Overnight batch inference: batch results arrive hours later, which breaks the 300 ms live moderation requirement.
• ✗ B. Provisioned Throughput for the large model: reserved capacity can steady throughput and change pricing, but it keeps paying for a much larger model than the task needs and will not approach a 15x saving.
• ✗ C. Prompt caching on instructions: caching reduces the cost of the repeated prompt prefix, but each short review still needs a full large-model inference and output, so the savings are far smaller.
• ✓ D. Smaller fine-tuned or distilled model (correct): it meets the 92% floor and the 300 ms target at about one fifteenth of the cost; the extra 1% of the large model is not required.
🎯 Key idea: When a smaller specialized model meets the accuracy and latency bars, serve it; optimizing the price of an oversized model saves much less.
🔗 Related practice sets
• AWS Generative AI Developer – Professional (AIP-C01): 6 Practice Questions on FM Integration, Data, and Compliance
• All AIP-C01 practice sets
• Every exam on this blog
📚 More AIP-C01 practice
More AIP-C01 questions: https://hyunhp.github.io/cert-bank/#exam=AIP-C01
Study guide: https://hyunhp.github.io/cert-bank/study.html#exam=AIP-C01
These are original practice questions, not real exam questions, and cert-bank is not affiliated with any certification vendor.
Original practice questions written for this blog, not exam dumps.
Comments
Post a Comment