Kuis Arsitektur AI Cloud

Jawaban dan penjelasan muncul saat Anda mengikuti kuis.

1. Why separate interactive and background AI workloads?
   • Mereka selalu menggunakan model yang berbeda-beda
   • Background work needs no monitoring
   • They can have different response-time and processing requirements
   • Interactive work cannot use queues
2. What is the purpose of backpressure in an overloaded service?
   • Limit incoming work so a struggling dependency does not overwhelm the system
   • Increase every retry rate
   • Hide the queue from operators
   • Make a slow provider respond faster
3. Why is network location alone insufficient to establish authorization?
   • Anything inside a network is automatically trusted
   • Cloud services cannot enforce access control
   • Network boundaries replace secret management
   • Access roles and credentials still determine which data or actions are permitted
4. A provider has a rate limit and model loading is slow. What should architecture plan for?
   • Unlimited retries from every worker
   • Bounded retries, admission control, capacity behavior, and clear unavailable states
   • A larger prompt
   • Automatic acceptance of stale outputs
5. A service scales web workers, but every worker calls the same limited inference endpoint. What may happen?
   • More workers can create a longer queue without increasing model capacity
   • Inference capacity necessarily scales with web workers
   • The model stops using memory
   • User authorization is automatically improved
6. What should cost per useful task include?
   • Only the provider’s listed input-token price
   • Only GPU electricity
   • Storage, transfer, failed attempts, and idle resources in addition to API cost
   • The server’s initial purchase price
7. Why use clear queue status and retry behavior for background jobs?
   • They guarantee every job succeeds on the first try
   • They make long-running work and recovery observable
   • They remove the need for durable state
   • They reduce model output length
8. What does a rollback test need to consider?
   • Only the web page title
   • Whether the new deployment has a shorter name
   • The model’s training benchmark alone
   • Compatible model, preprocessing, and data-migration versions