AIMentor — AI Expertise Learning Platform

13 questions · pass mark 80% · 20 minute limit. The timer auto-submits at zero.

Time remaining: 20:00

Question 1easy
What is a message queue in an AI pipeline?
A vector databaseA durable to-do list between services: a producer adds jobs, a worker takes them one at a timeA type of promptA rate limit
Question 2medium
What do message queues buy an AI workload? (Select all that apply)
Decoupling producers from consumers (either can restart without losing work)Buffering bursts so a spike drains at a sustainable rateSmoothing traffic to stay within provider rate limitsMaking the model more accurate
Select all that apply.
Question 3medium
In the producer/consumer pattern, where does the slow LLM call happen?
In the producer, before ackingIn the worker/consumer, off the request path, after the producer has already ackedIn the queue itselfIt doesn't happen at all
Question 4medium
Why use exponential backoff when retrying a flaky LLM call?
To retry faster each timeWaiting progressively longer avoids hammering a struggling service and gives transient issues time to clearTo use more tokensBecause the API requires it
Question 5hard
Which errors should you retry, and which should you fail fast on?
Retry everything, including 400sRetry transient errors (429, timeout, 5xx); fail fast on permanent ones (400, invalid input)Never retry anythingRetry only 400s
Question 6hard
Why add jitter (randomness) to retry backoff across many workers?
To make logs prettierSo a fleet of workers doesn't retry in lockstep, creating a synchronized thundering herdTo increase latencyTo disable retries
Question 7hard
What is a dead-letter queue (DLQ)?
A queue for high-priority messagesWhere a message goes after exhausting its retries — parked for inspection instead of retrying foreverA backup of the vector indexA queue that deletes all messages
Question 8medium
Why must a dead-lettered message also be removed (acked) from the main queue?
To save tokensSo the poison message stops blocking and the rest of the queue keeps flowingBecause the DLQ requires itIt doesn't need removing
Question 9hard
Most queues guarantee 'at-least-once' delivery. What does that imply for your workers?
Messages are never duplicatedA message may be delivered more than once (e.g. a crash before ack), so workers must be idempotentMessages are always lostYou must disable retries
Question 10hard
How do you achieve exactly-once EFFECT despite at-least-once DELIVERY?
By using a faster queueIdempotent workers that dedupe on the message id, so processing a redelivered message twice is harmlessBy never retryingBy using a bigger model
Question 11medium
Which are event-driven pipeline pitfalls? (Select all that apply)
Doing slow LLM work inline on the request/webhook pathRetrying permanent errors like a 400Having no dead-letter queue for poison messagesMaking workers idempotent
Select all that apply.
Question 12medium
What is the queue (name or acronym) called where messages go after exhausting their retries?
Question 13hard
What retry strategy (two words) waits progressively longer (1,2,4,8s) between attempts?

0 / 13 answered

Event-Driven Workflows — Final Exam