Autoscaling
Why might CPU-based autoscaling fail for an API service?
Whether the candidate can identify what actually saturates first, and whether they treat autoscaling as a control loop with lag rather than as elasticity.
The situation behind the question
An API scales on 70% average CPU with a minimum of two instances. During the morning peak, p99 latency goes from 120ms to 4s, the pool of database connections is exhausted, and the autoscaler adds no instances the entire time because CPU never exceeds 45%.