PRODUCTION SCENARIO
A real-time SageMaker endpoint serves an interactive recommendation feature. Request volume triples during evening peaks and falls off overnight, but the peak does not start at the same hour each day. The team must hold p99 latency inside its SLA without paying for peak capacity overnight.
Which auto scaling configuration is MOST appropriate?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: A target tracking policy on InvocationsPerInstance
InvocationsPerInstance is the per-instance request rate, so target tracking on it adds and removes capacity as traffic moves, which holds latency without paying for the peak overnight. CPUUtilization can stay flat while requests queue behind a saturated model. A scheduled action assumes the peak arrives on a clock, and ModelLatency step scaling reacts only once the SLA is already being missed.