A company runs real-time inferences with SageMaker. Auto scaling for the EC2 instances backing the endpoint launches new instances before previous scale-out instances are fully ready, causing inefficiencies and delays. Which change will improve scaling behavior while preserving response-time performance?

Increasing the cooldown period after scale-out activities is the correct solution. The cooldown period prevents the auto-scaling policy from launching additional instances too soon after a scale-out event. If new instances aren't fully ready before the auto-scaling policy evaluates again, it can trigger unnecessary further scaling. A longer cooldown period gives newly launched instances sufficient time to initialize and become available, allowing the auto-scaling policy to make more accurate decisions based on the actual capacity. Changing to a multi-model endpoint configuration might improve resource utilization but doesn't directly address the issue of instances not being ready before subsequent scaling decisions. Integrating API Gateway and Lambda adds an orchestration layer but doesn't inherently fix the underlying auto-scaling behavior of SageMaker endpoints. Decreasing the cooldown for scale-in and increasing the max instances would exacerbate the problem by making scaling more aggressive or allowing more instances to launch unnecessarily.