An ML engineer configures auto scaling for an inference component behind an Amazon SageMaker endpoint with a target-tracking policy of 100 invocations per model per minute. The endpoint scales normally during business hours, but at each business day start there are zero instances available to handle requests, causing delays. How should the engineer ensure the endpoint can handle incoming requests at the start of each business day?

The correct solution addresses the problem of zero instances at the start of the business day. A step scaling policy driven by a CloudWatch alarm allows for proactive scaling based on a schedule or predicted load. By adding a second CloudWatch alarm and scaling policy, the minimum instance count can be increased from zero to one (or more) at the start of each business day, ensuring instances are ready to handle requests. Reducing the cooldown period or adding a lifecycle hook won't provision instances before requests arrive. Changing the target metric to CPU utilization might improve scaling during active periods but won't solve the cold start issue if instances are scaled down to zero. Setting the scaling policy target value to one is too aggressive and would lead to over-provisioning and increased costs, as it would try to maintain a single invocation per minute, not a minimum instance count.