You need to host a trained model that receives an uneven request rate during the day. The hosting must scale to peak demand but minimize cost when idle. Which hosting approach meets these constraints?

Deploying the model to an Amazon SageMaker endpoint with autoscaling policies is the most suitable solution. SageMaker endpoints are designed for hosting models and can automatically adjust the number of instances based on CloudWatch metrics like CPU utilization or invocation count. This ensures the endpoint scales up to handle peak demand efficiently and scales down during idle periods, minimizing costs. Incorrect options: AWS Lambda with fixed concurrency won't scale automatically based on model invocation load, leading to potential performance issues or over-provisioning. Deploying to Amazon ECS on Fargate with a static number of tasks sized for peak load will incur unnecessary costs during idle times as resources remain provisioned. Hosting multiple copies behind an ALB without autoscaling still requires manual scaling or a separate autoscaling mechanism, which SageMaker endpoints provide natively and more integrated for ML models.