A company has a trained ML model in Amazon SageMaker that must be hosted in production with high availability, minimal latency, and the ability to handle unpredictable bursts of requests. Each request will be between 1 KB and 3 MB. The inference capacity must scale proportionally with demand. How should the company deploy the model?

A SageMaker real-time inference endpoint is the most suitable choice for this scenario. It is designed for high availability and low-latency inference, directly addressing the requirements. Enabling autoscaling for the endpoint allows it to scale proportionally with unpredictable bursts of requests, ensuring capacity meets demand. SageMaker handles the underlying infrastructure, simplifying deployment and management. Hosting in an Amazon ECS cluster would require more manual setup and management of the ML environment. Installing the SageMaker Operator on EKS is an option, but SageMaker real-time endpoints offer a more managed and integrated solution for SageMaker-trained models. Running inference on Spot Instances is cost-effective but not ideal for high availability and minimal latency due to their ephemeral nature.