The correct answer is Model inference service.

Running a model inference service inside a container on a managed Kubernetes cluster is a natural fit because inference workloads are typically stateless and require low latency and predictable scaling. Kubernetes provides autoscaling, load balancing, rolling updates and resource isolation that suit serving containerized models in production.

Containers make it easy to package the model code and its dependencies so the same image can be deployed across environments. A managed Kubernetes cluster also allows you to attach GPU nodes when needed and to integrate health checks and observability for continuous serving.

Data preparation is usually a batch oriented ETL step that runs as part of a data pipeline. It is better served by managed data processing tools and pipeline runners that handle long running jobs and large scale transformations.

Model training often requires distributed compute, long running GPU instances and specialized orchestration for checkpoints and hyperparameter tuning. Managed training services or dedicated training clusters are typically a better fit than a production inference deployment.

Data loading is an ingestion or streaming task that belongs in storage and pipeline services. It does not usually match the stateless, low latency serving pattern that you deploy as a containerized service on Kubernetes.

Full AWS Practitioner Certification Question