Cloud-agnostic architecture keeps application logic independent from a particular cloud provider. It favors foundation services such as virtual machines and storage, then layers portable services above them. This approach supports customer-data requirements, provider migration, and control over infrastructure costs. section 3.5.2
A practical reference architecture separates provider-specific infrastructure from portable service logic:
- Use cloud virtual machines for computing resources and storage for durable data. section 3.5.2
- Run Kubernetes on the provider infrastructure and expose training or optimization workflows through Kubernetes interfaces. section 3.5.2 section 5.3.1
- Keep application and workflow logic above the cluster boundary so the same service can target different managed Kubernetes clusters. section 3.5.2
AWS-to-GCP training migration
Provision compute with Amazon EC2, run an Amazon EKS cluster, and build a training service against Kubernetes interfaces. Migration to Google Cloud replaces EKS with Google Kubernetes Engine while most of the training service remains unchanged. The same pattern supports HPO workloads across AWS, Google Cloud Platform, and Azure. section 3.5.2 section 5.3.1
