Microsoft's AKS team has published guidance for running Anyscale's managed Ray service at scale on Azure Kubernetes Service, addressing three core operational challenges. For GPU scarcity, they recommend a multi-cluster, multi-region setup using Azure Arc to aggregate quota and reroute workloads. For scattered ML storage, they propose Azure BlobFuse2 to mount Blob Storage as a POSIX filesystem into Ray worker pods, enabling shared data access with local caching. For credential expiry, the new approach uses Microsoft Entra service principals with AKS workload identity to issue short-lived tokens automatically, eliminating manual rotation. The integration is currently in private preview. Notably, AWS and Google Cloud have made similar Anyscale partnerships, suggesting the industry is converging on Kubernetes-plus-Ray for AI workloads, with competition shifting to which cloud best streamlines the surrounding infrastructure.
Sort: