KubeRay joins CNCF, completing Linux Foundation governance for PARK AI stack

The Cloud Native Computing Foundation (CNCF) has accepted KubeRay as a sandbox project, filling the final governance gap in the emerging PARK stack for production AI. KubeRay is a Kubernetes operator that bridges Ray's distributed Python-first AI framework with Kubernetes orchestration, providing three CRDs: RayCluster for lifecycle management, RayJob for ephemeral batch workloads, and RayService for long-lived inference with Ray Serve.

The donation completes the Linux Foundation umbrella over all core PARK components: PyTorch and Ray through the PyTorch Foundation, Kubernetes through CNCF, and now KubeRay. This gives enterprises vendor-neutral assurance for building ML platforms, with production users like Spotify already using KubeRay to scale training and fine-tuning on Google Kubernetes Engine.

KubeRay architecture and CNCF integration

• RayCluster CRD: Manages full lifecycle of Ray clusters on Kubernetes, automating head/worker pod provisioning, health checks, Global Control Service fault tolerance, and dynamic autoscaling via the Ray Autoscaler.
• RayJob CRD: Provisions ephemeral RayClusters for batch and RL workloads, submits jobs automatically, and tears down clusters post-completion to maximize accelerator utilization.
• RayService CRD: Manages Ray Serve deployment graphs with zero-downtime rolling upgrades, automated failover, and traffic routing for production inference services.
• Observability integration: Native OpenTelemetry metrics pipeline with Prometheus exposition, plus Fluent Bit log aggregation across ephemeral and persistent clusters.
• Scheduler ecosystem: Native integration with Kueue, Volcano, and KAI for gang scheduling, multi-tenant GPU/TPU quotas, priority preemption, and topology-aware placement.

Why it matters

For platform engineers running AI infrastructure, KubeRay's CNCF status provides long-term governance stability comparable to Kubernetes itself, reducing vendor lock-in risk for organizations betting on Ray. The project's deep integration with established CNCF observability and scheduling tools means teams can adopt it without replacing existing operational stacks, accelerating standardization around a single open-source reference architecture for distributed AI training and serving.