KubeRay Joins CNCF, Bringing the PARK AI Stack Under Linux Foundation Governance

KubeRay connects Ray workloads with Kubernetes

KubeRay is an open source toolkit and Kubernetes operator for running distributed AI, machine learning and Python workloads powered by Ray on cloud-native infrastructure. Anyscale says it bridges Python-first AI frameworks with Kubernetes-native infrastructure, as distributed AI workloads become a core cloud-native construct.

The project provides three Kubernetes Custom Resource Definitions. RayCluster manages the lifecycle of Ray clusters, including head and worker pod provisioning, health checks, Global Control Service fault tolerance and dynamic autoscaling with the Ray Autoscaler.

RayJob is designed for batch processing, distributed training and reinforcement learning. It provisions an ephemeral RayCluster, submits a Ray job when the cluster is ready and removes the cluster after completion. RayService supports long-lived, high-availability online inference with Ray Serve, including rolling upgrades, automated failover and traffic routing.

The donation adds the missing PARK component

Anyscale says PyTorch and Ray already operate through the PyTorch Foundation, while Kubernetes is part of the CNCF. KubeRay was the remaining piece connecting Ray with Kubernetes. By joining the CNCF, the project brings the full PARK stack under Linux Foundation governance, which Anyscale says gives teams confidence in its neutrality and openness.

KubeRay is becoming a CNCF sandbox project. The company says the project has also gained adoption through integrations with the wider cloud-native ecosystem, rather than recreating existing infrastructure capabilities.

For observability, KubeRay works with OpenTelemetry, Prometheus and Fluent Bit. For scheduling and quota management, it integrates with Kueue, Volcano and KAI, supporting gang scheduling, multi-tenant resource quotas, priority preemption and topology-aware scheduling. The project also supports automated mTLS and Kubernetes network policies for workload security and isolation.

Community and production use will shape the next phase

Anyscale says Spotify uses KubeRay to scale Hendrix, its internal machine learning platform, for training and fine-tuning models with Ray on Google Kubernetes Engine. The company also highlights Hark's use of KubeRay for multimodal AI workloads and agentic workflows.

The announcement describes CNCF sandbox status as the beginning of KubeRay's next chapter under open, vendor-neutral governance. Maintainers are inviting users and contributors to discuss the roadmap, production use cases and Ray on Kubernetes at a Birds of a Feather session during KubeCon North America 2026. The community is also available through the KubeRay channel in CNCF or Ray Slack workspaces and biweekly community meetings.