The KubeRay development team announced that the tool has officially joined the Cloud Native Computing Foundation (CNCF) as a Sandbox project, a step that places the open AI tooling ecosystem under a unified and neutral governance umbrella led by the Linux Foundation.
Completion of the Open AI Infrastructure Stack
KubeRay represents the pivotal link in the software stack known as PARK, which brings together PyTorch, AI Models, the Ray framework, and Kubernetes. Today, this ecosystem serves as a de facto standard for running AI workloads at production scale. With PyTorch and Ray already under the supervision of the PyTorch Foundation (part of the Linux Foundation) and Kubernetes within the CNCF, KubeRay's transition achieves fully neutral management for all components of this stack, giving enterprises and infrastructure teams greater confidence in the ecosystem's stability and software neutrality.
Capabilities for Operating and Scaling Ray on Kubernetes
KubeRay acts as a Kubernetes Operator and a toolset designed to run Python- and Ray-based AI and machine learning workloads seamlessly over cloud infrastructure. The project achieves this through three Custom Resource Definitions (CRDs):
- RayCluster: Represents the core unit for managing the complete lifecycle of a Ray cluster on Kubernetes, automating the provisioning of Head and Worker Pods, health checks, Global Control Store (GCS) fault tolerance, and dynamic auto-scaling in coordination with the Ray Autoscaler.
- RayJob: Specifically designed for batch processing, distributed training, and reinforcement learning tasks; it automatically creates a temporary cluster, submits the computing job once ready, and tears down the cluster immediately after execution to maximize the efficiency of GPUs and accelerators.
- RayService: Dedicated to running long-running, high-availability online inference services via Ray Serve, enabling zero-downtime rolling updates, automatic traffic rerouting, and precise model deployment pipeline management.
Expanded Integration with the Cloud Environment
Major platforms like Spotify rely on KubeRay to run their internal Hendrix AI platform and train models via Ray on Google Kubernetes Engine (GKE). In addition to cluster management, KubeRay integrates with monitoring tools such as OpenTelemetry, Prometheus, and Fluent Bit to track accelerator performance and inference latency, alongside batch scheduling and quota management tools like Kueue, Volcano, and KAI to fairly distribute GPU and TPU workloads while supporting advanced isolation and security policies.
What This Transition Means for Developers and Enterprises
KubeRay's joining of the CNCF ensures the project is protected from proprietary vendor lock-in and provides a multi-company collaborative development environment. Technology enterprises benefit from accelerated compatibility between Python-based AI frameworks and container orchestration platforms, reducing the engineering complexity required to manage distributed computing clusters and facilitating the large-scale deployment of AI agents and multimodal applications.