CoreWeave adds first-party Anyscale support for managed Ray on CKS

What CoreWeave and Anyscale are offering

Ray is an open-source distributed framework for parallelizing Python and AI applications. CoreWeave describes it as a Python-native framework that supports unstructured data formats and orchestration over heterogeneous clusters, meaning those with both CPUs and GPUs. It is meant to hide the complexity of distributed AI workloads such as training, fine-tuning, inference and multimodal data processing.

Anyscale was developed by the original creators of Ray. Per the announcement, it offers developer tooling, resilient clusters, proprietary performance improvements called RayTurbo, and built-in governance controls, all built for Ray.

By running fully managed Ray clusters in their own account, CoreWeave customers keep full control of their data in CoreWeave AI Object Storage while using a wide range of accelerated compute options, according to CoreWeave. The company also points to the Weights & Biases integration with Anyscale via Ray, which it says adds scalable experiment tracking and artifact management.

Reliability and observability claims

CoreWeave lists several benefits of running Anyscale workloads on its clusters:

• Cluster validation: a validation suite checks GPUs, CPUs, memory, storage and networking, and verifies functional health so the cluster is ready for large-scale, production-grade workloads at delivery.
• Proactive health checking: automated checks monitor Ray nodes and remediate the source of anomalous behavior before it affects workloads.
• Observability: metrics range from GPU temperatures to real-time ingress and egress traffic, with dashboards that cover entire fleets and help customers spot under-optimized workloads and tie job interruptions to underlying issues.
• Performance: CoreWeave says the combination allows up to 96% goodput.

Anyscale is also described as a fully managed platform with access control, single sign-on with SAML and quota management, which CoreWeave calls critical to scaling Ray beyond a few clusters.

Why not self-managed Ray on CKS

CoreWeave notes that Ray and KubeRay, the open-source Kubernetes operator, can help teams start with a first Ray workload, but call them only a starting foundation. To reach a production-ready platform, teams still need to build developer tooling, observability, cost controls, job queues and user management, according to the announcement.

Anyscale is positioned as the shortcut. The listed components are:

• Anyscale Workspaces: an interactive dev console with workload observability, letting developers self-serve infrastructure without Kubernetes expertise.
• Fully Managed Ray Clusters: unhealthy node draining and replacement, fast autoscaling, and monitoring with managed Prometheus and Grafana dashboards.
• RayTurbo: optimizations not available in the open-source distribution, including 5x faster data preprocessing with RayTurbo Data, plus RayTurbo Train and RayTurbo Serve libraries.
• Budget and cost dashboards: budgets at organization or project level, with alerts when limits are exceeded.
• SLA-backed support: guaranteed response times from Ray experts with paid support plans.

As an example, CoreWeave describes reinforcement learning with the RayTurbo distribution of RLlib. Developers can use Python syntax and a notebook-like IDE to run jobs that scale up to 64 GPUs and 1000's of simulations, including on RL frameworks not natively supported in the open-source distribution, such as BC-IRL and APPO with proprietary scale-out enhancements.

What it means for customers

For CoreWeave customers already running AI workloads on CKS, the announcement means a managed Ray option that runs inside their own account and keeps data in CoreWeave AI Object Storage. Teams that would otherwise assemble tooling around open-source Ray and KubeRay can use Anyscale's managed components instead.

To get started, whether for a multimodal data pipeline, model training or large-scale inference, customers should contact their CoreWeave account manager or write to [email protected] to arrange a demo or proof-of-concept engagement. The source does not give pricing.