KubeRay v1.7 Launches With Upgraded Features for Ray on Kubernetes

KubeRay, the Kubernetes operator for managing Ray distributed workloads, has released versions 1.6 and 1.7, marking significant upgrades for AI, ML, and other scalable computing applications. The releases, which include over 381 commits from approximately 75 contributors, bring enhancements like a beta History Server, safer RayService upgrades, improved RayJob management, and advanced fault tolerance.

Key Upgrades and Features

The standout feature in v1.7 is the beta History Server, which allows users to access job and cluster logs even after a RayCluster is deleted. The revamped server reduces memory usage and startup time with lazy loading and LRU caching, while integrating with Ray’s authentication tokens for improved security. Automated sidecar injections simplify its deployment, removing manual configurations for log collection.

Security also sees a boost with native Kubernetes NetworkPolicy support, automated mTLS certificate management via cert-manager, and token-based Kubernetes RBAC authentication. These updates cater to enterprises prioritizing safer operations in production environments.

RayJob management gets more robust with new features like RayCronJob for native cron scheduling, retry logic for transient failures in sidecar mode, and better deletion policies for managing worker pods and clusters post-job completion.

Another noteworthy addition is GCS fault tolerance through an embedded RocksDB backend, eliminating the need for external Redis dependencies. This feature streamlines cluster deletion and enhances persistence for critical workloads.

RayService and RayCluster Enhancements

RayService, a key resource for deploying Ray Serve applications, now supports incremental upgrades with rollback functionality, graduating the feature to beta. This enables smoother zero-downtime upgrades, critical for production environments. Additionally, new customizable Ingress options and Autoscaler V2’s priority-aware worker group selection improve deployment flexibility and resource allocation.

For RayCluster users, the new Recreate upgrade strategy simplifies updates by automatically deleting and recreating pods when configurations change, ideal for GitOps workflows.

Why This Matters

KubeRay’s updates solidify its position as the go-to toolkit for running Ray workloads on Kubernetes, a critical combination for scaling machine learning and AI workloads. By addressing pain points like job management, security, and fault tolerance, these releases enhance its appeal for enterprise users handling complex distributed computing tasks.

The open-source project’s continued evolution reflects growing demand for Kubernetes-native solutions in AI/ML. Ray is already widely adopted for applications such as model training, batch processing, and inference serving, and these KubeRay updates make it easier to operationalize these workloads at scale.

Looking Ahead

As KubeRay v1.7 rolls out, the focus shifts to the upcoming v1.8 release, with the Ray community actively shaping its roadmap. Users can engage with contributors via Slack, GitHub, and bi-weekly community meetings to influence future developments.

For those ready to adopt, the Helm chart for KubeRay v1.7 is available now, offering streamlined installation for Kubernetes clusters. With enhancements targeting both enterprise-grade reliability and developer efficiency, KubeRay’s trajectory underscores its critical role in the Kubernetes and AI/ML ecosystems.

Image source: Shutterstock