March 2026 performance summary for ray-project/kuberay and vllm-project/tpu-inference. The month focused on reliability, scalability, and distributed execution improvements across Kubernetes-based Ray deployments and TPU-enabled inference pipelines. Key reliability enhancement: RayCluster hash generation now ignores tolerations and scheduling gates, reducing unnecessary updates from external controllers and stabilizing reconciliations. Service platform expansion: Ray Serve gained gRPC support via serve-x ports, with preservation of user-defined protocols during port merging and removal of unnecessary port sorting/filtering to streamline configuration. Distributed inference improvements: Added Multi-Host TPU support for the vLLM Ray executor, enhancing resource management and error handling for TPU nodes in multi-host environments. Overall impact: lower operational churn, broader service capabilities, and improved performance for distributed Ray deployments and TPU-backed inference workloads.

February 2026 — RayProject/kuberay delivered governance and deployment flexibility improvements backed by tests and documentation. Key outcomes include: 1) RayService managedBy field introduced to designate the managing entity (immutable) with tests and docs; commits include ea46c77d3d5fc1e31a4e05f3807ec1f2f76b2bca. 2) BuildPod enhanced to respect explicit plasma-directory, enabling users to skip the shared memory mount; tests validating behavior; commit e201898fe0f04213a041060659fac87b1167a2ef. 3) Associated e2e test fixes and lint issue resolutions to improve reliability and code quality. 4) Documentation updates to reflect new configuration options. Overall impact: improved deployment reliability, clearer ownership, and greater configuration flexibility for multi-tenant scenarios.