INFRA Signal 470
Multi-cluster Kubernetes orchestration project Karmada reaches CNCF graduation and releases v1.19
↗
The Cloud Native Computing Foundation has moved Karmada to its highest maturity tier alongside the release of v1.19.
Karmada provides a production-ready way to manage workloads across multiple clouds and regions without requiring new resource APIs. This is particularly relevant for AI training and inference where GPU capacity is often fragmented across different providers.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
The v1.19 release introduces enhanced multi-component scheduling for AI training and moves priority-based scheduling to Beta.
Karmada uses PropagationPolicy and OverridePolicy APIs to handle workload distribution and cluster-specific configuration changes.
The project maintains compatibility with standard Kubernetes manifests, controllers, and tools.
THE READ
What the cluster adds up to.
Karmada has reached the CNCF graduation tier, signaling it is stable and ready for production use. This transition follows a third-party security audit and the establishment of a formal steering committee. The project now supports a wide range of production adopters, including Bloomberg and Alibaba Cloud, for tasks like cross-region resilience and fleet-wide configuration.
The system operates via a control plane consisting of an API Server, Controller Manager, and Scheduler, all backed by etcd. It manages member clusters through four internal controllers that handle the lifecycle of clusters, policy binding, and the execution of manifests. Operators can use Helm charts for installation and Prometheus for monitoring the control plane.
Adopting Karmada avoids the need to learn a new federated-resources API, as it builds on the standard Kubernetes API. However, users must manage the additional Karmada control plane components and define custom policies for workload placement. The tool is designed to solve the specific problem of fragmented accelerator capacity in distributed AI workloads.
Written by elseif from the cluster below · checked for specifics the sources never containedTHE CLUSTER