Karmada Federated Control Plane for Kubernetes Achieves CNCF Graduation

TL;DR — Key Takeaways

– Karmada has graduated from the CNCF, signaling its technical maturity and readiness for enterprise production environments.

– Karmada provides a unified control plane for multiple Kubernetes clusters, enabling centralized workload placement, failover, autoscaling and disaster recovery across cloud environments.

– Version 1.19 expands support for AI workloads, introducing enhancements to scheduling and resource management for distributed AI training and inference.

Karmada, the control plane for managing multiple Kubernetes clusters, has graduated from the Cloud Native Computing Foundation’s technology program, offering the CNCF’s stamp-of-approval for production use just in time for the AI inferencing work it needs to support these days.  

The latest release, version 1.19, expands support for AI workloads with enhancements to multi-cluster scheduling and resource management.

The graduation was announced at the KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026. 

The CNCF graduation indicates that the software has reached technical maturity that makes it suitable for enterprise usage. Karmada’s security was audited by a third-party firm. It established a formal steering committee to ensure transparent governance, as well as a set of best practices.  

Controlling Multiple Clusters from the Same Control Plane

When Huawei first open sourced the federated control plane software in 2021, the primary use case for Karmada was the ability to control multiple Kubernetes clusters across different environments, for purposes of multi-region resilience and disaster recovery.

Karmada is short for “Kubernetes Armada,” armada being a word that means a fleet of warships. Karmada does indeed turn multiple Kubernetes clusters into a coordinated fleet. 

Karmada exposes a standard Kubernetes API and offers centralized placement, propagation, failover, and multi-cluster autoscaling. (It is unrelated to the Armada project – also built on Kubernetes – which is used for managing batch workloads.)

It can be integrated into CNCF observability and deployment projects, exporting Prometheus metrics, leverage etcd to track control-plane state, and using Helm charts for installation. 

Overall, the project has attracted 1,214 contributors across 292 contributing organizations, and its users have given it more than 5,600 GitHub stars.

Use Cases

Why does one need Karmada when Kubernetes itself can be used to orchestrate containers? Kubernetes is largely designed to control a single cluster. Most organizations run more than one cluster, and often run multiple clusters across different cloud providers. Karmada provides a unified control plane to manage them all, using the standard Kubernetes API. 

Organizations such as Bloomberg, Wellhub, Alibaba Cloud, Huawei, and Trip.com all use the technology. The software is widely used across Chinese cloud, internet, telecom, AI, travel, logistics, device, and enterprise organizations. 

For Shanghai-based cloud native software distributor DaoCloud, Karmada provides a way to give its own customers multi-cloud deployments using a consistent Kubernetes interface. The software provides a way for Trip.com to operate multiple clusters as a unified resource pool, easily bringing new clusters and migrating workloads across different clusters. 

“By automating disaster recovery, improving resource utilization, and simplifying the management of individual Kubernetes clusters, [Karmada] has enabled our platform engineering teams to operate more efficiently, while also giving internal application teams a simpler, more consistent experience deploying jobs,” noted Karmada maintainer Michas Szacillo, who is the engineering team lead for Bloomberg’s streaming platform. 

Looking Ahead

Looking ahead, the project will expand its scope to become more of a resource-aware control plane for heterogeneous computational infrastructure. It will include features such as priority-based preemption, multi-cluster queuing, and support for Dynamic Resource Allocation (DRA) across GPUs and other accelerators.