The Cloud Native Computing Foundation (CNCF) announced on September 2026 that Karmada, a multi-cluster and multi-cloud Kubernetes orchestration project, has graduated. This multi-cluster and multi-cloud Kubernetes orchestration project reached CNCF's highest maturity tier. This tier is for stable projects that are widely adopted and ready for production. The announcement happened at KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026 in Shanghai. It coincided with the project's v1.19 release. This update enhances multi-component scheduling for AI training jobs. It also promotes priority-based scheduling to Beta, which is now on by default.
Running one app across multiple Kubernetes clusters is common for hybrid cloud use, regional failover, or avoiding vendor lock-in. This often required custom automation or the now-archived KubeFed project. This approach required learning a different federated-resources API and faced issues with flexibility. As GPU capacity gets spread out across regions and cloud providers, teams doing distributed AI training and inference face a big challenge. No single cluster has enough accelerators. So, workloads need to be split, scheduled, and shifted across many systems.
Karmada, short for "Kubernetes Armada," builds on the standard Kubernetes API. It doesn’t replace it, so existing manifests, controllers, and tools work without changes on a Karmada control plane. That control plane consists of three components: a Karmada API Server, a Karmada Controller Manager, and a Karmada Scheduler, backed by its own etcd instance for state. Placement logic is expressed through two custom APIs: a PropagationPolicy, which maps a policy to a set of workloads (1:n) and defines scheduling and spreading constraints such as cluster affinity, multi-cluster splitting/rebalancing, and multi-dimension high availability across region, availability zone, cluster, or provider; and an OverridePolicy, which lets operators rewrite cluster-specific configuration, for example, swapping container image prefixes by region or StorageClass by cloud provider, without touching the underlying resource template.
Internally, four controllers do the propagation work: a Cluster Controller manages the lifecycle of registered member clusters; a Policy Controller watches PropagationPolicy objects and binds matching resources into ResourceBinding objects; a Binding Controller turns each ResourceBinding into per-cluster Work objects; and an Execution Controller watches those Work objects and pushes the resulting manifests to each member cluster's own API server. Karmada also exports Prometheus metrics from its control-plane components and ships Helm charts for installation, plugging directly into existing CNCF observability and deployment tooling rather than requiring a separate stack.
Since joining the CNCF Sandbox in September 2021 and moving to Incubating in December 2023, Karmada has expanded to over 1,214 contributors from 292 organizations. It also boasts more than 5,600 GitHub stars. Its production adopter list spans Bloomberg, Wellhub, Alibaba Cloud, Huawei, Trip.com, Bilibili, iFLYTEK, JDCloud, Kuaishou, RedNote, SenseTime, Vivo, WPS, and ZTO, using it for hybrid cloud capacity, cross-region resilience, GPU/CPU scheduling for AI workloads, and fleet-wide configuration distribution. To graduate, the project did a third-party security audit. It also formed a formal steering committee, adopted the CNCF Code of Conduct, and keeps a CII Best Practices Badge.
As organizations scale beyond a single Kubernetes cluster, they need consistent management without added complexity. Karmada solves this by extending familiar Kubernetes APIs to work across clusters and clouds.
said Chad Beaudin, the project's TOC Sponsor. Honghui Yue, Senior Development Expert at Trip.com, said:
At Trip.com , Karmada has become a critical part of our multi-cluster infrastructure and has delivered significant value in production. Without changing existing Kubernetes resource definitions, it has enabled us to operate multiple clusters as a unified resource pool, support cross-cluster elasticity and failover, bring new clusters into production more efficiently, and perform large-scale workload migration with minimal disruption to applications.
Karmada is one of two CNCF projects that replaced the retired KubeFed (Federation v2). The other project is Open Cluster Management (OCM). OCM is still in Sandbox. It uses a hub-and-spoke, agent-based method. This method focuses on cluster inventory and add-on APIs. OCM supports Red Hat Advanced Cluster Management. OCM focuses on governance and policy distribution. Karmada, on the other hand, emphasizes workload placement and dynamic scheduling. It also includes caching layers for cross-cluster resource queries and multi-cluster service discovery. It exists within a larger context that features Cluster API, focusing on cluster lifecycle instead of workload placement. It also includes GitOps tools like Argo CD's ApplicationSets and Rancher Fleet. Additionally, there's Microsoft's KubeFleet, which forms the foundation for Azure Kubernetes Fleet Manager.