Karmada Multi-Cluster Kubernetes Orchestration Project Achieves CNCF Graduation

Karmada Reaches CNCF Graduation

The Cloud Native Computing Foundation (CNCF) announced the graduation of Karmada, a multi-cluster and multi-cloud Kubernetes orchestration project, on September 2026. This achievement places Karmada in CNCF's highest maturity tier, indicating its stability, widespread adoption, and readiness for production environments. The announcement was made at KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026 in Shanghai.

Key Enhancements in v1.19 Release

Coinciding with its graduation, Karmada released version 1.19. This update introduces enhancements for multi-component scheduling, particularly benefiting AI training jobs. Additionally, priority-based scheduling has been promoted to Beta status and is now enabled by default, improving workload management.

Addressing Multi-Cluster Challenges

Running applications across multiple Kubernetes clusters is a common practice for hybrid cloud deployments, regional failover strategies, and avoiding vendor lock-in. Previously, this often required custom automation or relied on projects like the now-archived KubeFed, which introduced different federated-resources APIs and lacked flexibility. For distributed AI training and inference, where GPU capacity is spread across regions and cloud providers, Karmada helps manage workloads that need to be split, scheduled, and shifted across numerous systems due to the lack of sufficient accelerators in a single cluster.

Karmada's Architecture and Functionality

Karmada, short for 'Kubernetes Armada,' extends the standard Kubernetes API rather than replacing it. This design allows existing Kubernetes manifests, controllers, and tools to function without modification on a Karmada control plane. The control plane comprises a Karmada API Server, a Karmada Controller Manager, and a Karmada Scheduler, all backed by an etcd instance for state management.

Placement logic in Karmada is defined through two custom APIs: PropagationPolicy and OverridePolicy. PropagationPolicy maps policies to workloads, setting scheduling and spreading constraints like cluster affinity, multi-cluster splitting, rebalancing, and multi-dimension high availability across regions, availability zones, clusters, or providers. OverridePolicy enables operators to rewrite cluster-specific configurations, such as swapping container image prefixes by region or StorageClass by cloud provider, without altering the underlying resource template.