Karmada — Multi-Cloud & Multi-Cluster Kubernetes Orchestration

Introduction

Karmada (Kubernetes Armada) was open sourced by Huawei and is now a CNCF project. It lets you run one "host" Kubernetes control plane that schedules workloads onto many "member" clusters — across clouds, regions, or on-prem sites — while each workload stays a normal Deployment, Service, or CRD.

What Karmada Does

• Provides a super-API server that aggregates Pods, Services, and CRDs across clusters.
• Schedules workloads with placement policies (label, region, taint, weight, resource fit).
• Fails over or rebalances Deployments when a cluster goes unhealthy.
• Mirrors ConfigMaps, Secrets, and CRDs to target clusters with override rules.
• Supports pull mode (edge) and push mode (central) for member-cluster control.

Architecture Overview

The Karmada control plane is a Kubernetes itself — etcd, kube-apiserver, scheduler, controller-manager, plus Karmada-specific components (karmada-webhook, karmada-aggregated-apiserver, karmada-scheduler, karmada-controller-manager). Member clusters run a tiny agent (in pull mode) or are reached via kubeconfig (push mode). Propagation policies resolve which clusters get which objects.

Self-Hosting & Configuration

• kubectl karmada init produces a ready control plane on any kubeadm cluster.
• Helm charts and the Karmada Operator cover production deployments.
• Store cluster kubeconfigs as Secrets; rotate automatically with cert-manager.
• Use PropagationPolicy and ClusterPropagationPolicy for namespace vs. cluster scope.
• OverridePolicy rewrites image tags, resources, or args per target cluster.

Key Features

• Native kubectl experience across dozens of clusters.
• Declarative cross-cluster service discovery via MCS (Multi-Cluster Services).
• Differential config per cluster through strategic-merge or JSON-patch overrides.
• Works with any CNCF-conformant Kubernetes — EKS, GKE, AKS, OpenShift, k3s.
• Pluggable scheduling: priority, capacity-aware, CPU/GPU topology-aware.

Comparison with Similar Tools

• Kubefed v2 — archived predecessor; Karmada is considered its spiritual successor.
• Cluster API — cluster lifecycle; Karmada is workload lifecycle across clusters.
• Rancher Fleet — GitOps-first multi-cluster; Karmada is API-first.
• Google Anthos / Azure Arc — managed multi-cluster, vendor-locked.
• OpenShift ACM — Red Hat multi-cluster with a GUI; Karmada is upstream and free.

FAQ

Q: Do pods move between clusters automatically? A: Deployments can reschedule on cluster failure; individual pods do not migrate because Kubernetes pods are not live-migratable.

Q: How is cross-cluster networking solved? A: Karmada handles orchestration; pair it with Submariner, Cilium Cluster Mesh, or Istio multicluster for the data plane.

Q: Can I mix K8s versions? A: Yes, within two minor versions of the control plane per project guidance.

Q: CNCF status? A: Incubating as of 2023 — used in production by Huawei Cloud, ICBC, and others.

Sources

• https://github.com/karmada-io/karmada
• https://karmada.io/docs/