Skip to content

Getting started with Kamaji

This guide will lead you through the process of creating a working Kamaji setup on a generic infrastructure.

The material here is relatively dense. We strongly encourage you to dedicate time to walk through these instructions, with a mind to learning. We do NOT provide any "one-click" deployment here. However, once you've understood the components involved it is encouraged that you build suitable, auditable GitOps deployment processes around your final infrastructure.

The guide requires:

  • a bootstrap machine
  • a Kubernetes cluster to run the Admin and Tenant Control Planes
  • an arbitrary number of machines to host Tenants' workloads


Prepare the bootstrap workspace

On the bootstrap machine, clone the repo and prepare the workspace directory:

git clone
cd kamaji/deploy

We assume you have installed on the bootstrap workstation:

Access Management Cluster

In Kamaji, the Management Cluster is a regular Kubernetes cluster which hosts zero to many Tenant Cluster Control Planes. The Management Cluster acts as cockpit for all the Tenant Clusters as it hosts monitoring, logging, and governance of Kamaji setup, including all Tenant Clusters.

Throughout the following instructions, shell variables are used to indicate values that you should adjust to your environment:

source kamaji.env

Any regular and conformant Kubernetes v1.22+ cluster can be turned into a Kamaji setup. To work properly, the Management Clusterr should provide:

  • CNI module installed, eg. Calico, Cilium.
  • CSI module installed with a Storage Class for the Tenant datastores. Local Persistent Volumes are an option.
  • Support for LoadBalancer service type, eg. MetalLB, or a Cloud based controller.
  • Optionally, a Monitoring Stack installed, eg. Prometheus.

Make sure you have a kubeconfig file with admin permissions on the cluster you want to turn into Kamaji Management Cluster and check you can access:

kubectl cluster-info

Install Cert Manager

Kamaji takes advantage of the dynamic admission control, such as validating and mutating webhook configurations. These webhooks are secured by a TLS communication, and the certificates are managed by cert-manager, making it a prerequisite that must be installed:

helm repo add jetstack
helm repo update
helm install \
  cert-manager jetstack/cert-manager \
  --namespace cert-manager \
  --create-namespace \
  --version v1.11.0 \
  --set installCRDs=true

Install Kamaji Controller

Installing Kamaji via Helm charts is the preferred way. The Kamaji controller needs to access a Datastore in order to save data of the tenants' clusters. The Kamaji Helm Chart provides the installation of a basic unmanaged etcd as datastore, out of box.

Install Kamaji with helm using an unmanaged etcd as default datastore:

helm repo add clastix
helm repo update
helm install kamaji clastix/kamaji -n kamaji-system --create-namespace

A managed datastore is highly recommended in production

The kamaji-etcd project provides the code to setup a multi-tenant etcd running as StatefulSet made of three replicas. Optionally, Kamaji offers support for a more robust storage system, as MySQL or PostgreSQL compatible database, thanks to the native kine integration.

Now you should end up with a working Kamaji instance, including the default datastore:

kubectl -n kamaji-system get pods
NAME                         READY   STATUS      RESTARTS      AGE
etcd-0                       1/1     Running     0             50s
etcd-1                       1/1     Running     0             60s
etcd-2                       1/1     Running     0             90s
kamaji-7949578bfb-lj44p      1/1     Running     0             12s

An unsuccessful first installation could fail for several reasons, such as missing a StorageClass, or even for a trivial Ctrl+C during the installation phase.

See the Cleanup section before to retry an aborted installation.

Create Tenant Cluster

Tenant Control Plane

A tenant control plane of example looks like:

kind: TenantControlPlane
  name: ${TENANT_NAME}
  namespace: ${TENANT_NAMESPACE}
  labels: ${TENANT_NAME}
  dataStore: default
      replicas: 3
        apiServer: []
        controllerManager: []
        scheduler: []
            cpu: 250m
            memory: 512Mi
          limits: {}
            cpu: 125m
            memory: 256Mi
          limits: {}
            cpu: 125m
            memory: 256Mi
          limits: {}
      serviceType: LoadBalancer
    version: ${TENANT_VERSION}
      cgroupfs: systemd
      - ResourceQuota
      - LimitRanger
    port: ${TENANT_PORT}
    serviceCidr: ${TENANT_SVC_CIDR}
    podCidr: ${TENANT_POD_CIDR}
    coreDNS: {}
    kubeProxy: {}
        port: ${TENANT_PROXY_PORT}
            cpu: 100m
            memory: 128Mi
          limits: {}

kubectl -n ${TENANT_NAMESPACE} apply -f ${TENANT_NAMESPACE}-${TENANT_NAME}-tcp.yaml

After a few seconds, check the created resources in the tenants namespace and when ready it will look similar to the following:

kubectl -n ${TENANT_NAMESPACE} get tcp,deploy,pods,svc

NAME                           VERSION   STATUS   CONTROL-PLANE ENDPOINT   KUBECONFIG                   DATASTORE   AGE
tenantcontrolplane/tenant-00   v1.25.2   Ready      tenant-00-admin-kubeconfig   default     2m20s

NAME                        READY   UP-TO-DATE   AVAILABLE   AGE
deployment.apps/tenant-00   3/3     3            3           118s

NAME                             READY   STATUS    RESTARTS   AGE
pod/tenant-00-58847c8cdd-7hc4n   4/4     Running   0          82s
pod/tenant-00-58847c8cdd-ft5xt   4/4     Running   0          82s
pod/tenant-00-58847c8cdd-shc7t   4/4     Running   0          82s

NAME                TYPE           CLUSTER-IP      EXTERNAL-IP      PORT(S)                         AGE
service/tenant-00   LoadBalancer   6443:32152/TCP,8132:32713/TCP   2m20s

The regular Tenant Control Plane containers: kube-apiserver, kube-controller-manager, kube-scheduler are running unchanged in the tcp pods instead of dedicated machines and they are exposed through a service on the port 6443 of worker nodes in the Management Cluster.

The LoadBalancer service type is used to expose the Tenant Control Plane on the assigned loadBalancerIP acting as ControlPlaneEndpoint for the worker nodes and other clients as, for example, kubectl. Service types NodePort and ClusterIP are still viable options to expose the Tenant Control Plane, depending on the case. High Availability and rolling updates of the Tenant Control Planes are provided by the tcp Deployment and all the resources reconcilied by the Kamaji controller.

Working with Tenant Control Plane

Collect the external IP address of the tcp service:

TENANT_ADDR=$(kubectl -n ${TENANT_NAMESPACE} get svc ${TENANT_NAME} -o json | jq -r ."spec.loadBalancerIP")

and check it out:

curl -k https://${TENANT_ADDR}:${TENANT_PORT}/healthz
curl -k https://${TENANT_ADDR}:${TENANT_PORT}/version

The kubeconfig required to access the Tenant Control Plane is stored in a secret:

kubectl get secrets -n ${TENANT_NAMESPACE} ${TENANT_NAME}-admin-kubeconfig -o json \
  | jq -r '.data["admin.conf"]' \
  | base64 --decode \

and let's check it out:

kubectl --kubeconfig=${TENANT_NAMESPACE}-${TENANT_NAME}.kubeconfig cluster-info

Kubernetes control plane is running at
CoreDNS is running at

Check out how the Tenant Control Plane advertises itself to workloads:

kubectl --kubeconfig=${TENANT_NAMESPACE}-${TENANT_NAME}.kubeconfig get svc

default       kubernetes   ClusterIP    <none>        443/TCP   6m
kubectl --kubeconfig=${TENANT_NAMESPACE}-${TENANT_NAME}.kubeconfig get ep

NAME         ENDPOINTS             AGE
kubernetes   18m

And make sure it is ${TENANT_ADDR}:${TENANT_PORT}.

Join worker nodes

The Tenant Control Plane is made of pods running in the Kamaji Management Cluster. At this point, the Tenant Cluster has no worker nodes. So, the next step is to join some worker nodes to the Tenant Control Plane.

Kamaji does not provide any helper for creation of tenant worker nodes, instead it leverages the Cluster Management API. This allows you to create the Tenant Clusters, including worker nodes, in a completely declarative way. Refer to the Cluster API guide to learn more about supported providers.

An alternative approach for joining nodes is to use the kubeadm command on each node. Follow the related documentation in order to:

  • install containerd as container runtime
  • install crictl, the command line for working with containerd
  • install kubectl, kubelet, and kubeadm in the desired version

After the installation is complete on all the nodes, open the command line on your Linux workstation and store the IP address of each node in an environment variable:

WORKER0=<address of first node>
WORKER1=<address of second node>
WORKER2=<address of third node>

Store the join command in a variable:

JOIN_CMD=$(echo "sudo ")$(kubeadm --kubeconfig=${TENANT_NAMESPACE}-${TENANT_NAME}.kubeconfig token create --print-join-command)

Use a loop to log in to and run the join command on each node:

for i in "${!HOSTS[@]}"; do
  ssh ${USER}@${HOST} -t ${JOIN_CMD};


This manual process can be further automated to handle the node prerequisites and joining. See yaki script, which you could modify for your preferred operating system and version. The provided script is just a facility: it assumes all worker nodes are running Ubuntu 22.04. Make sure to adapt the script if you're using a different distribution.

Checking the nodes:

kubectl --kubeconfig=${TENANT_NAMESPACE}-${TENANT_NAME}.kubeconfig get nodes 

NAME                  STATUS     ROLES    AGE   VERSION
tenant-00-worker-00   NotReady   <none>   25s   v1.25.0
tenant-00-worker-01   NotReady   <none>   17s   v1.25.0
tenant-00-worker-02   NotReady   <none>   9s    v1.25.0

The cluster needs a CNI plugin to get the nodes ready. In this guide, we are going to install calico, but feel free to use one of your taste.

Download the latest stable Calico manifest:

curl -O

Before to apply the Calico manifest, you can customize it as necessary according to your preferences.

Apply to the Tenant Cluster:

kubectl --kubeconfig=${TENANT_NAMESPACE}-${TENANT_NAME}.kubeconfig apply -f calico.yaml

And after a while, nodes will be ready

kubectl --kubeconfig=${TENANT_NAMESPACE}-${TENANT_NAME}.kubeconfig get nodes 
NAME                  STATUS   ROLES    AGE     VERSION
tenant-00-worker-00   Ready    <none>   2m48s   v1.25.0
tenant-00-worker-01   Ready    <none>   2m40s   v1.25.0
tenant-00-worker-02   Ready    <none>   2m32s   v1.25.0


Delete a Tenant Cluster

First, remove the worker nodes joined the tenant control plane

kubectl --kubeconfig=${TENANT_NAMESPACE}-${TENANT_NAME}.kubeconfig delete nodes --all 

For each worker node, login and clean it

for i in "${!HOSTS[@]}"; do
  ssh ${USER}@${HOST} -t 'sudo kubeadm reset -f';
  ssh ${USER}@${HOST} -t 'sudo rm -rf /etc/cni/net.d';
  ssh ${USER}@${HOST} -t 'sudo systemctl reboot';

Delete the tenant control plane from Kamaji

kubectl delete -f ${TENANT_NAMESPACE}-${TENANT_NAME}-tcp.yaml

Uninstall Kamaji

Uninstall the Kamaji controller by removing the Helm release

helm uninstall kamaji -n kamaji-system

The default datastore installed three etcd replicas with persistent volumes, so remove the PersistentVolumeClaims resources:

kubectl -n kamaji-system delete pvc --all

Also delete the custom resources:

kubectl delete crd
kubectl delete crd

In case of a broken installation, manually remove the hooks installed by Kamaji:

kubectl delete ValidatingWebhookConfiguration kamaji-validating-webhook-configuration
kubectl delete MutatingWebhookConfiguration kamaji-mutating-webhook-configuration

That's all folks!