Skip to main content

Access the K3s cluster

Control plane high availability​

To ensure high availability of the Kubernetes control plane, we will set up a load balancer using HAProxy on a separate server (serverLb) that will distribute traffic to the K3s server nodes.

ansible-playbook -i ./ansible/inventory.yaml ./ansible/playbooks/105_install_haproxy.yaml

Get kubeconfig from the server​

Copy kube config from server1. You run this command ONCE during the initial setup.

mkdir -p ~/.kube
scp root@$(pulumi --cwd $PULUMI_CWD stack output server1.Ipv4Address):/etc/rancher/k3s/k3s.yaml ~/.kube/config-pulumi-hetzner-k3s

Configure DNS​

Execute the following command for instructions to point your DNS to the server load balancer IP

echo "Point your $K8S_URL DNS record to $(pulumi --cwd $PULUMI_CWD stack output serverLb.Ipv4Address) IP address"
info

Note that the wildcard DNS record won't work in this case because wildcard DNS points to the main load balancer IP which is used for HTTP/HTTPS traffic. Kube API server uses a different port (6443) which is handled by the server load balancer.

Replace the server url in the kubeconfig file

sed -i "s#https://127.0.0.1:6443#https://${K8S_URL}:6443#g" ~/.kube/config-pulumi-hetzner-k3s

Configure local kubectl​

Export KUBECONFIG env variable to point to the above config file.

export KUBECONFIG=~/.kube/config-pulumi-hetzner-k3s
note

Optionally, you can rename the file to config and place it in ~/.kube/ directory. This way you don't need to export KUBECONFIG env variable.

mv ~/.kube/config-pulumi-hetzner-k3s ~/.kube/config

Verify the cluster is working​

note

If at any point a pod gets stuck in ImagePullBackOff state it is most likely because the server's IP is blacklisted. Try deleting the affected pods and let them try to pull from the internal registry mirror.

warning

Some resources take a while to start. Be patient and wait for the resources to stabilize. As a rule of a thumb you can wait for all pods to become ready by executing the following command by replacing {deployment_namespace} with the namespace of the deployment. kubectl wait --for=condition=Ready pod --all -n {deployment_namespace} --timeout=60s

For example, to wait for all pods in the loki namespace to become ready you would run: kubectl wait --for=condition=Ready pod --all -n loki --timeout=60s

In addition, I recommend having k9s tool where you can easily view all pods in all namespaces or a very popular desktop GUI k8s management app Lens.

Ensure you are pointing to the correct cluster

kubectl get node

The output should be similar to:

NAME      STATUS   ROLES                       AGE    VERSION
agent1    Ready    <none>                      228d   v1.32.3+k3s1
agent2    Ready    <none>                      228d   v1.32.3+k3s1
agent3    Ready    <none>                      228d   v1.32.3+k3s1
server1   Ready    control-plane,etcd,master   228d   v1.32.3+k3s1
server2   Ready    control-plane,etcd,master   228d   v1.32.3+k3s1
server3   Ready    control-plane,etcd,master   228d   v1.32.3+k3s1