kubeadm

Run kubeadm in HA mode with Ignite VMs

This short guide shows you how to setup Kubernetes in HA mode with Ignite VMs.

NOTE: At the moment, you need to execute all these commands as root.

NOTE: It is assumed that you start no new VMs between running prepare.sh and starting the masters, as IP addresses are computed consecutively

First set up some files and certificates using prepare.sh from this directory:

./prepare.sh

This will create a kubeadm configuration file, generate the CA cert, give you a kubeconfig file, etc.

Start the seed master

For the bootstap master, copy over the CA cert and key to use, and the kubeadm config file:

ignite run weaveworks/ignite-kubeadm:latest \
    --cpus 2 \
    --memory 1GB \
    --ssh \
    --copy-files $(pwd)/run/config.yaml:/kubeadm.yaml \
    --copy-files $(pwd)/run/pki/ca.crt:/etc/kubernetes/pki/ca.crt \
    --copy-files $(pwd)/run/pki/ca.key:/etc/kubernetes/pki/ca.key \
    --name master-0

# Get the IP address of the initial master, for the kubeadm join command below
export MASTER_IP=$($ignite inspect vm master-0 | jq -r ".status.ipAddresses[0]")

Initialize it with kubeadm using ignite exec:

ignite exec master-0 -- kubeadm init --config /kubeadm.yaml --upload-certs

Join additional masters

Create more master VMs, but copy only the variables we need for joining:

for i in {1..2}; do
    ignite run weaveworks/ignite-kubeadm:latest \
        --cpus 2 \
        --memory 1GB \
        --ssh \
        --name master-${i}
done

Use ignite exec to join each VM to the control plane:

source run/k8s-vars.sh
for i in {1..2}; do
    ignite exec master-${i} -- kubeadm join ${MASTER_IP}.xip.io:6443 \
        --token ${TOKEN} \
        --discovery-token-ca-cert-hash sha256:${CA_HASH} \
        --certificate-key ${CERT_KEY} \
        --control-plane
done

Set up a HAProxy loadbalancer locally

docker run -d -v $(pwd)/run/haproxy.cfg:/usr/local/etc/haproxy/haproxy.cfg -p 6443:443 haproxy:alpine

Use kubectl

This will make kubectl talk to any of the three masters you've set up, via HAproxy.

export KUBECONFIG=$(pwd)/run/admin.conf

kubectl get nodes

Right now it's expected that the nodes are in state NotReady, as CNI networking isn't set up.

Install a CNI Network -- Weave Net

We're going to use Weave Net.

kubectl apply -f "https://cloud.weave.works/k8s/net?k8s-version=$(kubectl version | base64 | tr -d '\n')"

With this, the nodes should transition into the Ready state in a minute or so.

Watch the cluster heal

Kill the bootstrap master and see the cluster recover:

ignite rm -f master-0

kubectl get nodes

What's happening underneath here is that HAproxy (or any other loadbalancer) notices that master-0 is unhealthy, and removes it from the roundrobin list. etcd also realizes that one peer is lost, and re-elects a leader amongst the two that are still standing.