In a standard Docker Kubernetes cluster, kubelet is running on each node as systemd
service and is taking care of communication between runtime and the API service.
It is responsible for starting microservices pods (such as kube-proxy
, kubedns
,
etc. - they can be different for various ways of deploying k8s) and user pods.
Configuration of kubelet determines which runtime is used and in what way.
Kubelet itself is executed in Docker container (as we can see in kubelet.service
),
but, what is important, it's not a Kubernetes pod (at least for now),
so we can keep kubelet running inside the container (as well as directly on the host),
and regardless of this, run pods in the chosen runtime.
Below, you can find an instruction how to switch one or more nodes on running Kubernetes cluster from Docker to CRI-O.
At first, you need to stop kubelet service working on the node:
systemctl stop kubelet
and stop all kubelet Docker containers that are still running.
docker stop $(docker ps | grep k8s_ | awk '{print $1}')
We have to be sure that kubelet.service
will start after crio.service
.
It can be done by adding crio.service
to Wants=
section in /etc/systemd/system/kubelet.service
:
$ cat /etc/systemd/system/kubelet.service | grep Wants
Wants=docker.socket crio.service
If you'd like to change the way of starting kubelet (e.g., directly on the host instead of in a container), you can change it here, but, as mentioned, it's not necessary.
Kubelet parameters are stored in /etc/kubernetes/kubelet.env
file.
$ cat /etc/kubernetes/kubelet.env | grep KUBELET_ARGS
KUBELET_ARGS="--pod-manifest-path=/etc/kubernetes/manifests
--pod-infra-container-image=gcr.io/google_containers/pause-amd64:3.0
--cluster_dns=10.233.0.3 --cluster_domain=cluster.local
--resolv-conf=/etc/resolv.conf --kubeconfig=/etc/kubernetes/node-kubeconfig.yaml
--require-kubeconfig"
You need to add following parameters to KUBELET_ARGS
:
--container-runtime-endpoint=unix:///var/run/crio/crio.sock
Socket for remote runtime (default crio
socket localization).
--runtime-request-timeout=10m
- Optional but useful. Some requests, especially pulling huge images, may take longer than default (2 minutes) and will cause an error.
You may need to add following parameter to KUBELET_ARGS
(prior to Kubernetes
v1.24.0-alpha.2
). This flag is deprecated since v1.24.0-alpha.2
, and will no
longer be available starting from v1.27.0
:
--container-runtime=remote
- Use remote runtime with provided socket.
Kubelet is prepared now.
If your cluster is using flannel network, your network configuration should be like:
$ cat /etc/cni/net.d/10-crio.conf
{
"name": "crio",
"type": "flannel"
}
Then, kubelet will take parameters from /run/flannel/subnet.env
- file generated
by flannel kubelet microservice.
Start crio first, then kubelet. If you created crio
service:
systemctl start crio
systemctl start kubelet
You can follow the progress of preparing node using kubectl get nodes
or
kubectl get pods --all-namespaces
on Kubernetes control-plane.