Skip to content
Permalink
Branch: master
Find file Copy path
104 lines (80 sloc) 4.56 KB

Running CRI-O on kubernetes cluster

Switching runtime from docker to CRI-O

In standard docker kubernetes cluster, kubelet is running on each node as systemd service and is taking care of communication between runtime and api service. It is reponsible 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 container (as well as directly on the host), and regardless of this, run pods in chosen runtime.

Below, you can find an instruction how to switch one or more nodes on running kubernetes cluster from docker to CRI-O.

Preparing crio

You must prepare and install crio on each node you would like to switch. Besides the files installed by make install install.config, here's the list of files that must be provided:

File path Description Location
/etc/containers/policy.json containers policy Example stored in cri-o repository
/bin/runc runc or other OCI runtime Can be build from sources opencontainers/runc
/opt/cni/bin/{flannel, bridge,...} CNI plugins binaries Can be built from sources containernetworking/plugins
/etc/cni/net.d/... CNI network config Example here

crio binary can be executed directly on host, inside the container or in any way. However, recommended way is to set it as a systemd service. Here's the example of unit file:

# cat /etc/systemd/system/crio.service
[Unit]
Description=CRI-O daemon
Documentation=https://github.com/kubernetes-sigs/cri-o

[Service]
ExecStart=/bin/crio --runtime /bin/runc --log /root/crio.log --log-level debug
Restart=always
RestartSec=10s

[Install]
WantedBy=multi-user.target

Preparing kubelet

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 host instead of docker 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:

  • --experimental-cri=true - Use Container Runtime Interface. Will be true by default from kubernetes 1.6 release.
  • --container-runtime=remote - Use remote runtime with provided socket.
  • --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.

Kubelet is prepared now.

Flannel network

If your cluster is using flannel network, your network configuration should be like:

# cat /etc/cni/net.d/10-mynet.conf
{
    "name": "mynet",
    "type": "flannel"
}

Then, kubelet will take parameters from /run/flannel/subnet.env - file generated by flannel kubelet microservice.

Starting kubelet with CRI-O

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 master.

You can’t perform that action at this time.
You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session.