Differences between "plain" Kubernetes pods and VM pods
Virtlet tries hard to make VM pods appear as plain Kubernetes pods. Still, there are some important differences, including:
- VM pods can have just one "container"
- You can't use container images for VM pods
- Some container-specific settings such as network/PID/IPC namespaces, SELinux/AppArmor settings, privileged flag etc. aren't applicable to VM pods
- Some volume types are handled differently. There are VM-pod-specific settings such as Cloud-Init, persistent rootfs, etc. For more information, see Volumes.
kubectl execand exec readiness/liveness probes aren't supported for VM pods yet- There are VM-pod-specific settings such as Cloud-Init, persistent rootfs, etc.
Another important point is that when using a persistent root filesystem, the lifetime of the VM is not limited to that of the pod.
Despite these differences, there are quite a few Kubernetes features
that work for VM pods just as well as for "plain" pods, for example,
most kubectl commands, pointing services at VM pods, and so on.
Besides kubectl, a Virtlet-specific tool called
virtletctl can be used to perform Virtlet-specific
actions on VM pods and Virtlet processes in the cluster, such as
connecting to the VM using SSH and providing VNC connection and
dumping cluster diagnostic info.
Supported kubectl commands
Most kubectl commands' behavior doesn't differ between "plain" and
VM pods. Exceptions are kubectl exec which isn't supported at the
moment, and kubectl attach / kubectl logs which work only if the
VM has serial console configured.
kubectl attach attaches to the VM serial console. Detaching from the
console is done via Ctrl-].
kubectl logs displays the logs for the pod. In case of VM pod, the
log is the serial console output. kubectl logs -f, which follows the
log as it grows, is supported, too.
Using higher-level Kubernetes objects
One of the advantages of pod-based approach to running VMs on Kubernetes is an ability to use higher-level Kubernetes objects such as StatefulSets, Deployments, DaemonSets etc. trivially with VMs.
Virtlet includes a nested Kubernetes example which makes a nested Kubernetes cluster using a StatefulSet of 3 VM pods, which are initialized using kubeadm.
In the k8s-in-k8s example, first we create a
headless service
that will point domain names k8s-0, k8s-1 and k8s-2 as resolved
by cluster's DNS to the corresponding StatefulSet replicas:
apiVersion: v1
kind: Service
metadata:
name: k8s
labels:
app: k8s
spec:
ports:
- port: 22
name: ssh
clusterIP: None
selector:
app: inner-k8s
Then, we begin defining a StatefulSet with 3 replicas:
apiVersion: apps/v1
kind: StatefulSet
metadata:
name: k8s
spec:
serviceName: k8s
replicas: 3
selector:
matchLabels:
app: inner-k8s
The pods that comprise the StatefulSet are VM pods and thus
must have kubernetes.io/target-runtime: virtlet.cloud annotation.
Also, we set the
root volume size to 4Gi to have some place for Docker images:
template:
metadata:
labels:
app: inner-k8s
annotations:
kubernetes.io/target-runtime: virtlet.cloud
# set root volume size
VirtletRootVolumeSize: 4Gi
Then we add another annotation that will contain
Cloud-Init user-data, in which we write some files
to adjust Docker settings and add the Kubernetes repository for apt,
as well as a provisioning script that will install the necessary
packages and then run kubeadm init on k8s-0 and kubeadm join on
k8s-1 and k8s-2. The script makes use of StatefulSet's
stable network IDs,
so the nodes have names k8s-0, k8s-1 and k8s-2 that can be
resolved by cluster DNS.
VirtletCloudInitUserData: |
write_files:
- path: /etc/systemd/system/docker.service.d/env.conf
permissions: "0644"
owner: root
content: |
[Service]
Environment="DOCKER_OPTS=--storage-driver=overlay2"
- path: /etc/apt/sources.list.d/kubernetes.list
permissions: "0644"
owner: root
content: |
deb http://apt.kubernetes.io/ kubernetes-xenial main
- path: /usr/local/bin/provision.sh
permissions: "0755"
owner: root
content: |
#!/bin/bash
set -u -e
set -o pipefail
curl -s https://packages.cloud.google.com/apt/doc/apt-key.gpg | apt-key add -
apt-get update
apt-get install -y docker.io kubelet kubeadm kubectl kubernetes-cni
sed -i 's/--cluster-dns=10\.96\.0\.10/--cluster-dns=10.97.0.10/' \
/etc/systemd/system/kubelet.service.d/10-kubeadm.conf
systemctl daemon-reload
if [[ $(hostname) =~ -0$ ]]; then
# master node
kubeadm init --token adcb82.4eae29627dc4c5a6 \
--pod-network-cidr=10.200.0.0/16 \
--service-cidr=10.97.0.0/16 \
--apiserver-cert-extra-sans=127.0.0.1,localhost
export KUBECONFIG=/etc/kubernetes/admin.conf
export kubever=$(kubectl version | base64 | tr -d '\n')
kubectl apply -f "https://cloud.weave.works/k8s/net?k8s-version=$kubever"
while ! kubectl get pods -n kube-system -l k8s-app=kube-dns |
grep ' 1/1'; do
sleep 1
done
mkdir -p /root/.kube
chmod 700 /root/.kube
cp "${KUBECONFIG}" /root/.kube/config
echo "Master setup complete." >&2
else
# worker node
kubeadm join --token adcb82.4eae29627dc4c5a6 \
--discovery-token-unsafe-skip-ca-verification k8s-0.k8s:6443
echo "Node setup complete." >&2
fi
We then add an ssh public key (corresponding to examples/vmkey) for
the user root:
users:
- name: root
# VirtletSSHKeys only affects 'ubuntu' user for this image, but we want root access
ssh-authorized-keys:
- ssh-rsa AAAAB3NzaC1yc2EAAAADAQABAAABAQCaJEcFDXEK2ZbX0ZLS1EIYFZRbDAcRfuVjpstSc0De8+sV1aiu+dePxdkuDRwqFtCyk6dEZkssjOkBXtri00MECLkir6FcH3kKOJtbJ6vy3uaJc9w1ERo+wyl6SkAh/+JTJkp7QRXj8oylW5E20LsbnA/dIwWzAF51PPwF7A7FtNg9DnwPqMkxFo1Th/buOMKbP5ZA1mmNNtmzbMpMfJATvVyiv3ccsSJKOiyQr6UG+j7sc/7jMVz5Xk34Vd0l8GwcB0334MchHckmqDB142h/NCWTr8oLakDNvkfC1YneAfAO41hDkUbxPtVBG5M/o7P4fxoqiHEX+ZLfRxDtHB53 [email protected]
and make Cloud-Init run the provisioning script when the VM boots:
runcmd:
- /usr/local/bin/provision.sh
After that, we get to the pod spec, in which we limit the
possibilities of running the pod to the nodes with
extraRuntime=virtlet label where Virtlet daemon pods are placed and
use Ubuntu 16.04 image:
spec:
nodeSelector:
extraRuntime: virtlet
containers:
- name: ubuntu-vm
image: virtlet.cloud/cloud-images.ubuntu.com/xenial/current/xenial-server-cloudimg-amd64-disk1.img
imagePullPolicy: IfNotPresent
# tty and stdin required for `kubectl attach -t` to work
tty: true
stdin: true
We then specify a readiness probe which will mark each pod as Ready
once its ssh port becomes active:
readinessProbe:
tcpSocket:
port: 22
initialDelaySeconds: 5
We could use a more sophisticated check, e.g. by making sure that
apiserver is accessible, but for the purpose of the example we use
this trivial scheme to keep things simple. Note that kubeadm join on
k8s-1 and k8s-2 will keep retrying till kubeadm init on k8s-0
completes its task.
In order to test the example, we start it and wait till k8s-0,
k8s-1 and k8s-2 pods appear:
$ kubectl apply -f examples/k8s.yaml
$ kubectl get pods -w
Then we can view the logs on each of the node to see the progress of Kubernetes setup, e.g.
$ kubectl logs -f k8s-0
...
[ 226.115652] cloud-init[1513]: Master setup complete.
...
After the setup is complete, we can use virtletctl to ssh into the master VM and check the cluster:
$ virtletctl ssh [email protected] -- -i examples/vmkey
Welcome to Ubuntu 16.04.5 LTS (GNU/Linux 4.4.0-138-generic x86_64)
...
[email protected]:~# kubectl get pods --all-namespaces
NAMESPACE NAME READY STATUS RESTARTS AGE
kube-system coredns-576cbf47c7-8dnh6 1/1 Running 0 69m
kube-system coredns-576cbf47c7-q622f 1/1 Running 0 69m
kube-system etcd-k8s-0 1/1 Running 0 69m
kube-system kube-apiserver-k8s-0 1/1 Running 0 68m
kube-system kube-controller-manager-k8s-0 1/1 Running 0 68m
kube-system kube-proxy-4dgfx 1/1 Running 0 69m
kube-system kube-proxy-jmw6c 1/1 Running 0 69m
kube-system kube-proxy-qwbw7 1/1 Running 0 69m
kube-system kube-scheduler-k8s-0 1/1 Running 0 68m
kube-system weave-net-88jv4 2/2 Running 1 69m
kube-system weave-net-kz698 2/2 Running 0 69m
kube-system weave-net-rbnmf 2/2 Running 1 69m
[email protected]:~# kubectl get nodes
NAME STATUS ROLES AGE VERSION
kube-master Ready master 29m v1.14.1
kube-node-1 Ready <none> 28m v1.14.1
kube-node-2 Ready <none> 28m v1.14.1