k3s - 5 less than k8s

Lightweight Kubernetes. Easy to install, half the memory, all in a binary less than 40mb.

Great for

• Edge
• IoT
• CI
• ARM
• Situations where a PhD in k8s clusterology is infeasible

What is this?

k3s is intended to be a fully compliant Kubernetes distribution with the following changes:

1. Legacy, alpha, non-default features are removed. Hopefully, you shouldn't notice the stuff that has been removed.
2. Removed most in-tree plugins (cloud providers and storage plugins) which can be replaced with out of tree addons.
3. Add sqlite3 as the default storage mechanism. etcd3 is still available, but not the default.
4. Wrapped in simple launcher that handles a lot of the complexity of TLS and options.
5. Minimal to no OS dependencies (just a sane kernel and cgroup mounts needed). k3s packages required dependencies
   • containerd
   • Flannel
   • CoreDNS
   • CNI
   • Host utilities (iptables, socat, etc)

Quick start

1. Download k3s from latest release, x86_64, armhf, and arm64 are supported
2. Run server

sudo k3s server &
# Kubeconfig is written to /etc/rancher/k3s/k3s.yaml
sudo k3s kubectl get node

# On a different node run the below. NODE_TOKEN comes from /var/lib/rancher/k3s/server/node-token 
# on your server
sudo k3s agent --server https://myserver:6443 --token ${NODE_TOKEN}

Running server

To run the server just do

k3s server

You should get an output similar to

INFO[2019-01-22T15:16:19.908493986-07:00] Starting k3s dev                             
INFO[2019-01-22T15:16:19.908934479-07:00] Running kube-apiserver --allow-privileged=true --authorization-mode Node,RBAC --service-account-signing-key-file /var/lib/rancher/k3s/server/tls/service.key --service-cluster-ip-range 10.43.0.0/16 --advertise-port 6445 --advertise-address 127.0.0.1 --insecure-port 0 --secure-port 6444 --bind-address 127.0.0.1 --tls-cert-file /var/lib/rancher/k3s/server/tls/localhost.crt --tls-private-key-file /var/lib/rancher/k3s/server/tls/localhost.key --service-account-key-file /var/lib/rancher/k3s/server/tls/service.key --service-account-issuer k3s --api-audiences unknown --basic-auth-file /var/lib/rancher/k3s/server/cred/passwd --kubelet-client-certificate /var/lib/rancher/k3s/server/tls/token-node.crt --kubelet-client-key /var/lib/rancher/k3s/server/tls/token-node.key 
Flag --insecure-port has been deprecated, This flag will be removed in a future version.
INFO[2019-01-22T15:16:20.196766005-07:00] Running kube-scheduler --kubeconfig /var/lib/rancher/k3s/server/cred/kubeconfig-system.yaml --port 0 --secure-port 0 --leader-elect=false 
INFO[2019-01-22T15:16:20.196880841-07:00] Running kube-controller-manager --kubeconfig /var/lib/rancher/k3s/server/cred/kubeconfig-system.yaml --service-account-private-key-file /var/lib/rancher/k3s/server/tls/service.key --allocate-node-cidrs --cluster-cidr 10.42.0.0/16 --root-ca-file /var/lib/rancher/k3s/server/tls/token-ca.crt --port 0 --secure-port 0 --leader-elect=false 
Flag --port has been deprecated, see --secure-port instead.
INFO[2019-01-22T15:16:20.273441984-07:00] Listening on :6443                           
INFO[2019-01-22T15:16:20.278383446-07:00] Writing manifest: /var/lib/rancher/k3s/server/manifests/coredns.yaml 
INFO[2019-01-22T15:16:20.474454524-07:00] Node token is available at /var/lib/rancher/k3s/server/node-token 
INFO[2019-01-22T15:16:20.474471391-07:00] To join node to cluster: k3s agent -s https://10.20.0.3:6443 -t ${NODE_TOKEN} 
INFO[2019-01-22T15:16:20.541027133-07:00] Wrote kubeconfig /etc/rancher/k3s/k3s.yaml
INFO[2019-01-22T15:16:20.541049100-07:00] Run: k3s kubectl                             

The output will probably be much longer as the agent will spew a lot of logs. By default the server will register itself as a node (run the agent). It is common and almost required these days that the control plane be part of the cluster. To not run the agent by default use the --disable-agent flag

k3s server --disable-agent

At this point, you can run the agent as a separate process or not run it on this node at all.

Joining nodes

When the server starts it creates a file /var/lib/rancher/k3s/server/node-token. Use the contents of that file as NODE_TOKEN and then run the agent as follows

k3s agent --server https://myserver:6443 --token ${NODE_TOKEN}

That's it.

Accessing cluster from outside

Copy /etc/rancher/k3s/k3s.yaml on your machine located outside the cluster as ~/.kube/config. Then replace "localhost" with the IP or name of your k3s server. kubectl can now manage your k3s cluster.

Auto-deploying manifests

Any file found in /var/lib/rancher/k3s/server/manifests will automatically be deployed to Kubernetes in a manner similar to kubectl apply.

It is also possible to deploy Helm charts. k3s supports a CRD controller for installing charts. A YAML file specification can look as following (example taken from /var/lib/rancher/k3s/server/manifests/traefik.yaml):

apiVersion: k3s.cattle.io/v1
kind: HelmChart
metadata:
  name: traefik
  namespace: kube-system
spec:
  chart: stable/traefik
  set:
    rbac.enabled: "true"
    ssl.enabled: "true"

Building from source

The clone will be much faster on this repo if you do

git clone --depth 1 https://github.com/rancher/k3s.git

This repo includes all of Kubernetes history so --depth 1 will avoid most of that.

For development, you just need go 1.11 and a sane GOPATH. To compile the binaries run

go build -o k3s
go build -o kubectl ./cmd/kubectl
go build -o hyperkube ./vendor/k8s.io/kubernetes/cmd/hyperkube

This will create the main executable, but it does not include the dependencies like containerd, CNI, etc. To run a server and agent with all the dependencies for development run the following helper scripts

# Server
./scripts/dev-server.sh

# Agent
./scripts/dev-agent.sh

To build the full release binary run make and that will create ./dist/artifacts/k3s

Uninstalling server

If you installed your k3s server with the help of install.sh script from the root directory, you may use the uninstall script generated during installation, which will be created on your server node at /usr/local/bin/k3s-uninstall.sh

Kubernetes source

The source code for Kubernetes is in vendor/ and the location from which that is copied is in ./vendor.conf. Go to the referenced repo/tag and you'll find all the patches applied to upstream Kubernetes.

Open ports / Network security

The server needs port 6443 to be accessible by the nodes. The nodes need to be able to reach other nodes over UDP port 4789. This is used for flannel VXLAN. If you don't use flannel and provide your own custom CNI, then 4789 is not needed by k3s. The node should not listen on any other port. k3s uses reverse tunneling such that the nodes make outbound connections to the server and all kubelet traffic runs through that tunnel.

IMPORTANT. The VXLAN port on nodes should not be exposed to the world, it opens up your cluster network to accessed by anyone. Run your nodes behind a firewall/security group that disables access to port 4789.

Server HA

Just don't right now :) It's currently broken.

Running in Docker (and docker-compose)

I wouldn't be me if I couldn't run my cluster in Docker. rancher/k3s images are available to run k3s server and agent from Docker. A docker-compose.yml is in the root of this repo that serves as an example of how to run k3s from Docker. To run from docker-compose from this repo run

docker-compose up --scale node=3
# kubeconfig is written to current dir
kubectl --kubeconfig kubeconfig.yaml get node

NAME           STATUS   ROLES    AGE   VERSION
497278a2d6a2   Ready    <none>   11s   v1.13.2-k3s2
d54c8b17c055   Ready    <none>   11s   v1.13.2-k3s2
db7a5a5a5bdd   Ready    <none>   12s   v1.13.2-k3s2

Hyperkube

k3s is bundled in a nice wrapper to remove the majority of the headache of running k8s. If you don't want that wrapper and just want a smaller k8s distro, the releases includes the hyperkube binary you can use. It's then up to you to know how to use hyperkube. If you want individual binaries you will need to compile them yourself from source

containerd and Docker

k3s includes and defaults to containerd. Why? Because it's just plain better. If you want to run with Docker first stop and think, "Really? Do I really want more headache?" If still yes then you just need to run the agent with the --docker flag

 k3s agent -u ${SERVER_URL} -t ${NODE_TOKEN} --docker &

systemd

If you are bound by the shackles of systemd (as most of us are), there is a sample unit file in the root of this repo k3s.service which is as follows

[Unit]
Description=Lightweight Kubernetes
Documentation=https://k3s.io
After=network.target

[Service]
ExecStartPre=-/sbin/modprobe br_netfilter
ExecStartPre=-/sbin/modprobe overlay
ExecStart=/usr/local/bin/k3s server
KillMode=process
Delegate=yes
LimitNOFILE=infinity
LimitNPROC=infinity
LimitCORE=infinity
TasksMax=infinity

[Install]
WantedBy=multi-user.target

Flannel

Flannel is included by default, if you don't want flannel then run the agent with --no-flannel as follows

 k3s agent -u ${SERVER_URL} -t ${NODE_TOKEN} --no-flannel &

In this setup you will still be required to install your own CNI driver. More info here

CoreDNS

CoreDNS is deployed on start of the agent, to disable add --no-deploy coredns to the server

 k3s server --no-deploy coredns

If you don't install CoreDNS you will need to install a cluster DNS provider yourself.

Traefik

Traefik is deployed by default when starting the server; to disable it, start the server with --no-deploy traefik like this

 k3s server --no-deploy traefik

Service load balancer

k3s includes a basic service load balancer that uses available host ports. If you try to create a load balancer that listens on port 80, for example, it will try to find a free host in the cluster for port 80. If no port is available the load balancer will stay in Pending.

To disable the embedded service load balancer (if you wish to use a different implementation like MetalLB) just add --no-deploy=servicelb to the server on startup.

TODO

Currently broken or stuff that needs to be done for this to be considered production quality.

1. Metrics API due to kube aggregation not being setup
2. HA
3. Work on e2e, sonobouy.
4. etcd doesn't actually work because args aren't exposed