This document describes how to build a high-availability (HA) Kubernetes cluster. This is a fairly advanced topic. Users who merely want to experiment with Kubernetes are encouraged to use configurations that are simpler to set up such as the simple Docker based single node cluster instructions, or try Google Container Engine for hosted Kubernetes.
Also, at this time high availability support for Kubernetes is not continuously tested in our end-to-end (e2e) testing. We will be working to add this continuous testing, but for now the single-node master installations are more heavily tested.
Setting up a truly reliable, highly available distributed system requires a number of steps, it is akin to wearing underwear, pants, a belt, suspenders, another pair of underwear, and another pair of pants. We go into each of these steps in detail, but a summary is given here to help guide and orient the user.
The steps involved are as follows:
Here’s what the system should look like when it’s finished:
The remainder of this guide assumes that you are setting up a 3-node clustered master, where each machine is running some flavor of Linux. Examples in the guide are given for Debian distributions, but they should be easily adaptable to other distributions. Likewise, this set up should work whether you are running in a public or private cloud provider, or if you are running on bare metal.
The easiest way to implement an HA Kubernetes cluster is to start with an existing single-master cluster. The instructions at https://get.k8s.io describe easy installation for single-master clusters on a variety of platforms.
On each master node, we are going to run a number of processes that implement the Kubernetes API. The first step in making these reliable is
to make sure that each automatically restarts when it fails. To achieve this, we need to install a process watcher. We choose to use
the kubelet
that we run on each of the worker nodes. This is convenient, since we can use containers to distribute our binaries, we can
establish resource limits, and introspect the resource usage of each daemon. Of course, we also need something to monitor the kubelet
itself (insert who watches the watcher jokes here). For Debian systems, we choose monit, but there are a number of alternate
choices. For example, on systemd-based systems (e.g. RHEL, CentOS), you can run ‘systemctl enable kubelet’.
If you are extending from a standard Kubernetes installation, the kubelet
binary should already be present on your system. You can run
which kubelet
to determine if the binary is in fact installed. If it is not installed,
you should install the kubelet binary, the
kubelet init file and default-kubelet
scripts.
If you are using monit, you should also install the monit daemon (apt-get install monit
) and the monit-kubelet and
monit-docker configs.
On systemd systems you systemctl enable kubelet
and systemctl enable docker
.
The central foundation of a highly available solution is a redundant, reliable storage layer. The number one rule of high-availability is to protect the data. Whatever else happens, whatever catches on fire, if you have the data, you can rebuild. If you lose the data, you’re done.
Clustered etcd already replicates your storage to all master instances in your cluster. This means that to lose data, all three nodes would need to have their physical (or virtual) disks fail at the same time. The probability that this occurs is relatively low, so for many people running a replicated etcd cluster is likely reliable enough. You can add additional reliability by increasing the size of the cluster from three to five nodes. If that is still insufficient, you can add even more redundancy to your storage layer.
The full details of clustering etcd are beyond the scope of this document, lots of details are given on the etcd clustering page. This example walks through a simple cluster set up, using etcd’s built in discovery to build our cluster.
First, hit the etcd discovery service to create a new token:
curl https://discovery.etcd.io/new?size=3
On each node, copy the etcd.yaml file into /etc/kubernetes/manifests/etcd.yaml
The kubelet on each node actively monitors the contents of that directory, and it will create an instance of the etcd
server from the definition of the pod specified in etcd.yaml
.
Note that in etcd.yaml
you should substitute the token URL you got above for ${DISCOVERY_TOKEN}
on all three machines,
and you should substitute a different name (e.g. node-1
) for ${NODE_NAME} and the correct IP address
for ${NODE_IP}
on each machine.
Once you copy this into all three nodes, you should have a clustered etcd set up. You can validate with
etcdctl member list
and
etcdctl cluster-health
You can also validate that this is working with etcdctl set foo bar
on one node, and etcd get foo
on a different node.
Of course, if you are interested in increased data reliability, there are further options which makes the place where etcd installs it’s data even more reliable than regular disks (belts and suspenders, ftw!).
If you use a cloud provider, then they usually provide this for you, for example Persistent Disk on the Google Cloud Platform. These are block-device persistent storage that can be mounted onto your virtual machine. Other cloud providers provide similar solutions.
If you are running on physical machines, you can also use network attached redundant storage using an iSCSI or NFS interface. Alternatively, you can run a clustered file system like Gluster or Ceph. Finally, you can also run a RAID array on each physical machine.
Regardless of how you choose to implement it, if you chose to use one of these options, you should make sure that your storage is mounted
to each machine. If your storage is shared between the three masters in your cluster, you should create a different directory on the storage
for each node. Throughout these instructions, we assume that this storage is mounted to your machine in /var/etcd/data
Once you have replicated etcd set up correctly, we will also install the apiserver using the kubelet.
First you need to create the initial log file, so that Docker mounts a file instead of a directory:
touch /var/log/kube-apiserver.log
Next, you need to create a /srv/kubernetes/
directory on each node. This directory includes:
The easiest way to create this directory, may be to copy it from the master node of a working cluster, or you can manually generate these files yourself.
Once these files exist, copy the kube-apiserver.yaml into /etc/kubernetes/manifests/
on each master node.
The kubelet monitors this directory, and will automatically create an instance of the kube-apiserver
container using the pod definition specified
in the file.
At this point, you should have 3 apiservers all working correctly. If you set up a network load balancer, you should be able to access your cluster via that load balancer, and see traffic balancing between the apiserver instances. Setting up a load balancer will depend on the specifics of your platform, for example instructions for the Google Cloud Platform can be found here
Note, if you are using authentication, you may need to regenerate your certificate to include the IP address of the balancer, in addition to the IP addresses of the individual nodes.
For pods that you deploy into the cluster, the kubernetes
service/dns name should provide a load balanced endpoint for the master automatically.
For external users of the API (e.g. the kubectl
command line interface, continuous build pipelines, or other clients) you will want to configure
them to talk to the external load balancer’s IP address.
So far we have set up state storage, and we have set up the API server, but we haven’t run anything that actually modifies
cluster state, such as the controller manager and scheduler. To achieve this reliably, we only want to have one actor modifying state at a time, but we want replicated
instances of these actors, in case a machine dies. To achieve this, we are going to use a lease-lock in the API to perform
master election. We will use the --leader-elect
flag for each scheduler and controller-manager, using a lease in the API will ensure that only 1 instance of the scheduler and controller-manager are running at once.
First, create empty log files on each node, so that Docker will mount the files not make new directories:
touch /var/log/kube-scheduler.log
touch /var/log/kube-controller-manager.log
Next, set up the descriptions of the scheduler and controller manager pods on each node.
by copying kube-scheduler.yaml and kube-controller-manager.yaml into the /etc/kubernetes/manifests/
directory.
At this point, you are done (yeah!) with the master components, but you still need to add worker nodes (boo!).
If you have an existing cluster, this is as simple as reconfiguring your kubelets to talk to the load-balanced endpoint, and restarting the kubelets on each node.
If you are turning up a fresh cluster, you will need to install the kubelet and kube-proxy on each worker node, and
set the --apiserver
flag to your replicated endpoint.