Upgrading

<span class="p-notification__title">Note:</span>
 <p class="p-notification__message">This page describes a specific upgrade process. It is important to follow the specific upgrade pages for each release, as these may include additional steps and workarounds for safely upgrading.</p>

It is recommended that you keep your Kubernetes deployment updated to the latest available stable version. You should also update the other applications which make up the Charmed Kubernetes. Keeping up to date ensures you have the latest bug-fixes and security patches for smooth operation of your cluster.

New minor versions of Kubernetes are set to release once per quarter. You can check the latest release version on the Kubernetes release page on GitHub. Charmed Kubernetes is kept in close sync with upstream Kubernetes: updated versions will be released within a week of a new upstream version of Kubernetes.

<span class="p-notification__title">Note:</span>
<p class="p-notification__message"><strong>Kubernetes</strong> will automatically handle patch releases. This means that the cluster will perform an unattended automatic upgrade between patch versions, e.g. 1.19.1 to 1.19.2. Attended upgrades are only required when you wish to upgrade a minor version, e.g. 1.18.x to 1.19.x.</p>

You can see which version of each application is currently deployed by running

juju status

The ‘App’ section of the output lists each application and its version number. Note that this is the version of the upstream application deployed. The version of the Juju charm is indicated under the column titled ‘Rev’. The charms may be updated in between new versions of the application.

Before you begin

As with all upgrades, there is a possibility that there may be unforeseen difficulties. It is highly recommended that you make a backup of any important data, including any running workloads. For more details on creating backups, see the separate documentation on backups.

You should also make sure:

• The machine from which you will perform the backup has sufficient internet access to retrieve updated software
• Your cluster is running normally
• Your Juju client and controller/models are running the latest versions (see the Juju docs)
• You read the Upgrade notes to see if any caveats apply to the versions you are upgrading to/from
• You read the Release notes for the version you are upgrading to, which will alert you to any important changes to the operation of your cluster
• You read the Upstream release notes for details of deprecation notices and API changes for Kubernetes 1.23 which may impact your workloads.

It is also important to understand that Charmed Kubernetes will only upgrade and if necessary migrate, components relating specifically to elements of Kubernetes installed and configured as part of Charmed Kubernetes. This may not include any customised configuration of Kubernetes, or user generated objects (e.g. storage classes) or deployments which rely on deprecated APIs.

Infrastructure updates

The applications which run alongside the core Kubernetes components can be upgraded at any time. These applications are widely used and may frequently receive upgrades outside of the cycle of new releases of Kubernetes.

This includes:

• Docker
• easyrsa
• etcd
• flannel, calico or other CNI

Note that this may include other applications which you may have installed, such as Elasticsearch, Prometheus, Nagios, Helm, etc.

Upgrading Containerd

By default, Versions 1.15 and later use Containerd as the container runtime. This subordinate charm can be upgraded with the command:

juju upgrade-charm containerd

Upgrading Docker (if used)

By default, versions of Charmed Kubernetes since 1.15 use the Containerd runtime. You will only need to upgrade the Docker runtime if you have explicitly set that to be the container runtime. If this is not the case, you should skip this section.

Charmed Kubernetes will use the latest stable version of Docker when it is deployed. Since upgrading Docker can cause service disruption, there will be no automatic upgrades and instead this process must be triggered by the operator.

Note that this upgrade step only applies to deployments which actually use the Docker container runtime. Versions 1.15 and later use containerd by default, and you should instead follow the instructions above.

Version 1.15 and later

The kubernetes-master and kubernetes-worker are related to the docker subordinate charm where present. Whether you are running Docker on its own, or mixed with Containerd, the upgrade process is the same:

juju upgrade-charm docker

Versions prior to 1.15

Only the kubernetes-master and kubernetes-worker units require Docker. The charms for each include an action to trigger the upgrade.

Before the upgrade, it is useful to list all the units effected:

juju status kubernetes-* --format=short

…will return a list of the current kubernetes-master and kubernetes-worker units.

Start with the kubernetes-master units and run the upgrade action on one unit at a time:

juju run-action kubernetes-master/0 upgrade-docker --wait

As Docker is restarted on the unit, pods will be terminated. Wait for them to respawn before moving on to the next unit:

juju run-action kubernetes-master/1 upgrade-docker --wait

Once all the kubernetes-master units have been upgraded and the pods have respawned, the same procedure can then be applied to the kubernetes-worker units.

juju run-action kubernetes-worker/0 upgrade-docker --wait

As previously, wait between running the action on sucessive units to allow pods to migrate.

Upgrading etcd

As etcd manages critical data for the cluster, it is advisable to create a snapshot of this data before running an upgrade. This is covered in more detail in the documentation on backups (including how to restore snapshots in case of problems).

Upgrade the charm with the command:

juju upgrade-charm etcd

To upgrade etcd itself, you will need to set the etcd charm’s channel config.

To determine the correct channel, go to the releases section of the bundle repository page and check the relevant Charmed Kubernetes bundle. Within the bundle, you should see which channel the etcd charm is configured to use.

Once you know the correct channel, set the etcd charm’s channel config:

juju config etcd channel=3.4/stable

Upgrading additional components

The other infrastructure applications can be upgraded by running the upgrade-charm command:

juju upgrade-charm easyrsa

Any other infrastructure charms should be upgraded in a similar way. For example, if you are using the flannel CNI:

juju upgrade-charm flannel

<span class="p-notification__title">Note:</span>
<p class="p-notification__message">Some services may be briefly interrupted during the upgrade process. Upgrading
your CNI (e.g. flannel) will cause a small amount of network downtime. Upgrading
<strong>easyrsa</strong> will not cause any downtime. The behaviour of other
components you have added to your cluster may vary - check individual documentation
for these charms for more information on upgrades.</p>

Upgrading Kubernetes

Before you upgrade the Kubernetes components, you should be aware of the exact release you wish to upgrade to.

The Kubernetes charms use snap channels to manage the version of Kubernetes to use. Channels are explained in more detail in the official snap documentation, but in terms of Kubernetes all you need to know are the major and minor version numbers and the ‘risk-level’:

For most use cases, it is strongly recommended to use the ‘stable’ version of charms.

Upgrading the kube-api-loadbalancer

A core part of Charmed Kubernetes is the kubeapi-load-balancer component. To ensure API service continuity this upgrade should precede any upgrades to the Kubernetes master and worker units.

juju upgrade-charm kubeapi-load-balancer

The load balancer itself is based on NGINX, and the version reported by juju status is that of NGINX rather than Kubernetes. Unlike the other Kubernetes components, there is no need to set a specific channel or version for this charm.

Upgrading the kubernetes-master units

To start upgrading the Kubernetes master units, first upgrade the charm:

juju upgrade-charm kubernetes-master

Once the charm has been upgraded, it can be configured to select the desired Kubernetes channel, which takes the form Major.Minor/risk-level. This is then passed as a configuration option to the charm. So, for example, to select the stable 1.23 version of Kubernetes, you would enter:

juju config kubernetes-master channel=1.23/stable

If you wanted to try a release candidate for 1.23, the channel would be 1.23/candidate.

<span class="p-notification__title">Note:</span>
<p class="p-notification__message">Once the configuration has been changed, the charms will be put into a `blocked` state.
You must continue the upgrade process, even if you revert the configuration to the
currently active version of Kubernetes.</p>

Once the desired version has been configured, the upgrades should be performed. This is done by running the upgrade action on each master unit in the cluster:

juju run-action kubernetes-master/0 upgrade
juju run-action kubernetes-master/1 upgrade

If you have more master units in your cluster, you should continue and run this process on every one of them.

You can check the progress of the upgrade by running:

juju status | grep master

Ensure that all the master units have upgraded and are reporting normal status before continuing to upgrade the worker units.

Upgrading the kubernetes-worker units

<span class="p-notification__title">Caution:</span>
<p class="p-notification__message">A <a href="https://github.com/kubernetes/kubernetes/issues/70044"> current bug in Kubernetes</a> could prevent the upgrade from properly deleting old pods. See the <a href="#known-issues"> Known issues section</a> at the bottom of this page.</p>

For a running cluster, there are two different ways to proceed:

• Blue-green upgrade - This requires more resources, but should ensure a safe, zero-downtime transition of workloads to an updated cluster
• In-place upgrade - this simply upgrades the workers in situ, which may involve some service interruption but doesn’t require extra resources

Both methods are outlined below. The blue-green method is recommended for production systems.

Blue-green upgrade

To begin, upgrade the kubernetes-worker charm itself:

juju upgrade-charm kubernetes-worker

Next, run the command to configure the workers for the version of Kubernetes you wish to run (as you did previously for the master units). For example:

juju config kubernetes-worker channel=1.23/stable

Now add additional units of the kubernetes-worker. You should add as many units as you are replacing. For example, to add three additional units:

juju add-unit kubernetes-worker -n 3

This will create new units to migrate the existing workload to. As you configured the version prior to adding the units, they will be using the newly-selected version of Kubernetes.

Now we can pause the existing workers, which will cause the workloads to migrate to the new units recently added. A worker unit is paused by running the corresponding action on that unit:

juju run-action kubernetes-worker/0 pause
juju run-action kubernetes-worker/1 pause
juju run-action kubernetes-worker/2 pause
...

Continue until all the ‘old’ units have been paused. You can check on the workload status by running the command:

kubectl get pod -o wide

Once the workloads are running on the new units, it is safe to remove the old units:

juju remove-unit kubernetes-worker/0

Removing these units from the model will also release the underlying machines/instances they were running on, so no further clean up is required.

<span class="p-notification__title">Note:</span>
<p class="p-notification__message">A variation on this method is to add, pause, remove  and recycle units one at a time. This reduces the resource overhead to a single extra instance.</p>

In-place upgrade

To proceed with an in-place upgrade, first upgrade the charm itself:

juju upgrade-charm kubernetes-worker

Next, run the command to configure the workers for the version of Kubernetes you wish to run (as you did previously for the master units). For example:

juju config kubernetes-worker channel=1.23/stable

All the units can now be upgraded by running the upgrade action on each one:

juju run-action kubernetes-worker/0 upgrade
juju run-action kubernetes-worker/1 upgrade
...

Upgrading the Machine’s Series

All of the charms support upgrading the machine’s series via Juju. As each machine is upgraded, the applications on that machine will be stopped and the unit will go into a blocked status until the upgrade is complete. For the worker units, pods will be drained from the node and onto one of the other nodes at the start of the upgrade, and the node will be removed from the pool until the upgrade is complete.

Verify an Upgrade

Once an upgrade is complete and units settle, the output from:

juju status

… should indicate that all units are active and the correct version of Kubernetes is running.

It is recommended that you run a cluster validation to ensure that the cluster is fully functional.

Known Issues

A current bug in Kubernetes could prevent the upgrade from properly deleting old pods. You can see such an issue here:

kubectl get po --all-namespaces

NAMESPACE                         NAME                                                          READY   STATUS        RESTARTS   AGE
default                           nginx-ingress-kubernetes-worker-controller-r8d2v              0/1     Terminating   0          17m
ingress-nginx-kubernetes-worker   default-http-backend-kubernetes-worker-5d9bb77bc5-76c8w       1/1     Running       0          10m
ingress-nginx-kubernetes-worker   nginx-ingress-controller-kubernetes-worker-5dcf47fc4c-q9mh6   1/1     Running       0          10m
kube-system                       heapster-v1.6.0-beta.1-6db4b87d-phjvb                         4/4     Running       0          16m
kube-system                       kube-dns-596fbb8fbd-bp8lz                                     3/3     Running       0          18m
kube-system                       kubernetes-dashboard-67d4c89764-nwxss                         1/1     Running       0          18m
kube-system                       metrics-server-v0.3.1-67bb5c8d7-x9nzx                         2/2     Running       0          17m
kube-system                       monitoring-influxdb-grafana-v4-65cc9bb8c8-mwvcm               2/2     Running       0          17m

In this case the nginx-ingress-kubernetes-worker-controller-r8d2v has been stuck in the Terminating state for roughly 10 minutes. The workaround for such a problem is to force a deletion:

kubectl delete po/nginx-ingress-kubernetes-worker-controller-r8d2v --force --grace-period=0

This will result in output similar to the following:

warning: Immediate deletion does not wait for confirmation that the running resource has been terminated. The resource may continue to run on the cluster indefinitely.
pod "nginx-ingress-kubernetes-worker-controller-r8d2v" force deleted

You should verify that the pod has been sucessfully removed:

kubectl get po --all-namespaces

NAMESPACE                         NAME                                                          READY   STATUS    RESTARTS   AGE
ingress-nginx-kubernetes-worker   default-http-backend-kubernetes-worker-5d9bb77bc5-76c8w       1/1     Running   0          11m
ingress-nginx-kubernetes-worker   nginx-ingress-controller-kubernetes-worker-5dcf47fc4c-q9mh6   1/1     Running   0          11m
kube-system                       heapster-v1.6.0-beta.1-6db4b87d-phjvb                         4/4     Running   0          17m
kube-system                       kube-dns-596fbb8fbd-bp8lz                                     3/3     Running   0          19m
kube-system                       kubernetes-dashboard-67d4c89764-nwxss                         1/1     Running   0          19m
kube-system                       metrics-server-v0.3.1-67bb5c8d7-x9nzx                         2/2     Running   0          18m
kube-system                       monitoring-influxdb-grafana-v4-65cc9bb8c8-mwvcm               2/2     Running   0          18m