Skip to content

Latest commit

 

History

History
518 lines (431 loc) · 26.5 KB

Upgrade-Guide.md

File metadata and controls

518 lines (431 loc) · 26.5 KB
Raw
title version authors state
SCS k8s-cluster-api-provider upgrade guide
2023-09-07
Kurt Garloff, Roman Hros, Matej Feder
Draft (v0.7)

SCS k8s-cluster-api-provider upgrade guide

This document explains the steps to upgrade the SCS Kubernetes cluster-API based cluster management solution as follows:

  • from the R2 (2022-03) to the R3 (2022-09) state
  • from the R3 (2022-09) to the R4 (2023-03) state
  • from the R4 (2023-03) to the R5 (2023-09) state
  • from the R5 (2023-09) to the R6 (2024-03) state

The document explains how the management cluster and the workload clusters can be upgraded without disruption. It is highly recommended to do a step-by-step upgrade across major releases i.e. upgrade from R2 to R3 and then to R4 in the case of upgrade from the R2 to the R4. Upgrades across major releases without step-by-step process is not recommended and could lead to undocumented issues.

The various steps are not very complicated, but there are numerous steps to take, and it is advisable that cluster operators get some experience with this kind of cluster management before applying this to customer clusters that carry important workloads.

Note that while the detailed steps are tested and targeted to an R2 -> R3 move, R3 -> R4 move, R4 -> R5 move or R5 -> R6 move, many of the steps are a generic approach that will apply also for other upgrades, so expect a lot of similar steps when moving beyond R6.

Upgrades from cluster management prior to R2 are difficult; many changes before R2 assumed that you would redeploy the management cluster. Redeploying the management cluster can of course always be done, but it's typically disruptive to your workload clusters, unless you move your cluster management state into a new management cluster with clusterctl move.

Management host (cluster) vs. Workload clusters

When you initially deployed the SCS k8s-cluster-api-provider, you created a VM with a kind cluster inside and a number of templates, scripts and binaries that are then used to do the cluster management. This is your management host (or more precisely your single-host management cluster). Currently, all cluster management including upgrading etc. is done by connecting to this host via ssh and performing commands there. (You don't need root privileges to do cluster management there, the normal ubuntu user rights are sufficient; there are obviously host management tasks such as installing package updates that do require root power and the user has the sudo rights to do so.)

When you create the management host, you have the option to create your first workload cluster. This cluster is no different from other workload clusters that you create by calling commands on the management host, so you can manage it there. (The default name of this cluster is typically testcluster, though that can be changed since a while, #264).

On the management host, you have the openstack and kubernetes tools installed and configured, so you can nicely manage all aspects of your CaaS setups as well as the underlying IaaS. The kubectl configuration is in ~/.kube/config while you will find the OpenStack configuration in ~/.config/openstack/clouds.yaml. These files are automatically managed; you can add entries to the files though, and they should persist.

Updating the management host

There are two different possibilities to upgrade the management host.

  1. You do a component-wise in-place upgrade of it.
  2. You deploy a new management host and clusterctl move the resources over to it from the old one. (Note: Config state in ~/CLUSTER_NAME/)

TODO: Advice when to do what, risks, limitations

In-place upgrade

Operating system

You should keep the host up-to-date with respect to normal operating system upgrades, so perform your normal sudo apt-get update && sudo apt-get upgrade. kubectl, kustomize, yq, lxd and a few other tools are installed as snaps, so you may want to upgrade these as well: sudo snap refresh. From R5 sudo apt-get install -y jq is also required as this is used by the diskless flavors feature, #424. Default operating system image was changed from Ubuntu 20.04 to Ubuntu 22.04 in R4.

k8s-cluster-api-provider git

The automation is deployed on the management host by cloning the relevant git repository. into the k8s-cluster-api-provider directory. Note that the checked out branch will be the one that has been used when creating the management host, and you might want to change branches from maintained/v3.x to maintained/v4.x in case of R2 to R3 upgrade, maintained/v5.x for R3 to R4 upgrade, maintained/v6.x for R4 to R5 upgrade or maintained/v7.x for R5 to R6 upgrade. Use git branch -rl to see available branches in the k8s-cluster-api-provider repository.

You can update the scripts and templates by checking out the relevant branch main, maintained/v4.x, maintained/v5.x, maintained/v6.x or maintained/v7.x and using a git pull to ensure the latest content is retrieved. Once you do that, the cluster-management scripts will be up-to-date. (The ~/bin directory in your search PATH is symlinked to the check-ed out scripts.)

Note however that the binaries and used templates are NOT automatically updated. This should not normally result in problems -- when new features are introduced in the management scripts, they ensure to continue to support older templates.

Updating cluster-API and openstack cluster-API provider

To get the latest version of cluster-API, you can download a new clusterctl binary from https://github.com/kubernetes-sigs/cluster-api/releases, make it executable chmod +x clusterctl and install it to /usr/local/bin/, possibly saving the old binary by renaming it. clusterctl version should now display the current version number (v1.6.2 at the time of this writing).

You can now issue the command clusterctl upgrade plan and clusterctl will list the components in your (kind) management cluster that can be upgraded. Here's an example output:

ubuntu@capi-old-mgmtcluster:~ [0]$ clusterctl upgrade plan
Checking cert-manager version...
Cert-Manager is already up to date

Checking new release availability...

Latest release available for the v1beta1 API Version of Cluster API (contract):

NAME                       NAMESPACE                           TYPE                     CURRENT VERSION   NEXT VERSION
bootstrap-kubeadm          capi-kubeadm-bootstrap-system       BootstrapProvider        v1.5.1            v1.6.2
control-plane-kubeadm      capi-kubeadm-control-plane-system   ControlPlaneProvider     v1.5.1            v1.6.2
cluster-api                capi-system                         CoreProvider             v1.5.1            v1.6.2
infrastructure-openstack   capo-system                         InfrastructureProvider   v0.7.3            v0.9.0

You can now apply the upgrade by executing the following command:

clusterctl upgrade apply --contract v1beta1

You can then upgrade the components:

  1. export CLUSTER_TOPOLOGY=true - this is needed from R5 to R6 upgrade due to ClusterClass feature #600
  2. Upgrade components
    • You can do them one-by-one, e.g.: 
      clusterctl upgrade apply --infrastructure capo-system/openstack:v0.9.0 --core capi-system/cluster-api:v1.6.2 -b capi-kubeadm-bootstrap-system/kubeadm:v1.6.2 -c capi-kubeadm-control-plane-system/kubeadm:v1.6.2
    • Or simply do clusterctl upgrade apply --contract v1beta1

New templates

The cluster-template.yaml template used for the workload clusters is located in ~/k8s-cluster-api-provider/terraform/files/template/ and copied from there into ~/cluster-defaults/. Then workload clusters are created, they will also have a copy of it in ~/${CLUSTER_NAME}/. If you have not changed it manually, you can copy it over the old templates. (Consider backing up the old one though.)

The next create_cluster.sh <CLUSTER_NAME> run will then use the new template. Note that create_cluster.sh is idempotent -- it will not perform any changes on the cluster unless you have changed its configuration by tweaking cluster-template.yaml (which you almost never do!) or clusterctl.yaml (which you do often).

The other template file that changed -- however, some opentofu logic is used to prefill it with values. So you can't copy it from git.

R2 to R3

For going from R2 to R3, there is just one real change that you want to apply: Add the variables CONTROL_PLANE_MACHINE_GEN: genc01 and WORKER_MACHINE_GEN: genw01 to it. If you have copied over the new cluster-template.yaml as described above, then you're done. Otherwise you can use the script update-R2-R3.sh <CLUSTER_NAME> to tweak both clusterctl.yaml and cluster-template.yaml for the relevant cluster. (You can use cluster-defaults to change the templates in ~/cluster-defaults/ which get used when creating new clusters.)

R3 to R4

In the R3 to R4, CALICO_VERSION was moved from .capi-settings to clusterctl.yaml. So before upgrading workload clusters, you must add it also to the ~/${CLUSTER_NAME}/clusterctl.yaml.

echo "CALICO_VERSION: v3.25.0" >> ~/cluster-defaults/clusterctl.yaml
echo "CALICO_VERSION: v3.25.0" >> ~/testcluster/clusterctl.yaml

In the R3 to R4 upgrade process, cluster-template.yaml changed etcd defrag process in the kubeadm control-planes and also security group names by adding ${PREFIX}- to them, so it has to be changed also in openstack project e.g. (PREFIX=capi):

openstack security group set --name capi-allow-ssh allow-ssh
openstack security group set --name capi-allow-icmp allow-icmp

We changed immutable fields in the Cluster API templates, so before running create_cluster.sh to upgrade existing workload cluster the CONTROL_PLANE_MACHINE_GEN and WORKER_MACHINE_GEN needs to be incremented in cluster specific clusterctl.yaml.

In the R3 to R4 process, also cloud.conf added enable-ingress-hostname=true to the LoadBalancer section. It is due to NGINX_INGRESS_PROXY defaulting to true now. So if you want to use, or you are already using this proxy functionality we recommend you to add this line to your cloud.conf, e.g.:

echo "enable-ingress-hostname=true" >> ~/cluster-defaults/cloud.conf
echo "enable-ingress-hostname=true" >> ~/testcluster/cloud.conf

Then, before upgrading workload cluster by create_cluster.sh, you should delete cloud-config secret in the kube-system namespace, so it can be recreated. E.g.: kubectl delete secret cloud-config -n kube-system --kubeconfig=testcluster/testcluster.yaml

Also, the default nginx-ingress version has changed, so we recommend before upgrading cluster to delete ingress-nginx jobs, so new job with new image can be created in the update process.

kubectl delete job ingress-nginx-admission-create -n ingress-nginx --kubeconfig=testcluster/testcluster.yaml
kubectl delete job ingress-nginx-admission-patch -n ingress-nginx --kubeconfig=testcluster/testcluster.yaml
R4 to R5

In R4 to R5, the cluster-template.yaml and clusterctl.yaml changed (see release notes). You can use script update-R4-to-R5.sh to update the cluster's cluster-template.yaml and clusterctl.yaml from R4 to R5. This script could update an existing Kubernetes cluster configuration files as well as cluster-defaults files that could be used for spawning new R5 clusters.

If you want to update an existing cluster configuration files from R4 to R5, just use script as follows:

update-R4-to-R5.sh <CLUSTER_NAME>

After you executed the above you will find that e.g. Calico version has been bumped from v3.25.0 to v3.26.1. Note that some software versions are not configurable and are not directly mentioned in the cluster configuration files, but they are hardcoded in R5 scripts (e.g. ingress nginx controller, metrics server), see new-defaults section. Note that the Kubernetes version was not updated as well the default CNI is not the Cilium yet. This two R5 features are out of scope this script when it is applied on the existing cluster configuration files as this features require advanced action such as CNI migration and step-by-step Kubernetes upgrade of +2 minor releases.

If you want to update cluster-defaults configuration files from R4 to R5, just use script as follows:

update-R4-to-R5.sh cluster-defaults

The above action updates a cluster-defaults configuration file, which is almost similar to updating an existing cluster configuration file described above. The distinction lies in the fact that both the Kubernetes version and the default CNI are also updated, specifically to Kubernetes version v1.27.5 and Cilium as a default CNI.

R5 to R6

In R5 to R6, the cluster-template.yaml and clusterctl.yaml changed (see release notes). You can use script update-R5-to-R6.sh to update the cluster's cluster-template.yaml and clusterctl.yaml from R5 to R6. This script could update an existing Kubernetes cluster configuration files as well as cluster-defaults files that could be used for spawning new R6 clusters.

If you want to update an existing cluster configuration files from R5 to R6, just use script as follows:

update-R5-to-R6.sh <CLUSTER_NAME>

After you executed the above you will find that e.g. Calico version has been bumped from v3.26.1 to v3.27.2 or Kubernetes version bumped from v1.27.5 to v1.28.7. Note that some software versions are not configurable and are not directly mentioned in the cluster configuration files, but they are hardcoded in R6 scripts (e.g. ingress nginx controller, metrics server, cilium), see new-defaults section.

If you want to update cluster-defaults configuration files from R5 to R6, just use script as follows:

update-R5-to-R6.sh cluster-defaults

If you are curious: In R2, doing rolling upgrades of k8s versions required edits in cluster-template.yaml -- this is no longer the case in R3, R4, R5 and R6. Just increase the generation counter for node and control plane nodes if you upgrade k8s versions -- or otherwise change the worker or control plane node specs, such as e.g. using a different flavor.

New defaults

You deploy a CNI (calico or cilium), the OpenStack Cloud Controller Manager (OCCM), the cinder CSI driver to clusters; optionally also a metrics server (default is true), a nginx ingress controller (also defaulting to true), the flux2 controller, the cert-manager. Some of these tools come with binaries that you can use for management purposes and which get installed on the management host in /usr/local/bin/.

The scripts that deploy these components into your workload clusters download the manifests into ~/kubernetes-manifests.d/ with a version specific name. If you request a new version, a new download will happen; already existing versions will not be re-downloaded.

Most binaries in /usr/local/bin/ are not stored under a version-specific name. You need to rename them to case a re-download of a newer version. (The reason for not having version specific names is that this would break scripts from users that assume the unversioned names; the good news is that most of these binaries have no trouble managing somewhat older deployments, so you can typically work with the latest binary tool even if you have a variety of versions deployed into various clusters.)

The defaults have changed as follows:

R2 R3 R4 R5 R6
kind v0.14.0 v0.14.0 v0.17.0 v0.20.0 v0.20.0
capi v1.0.5 v1.2.2 v1.3.5 v1.5.1 v1.6.2
capo v0.5.3 v0.6.3 v0.7.1 v0.7.3 v0.9.0
helm v3.8.1 v3.9.4 v3.11.1 v3.12.3 v3.14.1
sonobuoy v0.56.2 v0.56.10 v0.56.16 v0.56.17 v0.57.1
k9s unversioned unversioned unversioned v0.27.4 v0.31.9
calico v3.22.1 v3.24.1 v3.25.0 v3.26.1 v3.27.2
calico CLI v3.22.1 v3.24.1 v3.25.0 v3.26.1 v3.27.2
cilium unversioned unversioned v1.13.0 v1.14.1 v1.15.1
cilium CLI unversioned unversioned v0.13.1 v0.15.7 v0.15.23
hubble CLI unversioned unversioned v0.11.2 v0.12.0 v0.13.0
nginx-ingress v1.1.2 v1.3.0 v1.6.4 v1.8.1 v1.9.6
flux2 unversioned unversioned v0.40.2 v2.1.0 v2.2.3
cert-manager v1.7.1 v1.9.1 v1.11.0 v1.12.4 v1.14.2
metrics-server v0.6.1 v0.6.1 v0.6.1 v0.6.4 v0.7.0
kubectx v0.9.5
kube-ps1 v0.8.0

The clusterctl move approach

To be written

  1. Create new management host in same project -- avoid name conflicts with different prefix, to be tweaked later. Avoid testcluster creation
  2. Ensure it's up and running ...
  3. Tweak prefix
  4. Copy over configs (and a bit of state though that's uncritical) by using the directories
  5. Copy over the openstack credentials clouds.yaml and the kubectl config
  6. clusterctl move

Updating workload clusters

k8s version upgrade

On R2 clusters

The old way: Editing cluster-template.yaml. Or better use the update-R2-to-R3.sh script to convert first.

On R3 and R4 clusters

Edit ~/<CLUSTER_NAME>/clusterctl.yaml and put the wanted version into the fields KUBERNETES_VERSION and OPENSTACK_IMAGE_NAME. The node image will be downloaded from https://minio.services.osism.tech/openstack-k8s-capi-images and registered if needed. (If you have updated the k8s-cluster-api-provider repo, you can use a version v1.NN.x, where you fill in NN with the wanted k8s version, but leave a literal .x which will get translated to the newest tested version.)

In the same file, increase the generation counters for CONTROL_PLANE_MACHINE_GEN and WORKER_MACHINE_GEN.

Now do the normal create_cluster.sh <CLUSTER_NAME> and watch cluster-api replace your worker nodes and doing a rolling upgrade of your control plane. If you used a 3-node (or 5 or higher) control plane node setup, you will have uninterrupted access not just to your workloads but also the workload's cluster control plane. Use clusterctl describe cluster <CLUSTER_NAME> or simply kubectl --context <CLUSTER_NAME>-admin@<CLUSTER_NAME> get nodes -o wide to watch the progress of this.

On R5 clusters

If you decide to migrate your existing Kubernetes cluster from R4 to R5 be aware of the following:

  • R5 features such as per cluster namespaces and Cilium as a default CNI are supported only on new clusters and will not be migrated on the existing clusters
  • R4 default Kubernetes version v1.25.6 can not be directly migrated to the R5 default Kubernetes version v1.27.5, because +2 minor Kubernetes version upgrade is not allowed. See further migration steps below if you want to upgrade also Kubernetes version to R5

Follow the below steps if you want to migrate an existing cluster from R4 to R5:

  1. Access your management node
  2. Checkout R5 branch 
    cd k8s-cluster-api-provider
git pull
git checkout maintained/v6.x
  3. Backup an existing cluster configuration files (recommended) 
    cd ..
cp -R <CLUSTER_NAME> <CLUSTER_NAME>-backup
  4. Update an existing cluster configuration files from R4 to R5 
    update-R4-to-R5.sh <CLUSTER_NAME>
  5. Validate updated cluster configuration files. You will find that e.g. Calico version has been bumped from v3.25.0 to v3.26.1. Note that some software versions are not configurable and are not directly mentioned in the cluster configuration files, but they are hardcoded in R5 scripts (e.g. ingress nginx controller, metrics server). Hence, read carefully the R5 release notes too. Also see that Kubernetes version was not updated, and it is still v1.25.6.
  6. Update an existing cluster (except Kubernetes version) 
    create_cluster.sh <CLUSTER_NAME>
  7. Update cluster-API and openstack cluster-API provider, see related section for details 
    clusterctl upgrade plan
clusterctl upgrade apply --contract v1beta1
  8. Bump Kubernetes version +1 minor release (to v1.26.8) and increase the generation counter for node and control plane nodes 
    sed -i 's/^KUBERNETES_VERSION: v1.25.6/KUBERNETES_VERSION: v1.26.8/' <CLUSTER_NAME>/clusterctl.yaml
sed -i 's/^OPENSTACK_IMAGE_NAME: ubuntu-capi-image-v1.25.6/OPENSTACK_IMAGE_NAME: ubuntu-capi-image-v1.26.8/' <CLUSTER_NAME>/clusterctl.yaml
sed -r 's/(^CONTROL_PLANE_MACHINE_GEN: genc)([0-9][0-9])/printf "\1%02d" $((\2+1))/ge' -i <CLUSTER_NAME>/clusterctl.yaml
sed -r 's/(^WORKER_MACHINE_GEN: genw)([0-9][0-9])/printf "\1%02d" $((\2+1))/ge' -i <CLUSTER_NAME>/clusterctl.yaml
  9. Update an existing cluster Kubernetes version to v1.26.8 
    create_cluster.sh <CLUSTER_NAME>
  10. Bump Kubernetes version to R5 v1.27.5 and increase the generation counter for node and control plane nodes
sed -i 's/^KUBERNETES_VERSION: v1.26.8/KUBERNETES_VERSION: v1.27.5/' <CLUSTER_NAME>/clusterctl.yaml
sed -i 's/^OPENSTACK_IMAGE_NAME: ubuntu-capi-image-v1.26.8/OPENSTACK_IMAGE_NAME: ubuntu-capi-image-v1.27.5/' <CLUSTER_NAME>/clusterctl.yaml
sed -r 's/(^CONTROL_PLANE_MACHINE_GEN: genc)([0-9][0-9])/printf "\1%02d" $((\2+1))/ge' -i <CLUSTER_NAME>/clusterctl.yaml
sed -r 's/(^WORKER_MACHINE_GEN: genw)([0-9][0-9])/printf "\1%02d" $((\2+1))/ge' -i <CLUSTER_NAME>/clusterctl.yaml
  1. Update an existing cluster to the R5 Kubernetes version v1.27.5 
    create_cluster.sh <CLUSTER_NAME>

On R6 clusters

If you decide to migrate your existing Kubernetes cluster from R5 to R6 be aware of the following:

  • Kubernetes version will be upgraded from v1.27.5 to v1.28.7
  • You have to migrate from Cluster based templates to ClusterClass based templates
  • Upgrade of cilium needs to be done manually (for clusters with USE_CILIUM: true)

Follow the below steps if you want to migrate an existing cluster from R5 to R6:

  1. Access your management node
  2. Checkout R6 branch 
    cd k8s-cluster-api-provider
git pull
git checkout maintained/v7.x
  3. Backup an existing cluster configuration files (recommended) 
    cd ..
cp -R <CLUSTER_NAME> <CLUSTER_NAME>-backup
  4. Update an existing cluster configuration files from R5 to R6 
    update-R5-to-R6.sh <CLUSTER_NAME>
  5. Validate updated cluster configuration files. You will find that e.g. Calico version has been bumped from v3.26.1 to v3.27.2 or Kubernetes version bumped from v1.27.5 to v1.28.7. Note that some software versions are not configurable and are not directly mentioned in the cluster configuration files, but they are hardcoded in R6 scripts (e.g. ingress nginx controller, metrics server, cilium). Hence, read carefully the R6 release notes too.
  6. Update cluster-API and openstack cluster-API provider, see related section for details 
    clusterctl upgrade plan
export CLUSTER_TOPOLOGY=true
clusterctl upgrade apply --contract v1beta1
  7. Migrate to ClusterClass 
    migrate-to-cluster-class.sh <CLUSTER_NAME>
  8. Increase the generation counter for worker and control plane nodes 
    sed -r 's/(^CONTROL_PLANE_MACHINE_GEN: genc)([0-9][0-9])/printf "\1%02d" $((\2+1))/ge' -i <CLUSTER_NAME>/clusterctl.yaml
sed -r 's/(^WORKER_MACHINE_GEN: genw)([0-9][0-9])/printf "\1%02d" $((\2+1))/ge' -i <CLUSTER_NAME>/clusterctl.yaml
  9. Update an existing cluster to the R6 
    create_cluster.sh <CLUSTER_NAME>

    Note: You will probably experience a double rollout of nodes because the k8s version and templates are changed concurrently here. See https://cluster-api.sigs.k8s.io/tasks/experimental-features/cluster-class/operate-cluster#effects-of-concurrent-changes

  10. Upgrade cilium (for clusters with USE_CILIUM: true) 
    KUBECONFIG=<CLUSTER_NAME>/<CLUSTER_NAME>.yaml bash -c 'helm get values cilium -n kube-system -o yaml | cilium upgrade --version v1.15.1 -f -'

New versions for mandatory components

OCCM, CNI (calico/cilium), CSI

New versions for optional components

nginx, metrics server, cert-manager, flux

etcd leader changes

While testing clusters with >= 3 control nodes, we have observed occasional transient error messages that reported an etcd leader change preventing a command from succeeding. This could result in a dozen of random failed tests in a sonobuoy conformance run. (Retrying the commands would let them succeed.)

Too frequent etcd leader changes are detrimental to your control plane performance and can lead to transient failures. They are a sign that the infrastructure supporting your cluster is introducing too high latencies.

We recommend to deploy the control nodes (which run etcd) on instances with local SSD storage (which we reflect in the default flavor name) and recommend using flavors with dedicated cores and that your network does not introduce latencies by significant packet drop.

We now always use slower heartbeat (250ms) and increase CPU and IO priority which used to be controlled by ETCD_PRIO_BOOST. This is safe.

If you build multi-controller clusters and can not use a flavor with low latency local storage (ideally SSD), you can also work around this with ETCD_UNSAFE_FS. ETCD_UNSAFE_FS is using barrier=0 mount option, which violates filesystem ordering guarantees. This works around storage latencies, but introduces the risk of inconsistent filesystem state and inconsistent etcd data in case of an unclean shutdown. You may be able to live with this risk in a multi-controller etcd setup. If you don't have flavors that fulfill the requirements (low-latency storage attached), your choice is between a single-controller cluster (without ETCD_UNSAFE_FS) and a multi-controller cluster with ETCD_UNSAFE_FS. Neither option is perfect, but we deem the multi-controller cluster preferable in such a scenario.