Upgrade Bare Metal cluster | EKS Anywhere ·

25 May 2024, 00:00

How to perform Bare Metal cluster upgrades

Considerations

Only EKS Anywhere and Kubernetes version upgrades are supported for Bare Metal clusters. You cannot update other cluster configuration.
Upgrades should never be run from ephemeral nodes (short-lived systems that spin up and down on a regular basis). If the EKS Anywhere version is lower than v0.18.0 and upgrade fails, you must not delete the KinD bootstrap cluster Docker container. During an upgrade, the bootstrap cluster contains critical EKS Anywhere components. If it is deleted after a failed upgrade, they cannot be recovered.
It is highly recommended to run the eksctl anywhere upgrade cluster command with the --no-timeouts option when the command is executed through automation. This prevents the CLI from timing out and enables cluster operators to fix issues preventing the upgrade from completing while the process is running.
In EKS Anywhere version v0.15.0, we introduced the EKS Anywhere cluster lifecycle controller that runs on management clusters and manages workload clusters. The EKS Anywhere lifecycle controller enables you to use Kubernetes API-compatible clients such as kubectl, GitOps, or Terraform for managing workload clusters. In this EKS Anywhere version, the EKS Anywhere cluster lifecycle controller rolls out new nodes in workload clusters when management clusters are upgraded. In EKS Anywhere version v0.16.0, this behavior was changed such that management clusters can be upgraded separately from workload clusters.
When running workload cluster upgrades after upgrading a management cluster, a machine rollout may be triggered on workload clusters during the workload cluster upgrade, even if the changes to the workload cluster spec didn’t require one (for example scaling down a worker node group).
Starting with EKS Anywhere v0.18.0, the osImageURL must include the Kubernetes minor version (Cluster.Spec.KubernetesVersion or Cluster.Spec.WorkerNodeGroupConfiguration[].KubernetesVersion in the cluster spec). For example, if the Kubernetes version is 1.29, the osImageURL must include 1.29, 1_29, 1-29 or 129. If you are upgrading Kubernetes versions, you must have a new OS image with your target Kubernetes version components.
If you are running EKS Anywhere in an airgapped environment, you must download the new artifacts and images prior to initiating the upgrade. Reference the Airgapped Upgrades page page for more information.

Upgrade Version Skew

There are a few dimensions of versioning to consider in your EKS Anywhere deployments:

Management clusters to workload clusters: Management clusters can be at most 1 EKS Anywhere minor version greater than the EKS Anywhere version of workload clusters. Workload clusters cannot have an EKS Anywhere version greater than management clusters.
Management components to cluster components: Management components can be at most 1 EKS Anywhere minor version greater than the EKS Anywhere version of cluster components.
EKS Anywhere version upgrades: Skipping EKS Anywhere minor versions during upgrade is not supported (v0.17.x to v0.19.x). We recommend you upgrade one EKS Anywhere minor version at a time (v0.17.x to v0.18.x to v0.19.x).
Kubernetes version upgrades: Skipping Kubernetes minor versions during upgrade is not supported (v1.26.x to v1.28.x). You must upgrade one Kubernetes minor version at a time (v1.26.x to v1.27.x to v1.28.x).
Kubernetes control plane and worker nodes: As of Kubernetes v1.28, worker nodes can be up to 3 minor versions lower than the Kubernetes control plane minor version. In earlier Kubernetes versions, worker nodes could be up to 2 minor versions lower than the Kubernetes control plane minor version.

Prerequisites

EKS Anywhere upgrades on Bare Metal require at least one spare hardware server for control plane upgrade and one for each worker node group upgrade. During upgrade, the spare hardware server is provisioned with the new version and then an old server is deprovisioned. The deprovisioned server is then reprovisioned with the new version while another old server is deprovisioned. This happens one at a time until all the control plane components have been upgraded, followed by worker node upgrades.

Check upgrade components

Before you perform an upgrade, check the current and new versions of components that are ready to upgrade by typing:

eksctl anywhere upgrade plan cluster -f cluster.yaml

The output should appear similar to the following:

Checking new release availability... NAME CURRENT VERSION NEXT VERSION EKS-A Management v0.19.0-dev+build.20+a0037f0 v0.19.0-dev+build.26+3bc5008 cert-manager v1.13.2+129095a v1.13.2+bb56494 cluster-api v1.6.1+5efe087 v1.6.1+9cf3436 kubeadm v1.6.1+8ceb315 v1.6.1+82f1c0a tinkerbell v0.4.0+cdde180 v0.4.0+e848206 kubeadm v1.6.1+6420e1c v1.6.1+2f0b35f etcdadm-bootstrap v1.0.10+7094b99 v1.0.10+a3f0355 etcdadm-controller v1.0.17+0259550 v1.0.17+ba86997

To format the output in json, add -o json to the end of the command line.

Check hardware availability

Next, you must ensure you have enough available hardware for the rolling upgrade operation to function. This type of upgrade requires you to have one spare hardware server for control plane upgrade and one for each worker node group upgrade. Check prerequisites for more information. Available hardware could have been fed to the cluster as extra hardware during a prior create command, or could be fed to the cluster during the upgrade process by providing the hardware CSV file to the upgrade cluster command .

To check if you have enough available hardware for rolling upgrade, you can use the kubectl command below to check if there are hardware objects with the selector labels corresponding to the controlplane/worker node group and without the ownerName label.

kubectl get hardware -n eksa-system --show-labels

For example, if you want to perform upgrade on a cluster with one worker node group with selector label type=worker-group-1, then you must have an additional hardware object in your cluster with the label type=controlplane (for control plane upgrade) and one with type=worker-group-1 (for worker node group upgrade) that doesn’t have the ownerName label.

In the command shown below, eksa-worker2 matches the selector label and it doesn’t have the ownerName label. Thus, it can be used to perform rolling upgrade of worker-group-1. Similarly, eksa-controlplane-spare will be used for rolling upgrade of control plane.

kubectl get hardware -n eksa-system --show-labels NAME STATE LABELS eksa-controlplane type=controlplane,v1alpha1.tinkerbell.org/ownerName=abhnvp-control-plane-template-1656427179688-9rm5f,v1alpha1.tinkerbell.org/ownerNamespace=eksa-system eksa-controlplane-spare type=controlplane eksa-worker1 type=worker-group-1,v1alpha1.tinkerbell.org/ownerName=abhnvp-md-0-1656427179689-9fqnx,v1alpha1.tinkerbell.org/ownerNamespace=eksa-system eksa-worker2 type=worker-group-1

If you don’t have any available hardware that match this requirement in the cluster, you can setup a new hardware CSV . You can feed this hardware inventory file during the upgrade cluster command .

Performing a cluster upgrade

To perform a cluster upgrade you can modify your cluster specification kubernetesVersion field to the desired version.

As an example, to upgrade a cluster with version 1.24 to 1.25 you would change your spec as follows:

apiVersion: anywhere.eks.amazonaws.com/v1alpha1 kind: Cluster metadata: name: dev spec: controlPlaneConfiguration: count: 1 endpoint: host: "198.18.99.49" machineGroupRef: kind: TinkerbellMachineConfig name: dev ... kubernetesVersion: "1.25" ...

NOTE: If you have a custom machine image for your nodes in your cluster config yaml or to upgrade a node or group of nodes to a new operating system version (ie. RHEL 8.7 to RHEL 8.8), you may also need to update your TinkerbellDatacenterConfig or TinkerbellMachineConfig with the new operating system image URL osImageURL .

and then you will run the upgrade cluster command .

Upgrade cluster command

kubectl CLI: The cluster lifecycle feature lets you use kubectl to talk to the Kubernetes API to upgrade a workload cluster. To use kubectl, run:
```
kubectl apply -f eksa-w01-cluster.yaml --kubeconfig mgmt/mgmt-eks-a-cluster.kubeconfig 
```
To check the state of a cluster managed with the cluster lifecyle feature, use kubectl to show the cluster object with its status.
The status field on the cluster object field holds information about the current state of the cluster.
```
kubectl get clusters w01 -o yaml 
```
The cluster has been fully upgraded once the status of the Ready condition is marked True. See the cluster status guide for more information.
GitOps: See Manage separate workload clusters with GitOps
Terraform: See Manage separate workload clusters with Terraform
NOTE:For kubectl, GitOps and Terraform:
- The baremetal controller does not support scaling upgrades and Kubernetes version upgrades in the same request.
- While scaling a workload cluster if you need to add additional machines, run:
```
eksctl anywhere generate hardware -z updated-hardware.csv > updated-hardware.yaml kubectl apply -f updated-hardware.yaml 
```
- If you want to upgrade multiple workload clusters, make sure that the spare hardware that is available for new nodes to rollout has labels unique to the workload cluster you are trying to upgrade. For instance, for an EKSA cluster named eksa-workload1, the hardware that is assigned for this cluster should have labels that are only going to be used for this cluster like type=eksa-workload1-cp and type=eksa-workload1-worker. Another workload cluster named eksa-workload2 can have labels like type=eksa-workload2-cp and type=eksa-workload2-worker. Please note that even though labels can be arbitrary, they need to be unique for each workload cluster. Not specifying unique cluster labels can cause cluster upgrades to behave in unexpected ways which may lead to unsuccessful upgrades and unstable clusters.
eksctl CLI: To upgrade a workload cluster with eksctl, run:
```
eksctl anywhere upgrade cluster -f cluster.yaml # --hardware-csv <hardware.csv> \ # uncomment to add more hardware --kubeconfig mgmt/mgmt-eks-a-cluster.kubeconfig 
```
As noted earlier, adding the --kubeconfig option tells eksctl to use the management cluster identified by that kubeconfig file to create a different workload cluster.
This will upgrade the cluster specification (if specified), upgrade the core components to the latest available versions and apply the changes using the provisioner controllers.
Output
Example output:
```
✅ control plane ready ✅ worker nodes ready ✅ nodes ready ✅ cluster CRDs ready ✅ cluster object present on workload cluster ✅ upgrade cluster kubernetes version increment ✅ validate immutable fields 🎉 all cluster upgrade preflight validations passed Performing provider setup and validations Ensuring etcd CAPI providers exist on management cluster before upgrade Pausing GitOps cluster resources reconcile Upgrading core components Backing up management cluster's resources before upgrading Upgrading management cluster Updating Git Repo with new EKS-A cluster spec Forcing reconcile Git repo with latest commit Resuming GitOps cluster resources kustomization Writing cluster config file 🎉 Cluster upgraded! Cleaning up backup resources 
```
Starting in EKS Anywhere v0.18.0, when upgrading management cluster the CLI depends on the EKS Anywhere Controller to perform the upgrade. In the event an issue occurs and the CLI times out, it may be possible to fix the issue and have the upgrade complete as the EKS Anywhere Controller will continually attempt to complete the upgrade.
During the workload cluster upgrade process, EKS Anywhere pauses the cluster controller reconciliation by adding the paused annotation anywhere.eks.amazonaws.com/paused: true to the EKS Anywhere cluster, provider datacenterconfig and machineconfig resources, before the components upgrade. After upgrade completes, the annotations are removed so that the cluster controller resumes reconciling the cluster. If the CLI execution is interrupted or times out, the controller won’t reconcile changes to the EKS-A objects until these annotations are removed. You can re-run the CLI to restart the upgrade process or remove the annotations manually with kubectl.
Though not recommended, you can manually pause the EKS Anywhere cluster controller reconciliation to perform extended maintenance work or interact with Cluster API objects directly. To do it, you can add the paused annotation to the cluster resource:
```
kubectl annotate clusters.anywhere.eks.amazonaws.com ${CLUSTER_NAME} -n ${CLUSTER_NAMESPACE} anywhere.eks.amazonaws.com/paused=true 
```
After finishing the task, make sure you resume the cluster reconciliation by removing the paused annotation, so that EKS Anywhere cluster controller can continue working as expected.
```
kubectl annotate clusters.anywhere.eks.amazonaws.com ${CLUSTER_NAME} -n ${CLUSTER_NAMESPACE} anywhere.eks.amazonaws.com/paused- 
```

Upgradeable cluster attributes

Cluster:

kubernetesVersion
controlPlaneConfiguration.count
controlPlaneConfiguration.upgradeRolloutStrategy.rollingUpdate.maxSurge
workerNodeGroupConfigurations.count
workerNodeGroupConfigurations.kubernetesVersion (in case of modular upgrade)
workerNodeGroupConfigurations.upgradeRolloutStrategy.rollingUpdate.maxSurge
workerNodeGroupConfigurations.upgradeRolloutStrategy.rollingUpdate.maxUnavailable

TinkerbellDatacenterConfig:

osImageURL

Advanced configuration for upgrade rollout strategy

EKS Anywhere allows an optional configuration to customize the behavior of upgrades.

upgradeRolloutStrategy can be configured separately for control plane and for each worker node group. This template contains an example for control plane under the controlPlaneConfiguration section and for worker node group under workerNodeGroupConfigurations:

apiVersion: anywhere.eks.amazonaws.com/v1alpha1 kind: Cluster metadata: name: my-cluster-name spec: clusterNetwork: cniConfig: cilium: {} pods: cidrBlocks: - 192.168.0.0/16 services: cidrBlocks: - 10.96.0.0/12 controlPlaneConfiguration: count: 1 endpoint: host: "10.61.248.209" machineGroupRef: kind: TinkerbellMachineConfig name: my-cluster-name-cp upgradeRolloutStrategy: type: RollingUpdate rollingUpdate: maxSurge: 1 datacenterRef: kind: TinkerbellDatacenterConfig name: my-cluster-name kubernetesVersion: "1.25" managementCluster: name: my-cluster-name workerNodeGroupConfigurations: - count: 2 machineGroupRef: kind: TinkerbellMachineConfig name: my-cluster-name name: md-0 upgradeRolloutStrategy: type: RollingUpdate rollingUpdate: maxSurge: 1 maxUnavailable: 0 --- ...

upgradeRolloutStrategy

Configuration parameters for upgrade strategy.

upgradeRolloutStrategy.type

Default: RollingUpdate

Type of rollout strategy. Supported values: RollingUpdate,InPlace.

NOTE: The upgrade rollout strategy type must be the same for all control plane and worker nodes.

upgradeRolloutStrategy.rollingUpdate

Configuration parameters for customizing rolling upgrade behavior.

NOTE: The rolling update parameters can only be configured if upgradeRolloutStrategy.type is RollingUpdate.

upgradeRolloutStrategy.rollingUpdate.maxSurge

Default: 1

This can not be 0 if maxUnavailable is 0.

The maximum number of machines that can be scheduled above the desired number of machines.

Example: When this is set to n, the new worker node group can be scaled up immediately by n when the rolling upgrade starts. Total number of machines in the cluster (old + new) never exceeds (desired number of machines + n). Once scale down happens and old machines are brought down, the new worker node group can be scaled up further ensuring that the total number of machines running at any time does not exceed the desired number of machines + n.

upgradeRolloutStrategy.rollingUpdate.maxUnavailable

Default: 0

This can not be 0 if MaxSurge is 0.

The maximum number of machines that can be unavailable during the upgrade.

This can only be configured for worker nodes.

Example: When this is set to n, the old worker node group can be scaled down by n machines immediately when the rolling upgrade starts. Once new machines are ready, old worker node group can be scaled down further, followed by scaling up the new worker node group, ensuring that the total number of machines unavailable at all times during the upgrade never falls below n.

Rolling Upgrades

The RollingUpdate rollout strategy type allows the specification of two parameters that control the desired behavior of rolling upgrades:

maxSurge - The maximum number of machines that can be scheduled above the desired number of machines. When not specified, the current CAPI default of 1 is used.
maxUnavailable - The maximum number of machines that can be unavailable during the upgrade. When not specified, the current CAPI default of 0 is used.

Example configuration:

upgradeRolloutStrategy: type: RollingUpdate rollingUpdate: maxSurge: 1 maxUnavailable: 0 # only configurable for worker nodes

Rolling upgrades with no additional hardware

When maxSurge is set to 0 and maxUnavailable is set to 1, it allows for a rolling upgrade without need for additional hardware. Use this configuration if your workloads can tolerate node unavailability.

NOTE: This could ONLY be used if unavailability of a maximum of 1 node is acceptable. For single node clusters, an additional temporary machine is a must. Alternatively, you may recreate the single node cluster for upgrading and handle data recovery manually.

With this kind of configuration, the rolling upgrade will proceed node by node, deprovision and delete a node fully before re-provisioning it with upgraded version, and re-join it to the cluster. This means that any point during the course of the rolling upgrade, there could be one unavailable node.

In-Place Upgrades

As of EKS Anywhere version v0.19.0, the InPlace rollout strategy type can be used to upgrade the EKS Anywhere and Kubernetes versions by upgrading the components on the same physical machines without requiring additional server capacity. EKS Anywhere schedules a privileged pod that executes the upgrade logic as a sequence of init containers on each node to be upgraded. This upgrade logic includes updating the containerd, cri-tools, kubeadm, kubectl and kubelet binaries along with core Kubernetes components and restarting those services.

Due to the nature of this upgrade, temporary downtime of workloads can be expected. It is best practice to configure your clusters in a way that they are resilient to having one node down.

During in place upgrades, EKS Anywhere pauses machine health checks to ensure that new nodes are not rolled out while the node is temporarily down during the upgrade process. Moreover, autoscaler configuration is not supported when using InPlace upgrade rollout strategy to further ensure that no new nodes are rolled out unexpectedly.