Lifecycle and stages of cluster upgrades

When you upgrade GKE on Bare Metal, the upgrade process involves multiple steps and components. To help monitor the upgrade status or diagnose and troubleshoot problems, it's helpful to know what happens when you run the bmctl upgrade cluster command. This documents details the components and stages of a cluster upgrade.

Overview

The upgrade process moves your GKE on Bare Metal cluster from its current version to a higher version.

This version information is stored in the following locations as part of the cluster custom resource in the admin cluster:

  • status.anthosBareMetalVersion: defines the current version of the cluster.
  • spec.anthosBareMetalVersion: defines the target version, and is set when the upgrade process starts to run.

A successful upgrade operation reconciles status.anthosBareMetalVersion to spec.anthosBareMetalVersion so that both show the target version.

Version skew

The version skew is the difference in versions between an admin cluster and its managed user clusters. GKE on Bare Metal clusters follow the same style as Kubernetes: the admin cluster can be at most one minor version ahead of its managed clusters.

Version rules for upgrades

When you download and install a new version of bmctl, you can upgrade your admin, hybrid, standalone, and user clusters created or upgraded with an earlier version of bmctl. Clusters can't be downgraded to a lower version.

You can only upgrade a cluster to a version that matches the version of bmctl you are using. That is, if you are using version 1.28.100-gke.146 of bmctl, you can upgrade a cluster to version 1.28.100-gke.146 only.

Patch version upgrades

For a given minor version, you can upgrade to any higher patch version. That is, you can upgrade a 1.28.X version cluster to version 1.28.Y as long as Y is greater than X. For example, you can upgrade from 1.16.0 to 1.16.1 and you can upgrade from 1.16.1 to 1.16.3. We recommend that you upgrade to the latest patch version whenever possible to ensure your clusters have the latest security fixes.

Minor version upgrades

You can upgrade clusters from one minor version to the next, regardless of the patch version. That is, you can upgrade from 1.N.X to 1.N+1.Y, where 1.N.X is the version of your cluster and N+1 is the next available minor version. The patch versions, X and Y, don't affect the upgrade logic in this case. For example, you can upgrade from 1.16.3 to 1.28.100-gke.146.

You can't skip minor versions when upgrading clusters. If you attempt to upgrade to a minor version that is two or more minor versions higher than the cluster version, bmctl emits an error. For example, you can't upgrade a version 1.15.0 cluster to version 1.28.0.

An admin cluster can manage user clusters that are on the same or previous minor version. Managed user clusters can't be more than one minor version lower than the admin cluster, so before upgrading an admin cluster to a new minor version, make sure that all managed user clusters are at the same minor version as the admin cluster.

Node pool versioning rules

When you upgrade node pools selectively, the following version rules apply:

  • Cluster version must be greater than or equal to the worker node pool version.

  • With GKE on Bare Metal minor release 1.28, a worker node pool version can be up to two minor versions behind the cluster (control plane) version. This n-2 version skew is available as a (Preview) capability. This Preview capability is enabled with an annotation in the cluster configuration file. If you don't enable this Preview capability, the maximum version skew between a worker node pool and the cluster is one minor version.

  • Worker node pools can't be at a version that released chronologically later than the cluster version.

    For example, if version 1.16.4 releases on a date after version 1.28.0-gke.425, then you can't upgrade your cluster to version 1.28.0-gke.425 and leave a worker node pool at version 1.16.4. Similarly, if you upgrade your cluster to version 1.28.0-gke.425, but opted to leave a worker node pool at version 1.16.3, you can't later upgrade the worker node pool to version 1.16.4.

The following table lists the supported node pool versions that are allowed for a specific cluster version:

Cluster (control plane) version Supported worker node pool versions
1.28.100-gke.146
  • 1.28.100-gke.146
  • 1.28.0-gke.425
  • 1.16.4
  • 1.16.3
  • 1.16.2
  • 1.16.1
  • 1.16.0
  • 1.15.8
  • 1.15.7
  • 1.15.6
  • 1.15.5
  • 1.15.4
  • 1.15.3
  • 1.15.2
  • 1.15.1
  • 1.15.0
1.28.0-gke.425
  • 1.28.0-gke.425
  • 1.16.3
  • 1.16.2
  • 1.16.1
  • 1.16.0
  • 1.15.6
  • 1.15.5
  • 1.15.4
  • 1.15.3
  • 1.15.2
  • 1.15.1
  • 1.15.0

     

     

     

Upgrade components

Components are upgraded at both the node and the cluster the level. At the cluster level, the following components are upgraded:

  • Cluster components for networking, observability, and storage.
  • For admin, hybrid, and standalone clusters, the lifecycle controllers.
  • The gke-connect-agent.

Nodes in a cluster run as one of the following roles, with different components upgraded depending on the node's role:

Role of the node Function Components to upgrade
Worker Runs user workloads Kubelet, container runtime (Docker or containerd)
Control plane Runs the Kubernetes control plane, cluster lifecycle controllers, and Google Kubernetes Engine (GKE) Enterprise edition platform add-ons Kubernetes control plane static Pods (kubeapi-server, kube-scheduler, kube-controller-manager, etcd)

Lifecycle controllers like lifecycle-controllers-manager and anthos-cluster-operator

Google Kubernetes Engine (GKE) Enterprise edition platform add-ons like stackdriver-log-aggregator and gke-connect-agent
Control plane load balancer Runs HAProxy and Keepalived that serve traffic to kube-apiserver, and run MetalLB speakers to claim virtual IP addresses Control plane load balancer static Pods (HAProxy, Keepalived)

MetalLB speakers

Downtime expectation

The following table details the expected downtime and potential impact when you upgrade clusters. This table assumes you have multiple cluster nodes and an HA control plane. If you run a standalone cluster or don't have an HA control plane, expect additional downtime. Unless noted, this downtime applies to both admin and user cluster upgrades:

Components Downtime expectations When downtime happens
Kubernetes control plane API server (kube-apiserver), etcd, and scheduler No downtime N/A
Lifecycle controllers and ansible-runner job (admin cluster only) No downtime N/A
Kubernetes control plane loadbalancer-haproxy and keepalived Transient downtime (less than 1 to 2 minutes) when the load balancer redirects traffic. Start of the upgrade process.
Observability pipeline-stackdriver and metrics-server Operator drained and upgraded. Downtime should be less than 5 minutes.

DaemonSets continue to work with no downtime.
After control plane nodes finish upgrading.
Container network interface (CNI) No downtime for existing networking routes.

DaemonSet deployed two by two with no downtime.

Operator is drained and upgraded. Downtime less than 5 minutes.
After control plane nodes finish upgrading.
MetalLB (user cluster only) Operator drained and upgraded. Downtime is less than 5 minutes.

No downtime for existing service
After control plane nodes finish upgrading.
CoreDNS and DNS autoscaler (user cluster only) CoreDNS has multiple replicas with autoscaler. Usually no downtime. After control plane nodes finish upgrading.
Local volume provisioner No downtime for existing provisioned persistent volumes (PVs).

Operator might have 5 minutes downtime.
After control plane nodes finish upgrading.
Istio / ingress Istio operator is drained and upgraded. About 5 minutes of downtime.

Existing configured ingress continue to work.
After control plane nodes finish upgrading.
Other system operators 5 minutes downtime when drained and upgraded. After control plane nodes finish upgrading.
User workloads Depends on the setup, such as if highly available.

Review your own workload deployments to understand potential impact.
When the worker node(s) are upgraded.

User cluster upgrade details

This section details the order of component upgrades and status information for a user cluster upgrade. The following section details deviations from this flow for admin, hybrid, or standalone cluster upgrades.

The following diagram shows preflight check process for a user cluster upgrade:

The cluster preflight check runs additional health checks on the cluster before the upgrade process starts.

The preceding diagram details the steps that happen during an upgrade:

  • The bmctl upgrade cluster command creates a PreflightCheck custom resource.
  • This preflight check runs additional checks such as cluster upgrade checks, network health checks, and node health checks.
  • The results of these additional checks combine to report on the ability for the cluster to successfully upgrade to the target version.

If the preflight checks are successful and there are no blocking issues, the components in the cluster are upgraded in a specified order, as shown in the following diagram:

The control plane load balancers and control plane node pool and upgraded, then GKE connect, cluster add-ons, and the load balancer node pool and worker node pools are upgraded.

In the preceding diagram, components are upgraded in order as follows:

  1. The upgrade starts by updating the spec.anthosBareMetalVersion field.
  2. The control plane load balancers are upgraded.
  3. The control plane node pool is upgraded.
  4. In parallel, GKE connect is upgraded, cluster add-ons are upgraded, and the load balancer node pool is upgraded.
    1. After the load balancer node pool is successfully upgraded, the worker node pools are upgraded.
  5. When all components have upgraded, cluster health checks run.

    The health checks continue to run until all checks pass.

  6. When all health checks pass, the upgrade is finished.

Each component has its own status field inside the Cluster custom resource. You can check the status in these fields to understand the progress of the upgrade:

Sequence Field name Meaning
1 status.controlPlaneNodepoolStatus Status is copied from the control plane node pool status. The field includes the versions of the nodes of control plane node pools
2 status.anthosBareMetalLifecycleControllersManifestsVersion Version of lifecycles-controllers-manager applied to the cluster. This field is only available for admin, standalone, or hybrid clusters.
2 status.anthosBareMetalManifestsVersion Version of the cluster from the last applied manifest.
2 status.controlPlaneLoadBalancerNodepoolStatus Status is copied from the control plane load balancer node pool status. This field is empty if no separate control plane load balancer is specified in Cluster.Spec.
3 status.anthosBareMetalVersions An aggregated version map of version to node numbers.
4 status.anthosBareMetalVersion Final status of the upgraded version.

Admin, hybrid, and standalone cluster upgrade details

Starting with bmctl version 1.15.0, the default upgrade behavior for self-managed (admin, hybrid, or standalone) clusters is an in-place upgrade. That is, when you upgrade a cluster to version 1.15.0 or higher, the upgrade uses lifecycle controllers, instead of a bootstrap cluster, to manage the entire upgrade process. This change simplifies the process and reduces resource requirements, which makes cluster upgrades more reliable and scalable.

Although using a bootstrap cluster for upgrading isn't recommended, the option is still available. To use a bootstrap cluster when you upgrade, run the bmctl upgrade command with the --use-bootstrap=true flag. The stages of the upgrade are different, depending on which method you use.

In-place upgrades

The default, in-place upgrade process for self-managed clusters is similar to the user cluster upgrade process. However, when you use the in-place upgrade process, a new version of the preflightcheck-operator is deployed before the cluster preflight check and health checks run:

A new version of the preflightcheck-operator is deployed before the cluster preflight check runs additional health checks on the cluster.

Like the user cluster upgrade, the upgrade process starts by updating the Cluster.spec.anthosBareMetalVersion field to the target version. Two additional steps run before components are updated, as shown in the following diagram: the lifecycle-controller-manager upgrades itself to the target version, and then deploys the target version of anthos-cluster-operator. This anthos-cluster-operator performs the remaining steps of the upgrade process:

The lifecycle-controller-manager and anthos-cluster-operator are deployed before the rest of the cluster is upgraded in the same order as the components in the user cluster.

Upon success, the anthos-cluster-operator reconciles the target version from spec.anthosBareMetalVersion to status.anthosBareMetalVersion.

Upgrade with a bootstrap cluster

The process to upgrade an admin, hybrid, or standalone cluster is similar to a user cluster discussed in the previous section.

The main difference is that the bmctl upgrade cluster command starts a process to create a bootstrap cluster. This bootstrap cluster is a temporary cluster that manages the hybrid, admin, or standalone cluster during an upgrade.

The process to transfer the management ownership of the cluster to the bootstrap cluster is called a pivot. The rest of the upgrade follows the same process as the user cluster upgrade.

During the upgrade process, the resources in the target cluster remain stale. The upgrade progress is only reflected in the resources of the bootstrap cluster.

If needed, you can access the bootstrap cluster to help monitor and debug the upgrade process. The bootstrap cluster can be accessed through bmctl-workspace/.kindkubeconfig.

To transfer the management ownership of the cluster back after the upgrade is complete, the cluster pivots the resources from the bootstrap cluster to the upgraded cluster. There are no manual steps you perform to pivot the cluster during the upgrade process. The bootstrap cluster is deleted after the cluster upgrade succeeds.

Node draining

GKE on Bare Metal cluster upgrades might lead to application disruption as the nodes are drained. This draining process causes all Pods that run on a node to shut down and restart on remaining nodes in the cluster.

Deployments can be used to tolerate such disruption. A Deployment can specify multiple replicas of an application or service should run. An application with multiple replicas should experience little to no disruption during upgrades.

Pod disruption budgets (PDBs)

Pod disruption budgets (PDBs) can be used to ensure that a defined number of replicas always run in the cluster under normal running conditions. PDBs let you limit the disruption to a workload when its Pods need to be rescheduled. However, GKE on Bare Metal doesn't honor PDBs when nodes drain during an upgrade. Instead, the node draining process is best effort. Some Pods might get stuck in a Terminating state and refuse to vacate the node. The upgrade proceeds, even with stuck Pods, when the draining process on a node takes more than 20 minutes.

What's next