Upgrade options for Azure Kubernetes Service (AKS) clusters
Article
This article covers the different upgrade options for AKS clusters. To perform a basic Kubernetes version upgrade, see Upgrade an AKS cluster.
For AKS clusters that use multiple node pools or Windows Server nodes, see Upgrade a node pool in AKS. To upgrade a specific node pool without performing a Kubernetes cluster upgrade, see Upgrade a specific node pool.
Perform manual upgrades
You can perform manual upgrades to control when your cluster upgrades to a new Kubernetes version. Manual upgrades are useful when you want to test a new Kubernetes version before upgrading your production cluster. You can also use manual upgrades to upgrade your cluster to a specific Kubernetes version that isn't the latest available version.
To perform manual upgrades, see the following articles:
You can configure automatic upgrades to automatically upgrade your cluster to the latest available Kubernetes version. Automatic upgrades are useful when you want to ensure your cluster is always running the latest Kubernetes version. You can also use automatic upgrades to ensure your cluster is always running a supported Kubernetes version.
To configure automatic upgrades, see the following articles:
Special considerations for node pools that span multiple availability zones
AKS uses best-effort zone balancing in node groups. During an upgrade surge, the zones for the surge nodes in Virtual Machine Scale Sets are unknown ahead of time, which can temporarily cause an unbalanced zone configuration during an upgrade. However, AKS deletes surge nodes once the upgrade completes and preserves the original zone balance. If you want to keep your zones balanced during upgrades, you can increase the surge to a multiple of three nodes, and Virtual Machine Scale Sets balances your nodes across availability zones with best-effort zone balancing. With best-effort zone balance, the scale set attempts to scale in and out while maintaining balance. However, if for some reason this isn't possible (for example, if one zone goes down, the scale set can't create a new VM in that zone), the scale set allows temporary imbalance to successfully scale in or out.
Persistent volume claims (PVCs) backed by Azure locally redundant storage (LRS) Disks are bound to a particular zone and might fail to recover immediately if the surge node doesn't match the zone of the PVC. If the zones don't match, it can cause downtime on your application when the upgrade operation continues to drain nodes but the PVs are bound to a zone. To handle this case and maintain high availability, configure a Pod Disruption Budget on your application to allow Kubernetes to respect your availability requirements during the drain operation.
Optimize for undrainable node behavior (Preview)
You can configure the upgrade process behavior for drain failures. The default upgrade behavior is Schedule, which consists of a node drain failure causing the upgrade operation to fail, leaving the undrained nodes in a schedulable state. Alternatively, you can select the Cordon behavior, which skips nodes that fail to drain by placing them in a quarantined state, labels them kubernetes.azure.com/upgrade-status:Quarantined, and proceeds with upgrading the remaining nodes. This behavior ensures that all nodes are either upgraded or quarantined. This approach allows you to troubleshoot drain failures and gracefully manage the quarantined nodes.
How do I set the new Cordon behavior?
Use CLI preview and install aks-preview extension 9.0.0b3 or later.
You can use the following commands to update or install aks-preview extension:
az extension update --name aks-preview
az extension add --name aks-preview
Update the node pool undrainable node behavior to Cordon:
az aks nodepool update --cluster-name $CLUSTER_NAME --name $NODE_POOL_NAME --resource-group $RESOURCE_GROUP --max-surge 1 --undrainable-node-behavior Cordon
The following example output shows the undrainable node behavior updated:
Verify the label on any blocked nodes. When there's a drain node failure on upgrade using the following command:
kubectl get nodes --show-labels=true
The blocked nodes are unscheduled for pods and marked with the label "kubernetes.azure.com/upgrade-status: Quarantined". The maximum number of nodes that can be left blocked can't be more than the Max-Surge value.
How do I remove the blocked nodes?
First resolve the issue causing the drain. The following example removes the responsible PDB:
Then delete the blocked node using the az aks nodepool delete-machines command. This command is useful if you intend to reduce the node pool footprint by removing nodes left behind in older versions.
az aks nodepool delete-machines --cluster-name MyCluster --machine-names aks-nodepool1-test123-vmss000000 --name nodepool1 --resource-group TestRG
After you complete this step, you can reconcile the cluster status by performing any update operation without the optional fields as outlined here.
Example command:
az aks update --resource-group TestRG --name MyCluster
Alternatively, you can scale the node pool to the same number of nodes as the count of upgraded nodes. This action ensures the node pool gets to its intended original size. AKS prioritizes the removal of the blocked nodes. This command also restores the cluster provisioning status to Succeeded. In the example given, 2 is the total number of upgraded nodes.
az aks nodepool scale --resource-group TestRG --cluster-name MyCluster --name nodepool1 --node-count 2
Optimize upgrades to improve performance and minimize disruptions
Planned Maintenance Window enables service teams to schedule auto-upgrade during a predefined window, typically a low-traffic period, to minimize workload impact. We recommend a window duration of at least four hours.
Max Surge on the node pool allows requesting extra quota during the upgrade process and limits the number of nodes selected for upgrade simultaneously. A higher max surge results in a faster upgrade process. We don't recommend setting it at 100%, as it upgrades all nodes simultaneously, which can cause disruptions to running applications. We recommend a max surge quota of 33% for production node pools.
Pod Disruption Budget is set for service applications and limits the number of pods that can be down during voluntary disruptions, such as AKS-controlled node upgrades. It can be configured as minAvailable replicas, indicating the minimum number of application pods that need to be active, or maxUnavailable replicas, indicating the maximum number of application pods that can be terminated, ensuring high availability for the application. Refer to the guidance provided for configuring Pod Disruption Budgets (PDBs). PDB values should be validated to determine the settings that work best for your specific service.
Node drain timeout on the node pool allows you to configure the wait duration for eviction of pods and graceful termination per node during an upgrade. This option is useful when dealing with long running workloads. When the node drain timeout is specified (in minutes), AKS respects waiting on pod disruption budgets. If not specified, the default timeout is 30 minutes.
Node soak time helps stagger node upgrades in a controlled manner and can minimize application downtime during an upgrade. You can specify a wait time, preferably as reasonably close to 0 minutes as possible, to check application readiness between node upgrades. If not specified, the default value is 0 minutes. Node soak time works together with the max surge and node drain timeout properties available in the node pool to deliver the right outcomes in terms of upgrade speed and application availability.
Next steps
This article listed different upgrade options for AKS clusters. For a detailed discussion of upgrade best practices and other considerations, see AKS patch and upgrade guidance.
Collaborate with us on GitHub
The source for this content can be found on GitHub, where you can also create and review issues and pull requests. For more information, see our contributor guide.
Azure Kubernetes Service feedback
Azure Kubernetes Service is an open source project. Select a link to provide feedback: