Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
You might want to change the size of your virtual machines (VMs) to accommodate an increasing number of deployments or to run a larger workload. Always resize node pools through the AKS API. Don't resize the underlying Virtual Machine Scale Sets directly, as outlined in the support policies for AKS:
AKS agent nodes appear in the Azure portal as regular Azure IaaS resources. But these virtual machines are deployed into a custom Azure resource group (usually prefixed with MC_*). You can't make direct customizations to these nodes using the IaaS APIs or resources. Any custom changes that aren't done via the AKS API won't persist through an upgrade, scale, update, or reboot.
Important
This article applies to Virtual Machine Scale Sets-based node pools. When using Virtual Machines-based node pools, you can update the VM sizes in an existing node pool by using a single Azure CLI command and have multiple VM sizes in the same node pool. For more information, see the Virtual Machines node pools documentation.
Choose a resize method
We recommend that you resize the node pool in place with a single command. If in-place resize doesn't support your scenario, create a new node pool with the target VM size and move your workloads to it instead.
Option 1: Resize a node pool in place
When you change the VM size (SKU) of an existing node pool by using az aks nodepool update --node-vm-size <new-size>, AKS performs a rolling upgrade by:
- Surging new nodes with the target VM size.
- Cordoning and draining the old nodes.
- Deleting the old nodes.
How the resize rollout works
The resize rollout uses the same rolling upgrade engine as a node image upgrade and a Kubernetes version upgrade, so it honors the following upgrade-related settings already configured on the node pool. In particular, the resize respects:
- Max surge (
--max-surge): Controls how many extra nodes with the target VM size are added during the rollout. A higher value resizes the pool faster but consumes more compute and IP quota; a lower value is slower but less disruptive. The AKS default is1, and33%is recommended for production node pools. - Node drain timeout (
--drain-timeout): How long AKS waits for pod eviction on each old node before forcefully deleting it. The default is 30 minutes. Combine this with appropriate PodDisruptionBudgets so workloads can drain safely. - Node soak duration (
--node-soak-duration): How long AKS waits after a new node becomes Ready before moving on to the next batch. Useful for letting workloads stabilize on the new VM size before continuing the rollout.
Prerequisites
- Azure CLI version 2.93.0 or later installed and configured. Run
az --versionto find the version. If you need to install or upgrade, see Install Azure CLI. - AKS API version
2026-02-01or later. - Because resize reuses the max surge upgrade pipeline, the same prerequisites apply:
- Ensure your subscription has enough target VM size replacement capacity and available subnet IPs for surge nodes.
- Your PodDisruptionBudgets allow at least one replica to be evicted at a time, otherwise resize can be blocked during drain.
- For end-to-end recommendations, see Best practices for AKS node pool upgrades.
Resize the node pool
Use the az aks nodepool update command with the --node-vm-size parameter to change the VM size of an existing VMSS-based node pool:
az aks nodepool update \
--resource-group MyResourceGroup \
--cluster-name MyManagedCluster \
--name nodepool1 \
--node-vm-size Standard_D4s_v3
Note
In-place resize requires surge capacity to provision new nodes with the target VM size before draining the old ones. If the node pool is configured with --max-surge 0 (that is, --max-unavailable is in effect), the resize request is rejected with a 400 Bad Request. To proceed, set --max-surge to at least 1 when resize using
az aks nodepool update \
--resource-group MyResourceGroup \
--cluster-name MyManagedCluster \
--name nodepool1 \
--node-vm-size Standard_D4s_v3 \
--max-surge 33%
and optionally restore your original --max-surge and --max-unavailable values after the resize completes.
Understand if in-place node pool resize is supported
In-place resize only supports specific scenarios. It depends on your node pool configuration, your operating system, and whether the current and target VM sizes use the same disk controller type (SCSI or NVMe). AKS validates rejected requests before changing any nodes, so the node pool stays on its current VM size.
The following changes are not supported as part of an in-place VMSS resize. Use Option 2: Resize by creating a new node pool:
- Changing the CPU architecture (for example, x64 to ARM64).
- Changing confidential computing support (for example, enabling or disabling SNP).
- Changing OS disk type (for example, resizing a node pool that uses an ephemeral OS disk to a VM size that can't host the OS disk in the same placement).
- Combining the resize with a Kubernetes version upgrade or a node count change in the same operation.
Resize across disk controller types and hypervisor generations
When the target VM size requires a different disk controller or hypervisor generation than the node pool currently uses, AKS updates the disk controller. If needed, AKS switches the node pool to the Gen2 version of its node image as part of the same rolling upgrade.
Note
Disk controller type migration is available only to the node pools described in Node pool eligibility for disk controller migration.
The following table summarizes how AKS handles resizes that cross disk controller types or hypervisor generations:
| Current node pool | Target VM size | Result |
|---|---|---|
Gen2 node image on a SCSI size (for example, Standard_D4ds_v5) |
Gen2, NVMe-only size (for example, Standard_D4s_v6) |
Supported for eligible node pools. New nodes use the NVMe disk controller. The node image doesn't change. Other node pools fail with DiskControllerTypeIncompatible. |
Gen1 node image on a size that supports both Gen1 and Gen2, with SCSI (for example, Standard_D4ds_v5) |
Gen2, NVMe-only size (for example, Standard_D4s_v6) |
Supported for eligible node pools. AKS switches the node pool to the Gen2 version of the same OS image (for example, AKSUbuntu-2404gen2containerd) and to the NVMe disk controller. Other node pools fail with DiskControllerTypeIncompatible. |
Gen1 node image on a Gen1-only size (for example, Standard_D3_v2) |
Gen2, NVMe-only size (for example, Standard_D4s_v6) |
Not supported in a single step. Use a two-step resize. |
| Gen1 node image on a Gen1-only size | Size that supports both Gen1 and Gen2 | Supported (standard resize). The node pool keeps its Gen1 node image. |
| SCSI size | Size that supports both SCSI and NVMe (for example, Standard_E4bds_v5) |
Supported (standard resize). The node pool keeps the SCSI disk controller. |
| NVMe size | SCSI-only size | Not supported. The request fails with DiskControllerTypeIncompatible. |
| Gen2 node image | Gen1-only size | Not supported. The request fails with HypervisorGenerationIncompatible, or with DiskControllerTypeIncompatible if the current size is NVMe. |
When AKS switches a node pool from Gen1 to Gen2, the node pool keeps its OS SKU, OS version, and Kubernetes version. Only the node image generation, firmware (BIOS to UEFI), and disk controller change.
Node pool eligibility for disk controller migration
Important
The following requirements apply only to resizes that change the disk controller type. Node pools that don't meet them can still use standard resize between VM sizes that have the same disk controller type. To move these node pools to an NVMe-only size, create a new node pool instead.
To resize from a SCSI size to an NVMe-only size, with or without a Gen1-to-Gen2 switch, the node pool must meet all of the following requirements:
Operating system: The node pool uses the standard Gen1 or Gen2 node image for Ubuntu 22.04, Ubuntu 24.04, or Azure Linux 3.0. Other OS SKUs and image variants aren't supported, including Windows, Azure Container Linux, Ubuntu 26.04, and FIPS-enabled, Trusted Launch, confidential VM, Arm64, and Kata (pod sandboxing) node images.
Node pool configuration: The node pool and cluster must be in the
Runningstate, and the node pool must not:- Use Spot VMs.
- Be associated with a capacity reservation group.
- Use the blue-green upgrade strategy.
- Be a Static Egress Gateway node pool.
- Be created from a node pool snapshot.
- Be deployed to an Azure Extended Zone.
VM sizes: Neither the current nor the target VM size can be a GPU-enabled or confidential computing size (for example, the
DC,EC, orNPseries).Request: In addition to the restrictions that apply to every in-place resize, the request can't also change autoscaler settings, the node image version, availability zones, node labels, VM tags, NSG settings (allowedHostPorts, applicationSecurityGroups), or custom CA trust.
Resize from a Gen1-only size to an NVMe-only size
Important
You can't resize a node pool directly from a Gen1-only size (for example, Standard_D3_v2) to a Gen2, NVMe-only size (for example, Standard_D4s_v6). Resize in two steps instead. The second step is a disk controller migration, so the node pool must meet the eligibility requirements.
Resize the node pool to a SCSI size that supports both Gen1 and Gen2, such as
Standard_DS3_v2. The node pool keeps its Gen1 node image. This step is a standard resize.az aks nodepool update \ --resource-group MyResourceGroup \ --cluster-name MyManagedCluster \ --name nodepool1 \ --node-vm-size Standard_DS3_v2After the first resize finishes, resize the node pool to the NVMe-only size. AKS switches the node pool to the Gen2 node image and the NVMe disk controller.
az aks nodepool update \ --resource-group MyResourceGroup \ --cluster-name MyManagedCluster \ --name nodepool1 \ --node-vm-size Standard_D4s_v6
Note
Because the OS disk on NVMe-based VMs is exposed as an NVMe device (for example, /dev/nvme0n1) instead of a SCSI device (for example, /dev/sda), review any workloads, DaemonSets, or scripts that reference device paths directly before you resize.
Option 2: Resize by creating a new node pool
Use this method when in-place resize doesn't support your scenario. You create a new node pool with the target VM size, move your workloads to it by cordoning and draining the existing nodes, and then delete the existing node pool.
Create a new node pool with the desired SKU
Note
Every AKS cluster must contain at least one system node pool with at least one node. In this example, we use a --mode of System to add a system node pool to replace the system node pool we want to resize. You can update the mode of a node pool at any time. You can also add a user node pool by setting --mode to User.
When resizing, make sure you consider all workload requirements, such as availability zones, and configure your VMSS node pool accordingly. You might need to modify the following command to best fit your needs. For a full list of the configuration options, see the az aks nodepool add reference page.
Create a new node pool using the
az aks nodepool addcommand. In this example, we create a new node pool,mynodepool, with three nodes and theStandard_DS3_v2VM SKU to replace an existing node pool,nodepool1, that has theStandard_DS2_v2VM SKU.az aks nodepool add \ --resource-group myResourceGroup \ --cluster-name myAKSCluster \ --name mynodepool \ --node-count 3 \ --node-vm-size Standard_DS3_v2 \ --mode System \ --no-waitIt takes a few minutes for the new node pool to be created.
Get the status of the new node pool using the
kubectl get nodescommand.kubectl get nodesYour output should resemble the following example output, showing both the new node pool
mynodepooland the existing node poolnodepool1:NAME STATUS ROLES AGE VERSION aks-mynodepool-98765432-vmss000000 Ready agent 23m v1.21.9 aks-mynodepool-98765432-vmss000001 Ready agent 23m v1.21.9 aks-mynodepool-98765432-vmss000002 Ready agent 23m v1.21.9 aks-nodepool1-12345678-vmss000000 Ready agent 10d v1.21.9 aks-nodepool1-12345678-vmss000001 Ready agent 10d v1.21.9 aks-nodepool1-12345678-vmss000002 Ready agent 10d v1.21.9
Cordon the existing nodes
Cordoning marks specified nodes as unschedulable and prevents any more pods from being added to the nodes.
Get the names of the nodes you want to cordon using the
kubectl get nodescommand.kubectl get nodesYour output should resemble the following example output, showing the nodes in the existing node pool
nodepool1that you want to cordon:NAME STATUS ROLES AGE VERSION aks-nodepool1-12345678-vmss000000 Ready agent 7d21h v1.21.9 aks-nodepool1-12345678-vmss000001 Ready agent 7d21h v1.21.9 aks-nodepool1-12345678-vmss000002 Ready agent 7d21h v1.21.9Cordon the existing nodes using the
kubectl cordoncommand, specifying the desired nodes in a space-separated list. For example:kubectl cordon aks-nodepool1-12345678-vmss000000 aks-nodepool1-12345678-vmss000001 aks-nodepool1-12345678-vmss000002Your output should resemble the following example output, showing that the nodes are cordoned:
node/aks-nodepool1-12345678-vmss000000 cordoned node/aks-nodepool1-12345678-vmss000001 cordoned node/aks-nodepool1-12345678-vmss000002 cordoned
Drain the existing nodes
Important
To successfully drain nodes and evict running pods, ensure that any PodDisruptionBudgets (PDBs) allow for at least one pod replica to be moved at a time. Otherwise, the drain/evict operation fails. To check this, you can run kubectl get pdb -A and verify ALLOWED DISRUPTIONS is at least 1 or higher.
When you drain nodes, the pods running on them are evicted and recreated on the other schedulable nodes.
Drain the existing nodes using the
kubectl draincommand with the--ignore-daemonsetsand--delete-emptydir-dataflags, specifying the desired nodes in a space-separated list. For example:Important
Using
--delete-emptydir-datais required to evict the AKS-createdcorednsandmetrics-serverpods. If you don't use this flag, you get an error. For more information, see the documentation on emptydir.kubectl drain aks-nodepool1-12345678-vmss000000 aks-nodepool1-12345678-vmss000001 aks-nodepool1-12345678-vmss000002 --ignore-daemonsets --delete-emptydir-dataAfter the drain operation finishes, all pods (excluding the pods controlled by daemon sets) should be running on the new node pool. You can verify this using the
kubectl get podscommand.kubectl get pods -o wide -A
Troubleshoot pod eviction issues
You might encounter the following error when draining nodes:
Error when evicting pods/[podname] -n [namespace] (will retry after 5s): Cannot evict pod as it would violate the pod's disruption budget.
By default, your cluster has AKS-managed pod disruption budgets (such as coredns-pdb or konnectivity-agent) with a MinAvailable of 1. For example, if there are two coredns pods running, only one can be disrupted at a time. While one of them is getting recreated and is unavailable, the other coredns pod can't be evicted due to the pod disruption budget. This issue resolves itself after the initial coredns pod is scheduled and running, allowing the second pod to be properly evicted and recreated.
Tip
Consider draining nodes one by one for a smoother eviction experience and to avoid throttling. For more information, see:
Remove the existing node pool
Important
When you delete a node pool, AKS doesn't perform cordon and drain. To minimize the disruption of rescheduling pods currently running on the node pool you plan to delete, perform a cordon and drain on all nodes in the node pool before deleting.
Delete the original node pool using the
az aks nodepool deletecommand.az aks nodepool delete \ --resource-group myResourceGroup \ --cluster-name myAKSCluster \ --name nodepool1Verify that your AKS cluster has only the new node pool with the applications and pods properly running using the
kubectl get nodescommand.kubectl get nodesYour output should resemble the following example output, showing only the new node pool
mynodepool:NAME STATUS ROLES AGE VERSION aks-mynodepool-98765432-vmss000000 Ready agent 63m v1.21.9 aks-mynodepool-98765432-vmss000001 Ready agent 63m v1.21.9 aks-mynodepool-98765432-vmss000002 Ready agent 63m v1.21.9
Next steps
- Learn more about using multiple node pools.
- Configure upgrade settings for node pools, such as max surge and drain timeout, which in-place resize also uses.