如何在不删除Azure Kubernetes集群的情况下停止（取消分配）节点？

Got it, let’s work through this problem together. The reason you can only scale down to 1 node is almost certainly because you’re dealing with the system node pool—Azure requires these to have at least one running node to host critical Kubernetes system pods like kube-proxy, coredns, and the controller manager. But we’ve got ways to either eliminate weekend costs entirely or slash them to near-zero, depending on how your cluster is set up.

Option 1: You have separate user node pools (recommended setup) #

If you’ve deployed your application workloads on dedicated user node pools (not the system pool), you can scale those down to 0 completely, and optimize the system pool to minimize its cost.

Step-by-step actions: #

1. Identify your node pools first
   Run this Azure CLI command to list all node pools and spot which one is the system pool (look for the System: Yes entry):

   az aks nodepool list --cluster-name YOUR_CLUSTER_NAME --resource-group YOUR_RG_NAME -o table
   

2. Scale user node pools to 0
   For every user node pool you have, run this command to deallocate all its VMs (you won’t be charged for these over the weekend):

   az aks nodepool scale --cluster-name YOUR_CLUSTER_NAME --resource-group YOUR_RG_NAME --name USER_NODE_POOL_NAME --node-count 0
   

3. Optimize the system node pool (for extra savings)
   If you want to cut costs even more, resize the system pool’s VM to a cheaper SKU (like Standard_B1s or Standard_A1_v2) before locking it to 1 node:

   az aks nodepool upgrade --cluster-name YOUR_CLUSTER_NAME --resource-group YOUR_RG_NAME --name SYSTEM_NODE_POOL_NAME --node-vm-size Standard_B1s --no-wait
   

   Wait for the upgrade to finish, then confirm it’s scaled to 1 node:

   az aks nodepool scale --cluster-name YOUR_CLUSTER_NAME --resource-group YOUR_RG_NAME --name SYSTEM_NODE_POOL_NAME --node-count 1
   

Option 2: You only have a system node pool (no user pools) #

If all your workloads are running on the system node pool, you can’t scale it to 0—but you can stop and deallocate the underlying VMs to avoid VM compute charges. Note: This will make your cluster unavailable over the weekend, so only do this if you don’t need access to it during that time.

Step-by-step actions: #

1. Find the VM scale set tied to your system node pool
   First, grab the scale set ID with this command:

   az aks nodepool show --cluster-name YOUR_CLUSTER_NAME --resource-group YOUR_RG_NAME --name SYSTEM_NODE_POOL_NAME --query 'scaleSetId' -o tsv
   

   Pull the scale set name from the ID (it’ll look something like aks-nodepool1-123456-vmss).
2. Stop and deallocate all VMs in the scale set
   Run these commands to shut down the VMs—this stops charging you for their compute resources:

   az vmss stop --resource-group YOUR_RG_NAME --name SCALE_SET_NAME --no-wait
   az vmss deallocate --resource-group YOUR_RG_NAME --name SCALE_SET_NAME --no-wait
   

3. Restore your cluster on Monday
   Start the VMs back up:

   az vmss start --resource-group YOUR_RG_NAME --name SCALE_SET_NAME --no-wait
   

   If the node doesn’t rejoin the cluster automatically, run this to refresh it:

   az aks nodepool upgrade --cluster-name YOUR_CLUSTER_NAME --resource-group YOUR_RG_NAME --name SYSTEM_NODE_POOL_NAME --node-count 1
   

Bonus: Automate this to skip manual work #

To avoid doing this every weekend, use Azure Automation Runbooks to schedule the scale-down/stop commands for Friday evening, and scale-up/start commands for Monday morning. Here’s a quick breakdown:

• Create an Azure Automation account with a run-as account that has permissions to manage your AKS cluster and VM scale sets.
• Write two runbooks (PowerShell or Bash): one to scale user pools to 0 and stop system pool VMs, another to reverse the process.
• Schedule the runbooks to run at your desired off-hours.

Key things to keep in mind #

• When scaling user pools to 0, make sure your workloads can handle being shut down over the weekend (use persistent storage if you need to retain data).
• Stopping system pool VMs will take your cluster offline—only do this if you don’t need access during off-hours.
• Always test these steps in a non-production cluster first to avoid unexpected issues.

内容的提问来源于stack exchange，提问作者Dario Romero