The cluster autoscaler on Azure dynamically scales Kubernetes worker nodes. It runs as a deployment in your cluster.

This README will help you get cluster autoscaler running on your Azure Kubernetes cluster.

Kubernetes v1.10.x or later is required to use cluster autoscaler on Azure. See the "Releases" section in the README for more information.

Cluster autoscaler v1.2.x or later is required for Azure. See the "Releases" section in the README for more information.

NOTE: In the deployment manifests referenced below, be sure to replace the {{ ca_version }} placeholder with an actual release, such as v1.14.2.

Get Azure credentials by running the following Azure CLI command:

# replace <subscription-id> with yours.
az ad sp create-for-rbac --role="Contributor" --scopes="/subscriptions/<subscription-id>" --output json

This will create a new service principal with "Contributor" role scoped to your subscription. Save the JSON output, because it will be needed to configure the cluster autoscaler deployment in the next step.

If you are using nodeSelector, you need to tag the VMSS with a node-template key "k8s.io_cluster-autoscaler_node-template_label_" for using labels and "k8s.io_cluster-autoscaler_node-template_taint_" if you are using taints.

Note that these tags use the pipe _ character compared to a forward slash due to Azure tag name restrictions.

To add the label of foo=bar to a node from a VMSS pool, you would add the following tag to the VMSS k8s.io_cluster-autoscaler_node-template_label_foo: bar.

You can also use forward slashes in the labels by setting them as an underscore in the tag name. For example to add the label of k8s.io/foo=bar to a node from a VMSS pool, you would add the following tag to the VMSS k8s.io_cluster-autoscaler_node-template_label_k8s.io_foo: bar. To encode a tag name containing an underscore, use "2" (eg. "cpu2arch" gives "cpu_arch").

To add the taint of foo=bar:NoSchedule to a node from a VMSS pool, you would add the following tag to the VMSS k8s.io_cluster-autoscaler_node-template_taint_foo: bar:NoSchedule.

You can also use forward slashes in taints by setting them as an underscore in the tag name. For example to add the taint of k8s.io/foo=bar:NoSchedule to a node from a VMSS pool, you would add the following tag to the VMSS k8s.io_cluster-autoscaler_node-template_taint_k8s.io_foo: bar:NoSchedule. To encode a taint name containing an underscore, use "~2".

When scaling from an empty VM Scale Set (0 instances), Cluster Autoscaler will evaluate the provided resources (cpu, memory, ephemeral-storage) based on that VM Scale Set's backing instance type. This can be overridden (for instance, to account for system reserved resources) by specifying capacities with VMSS tags, formated as: k8s.io_cluster-autoscaler_node-template_resources_<resource name>: <resource value>. For instance:

k8s.io_cluster-autoscaler_node-template_resources_cpu: 3800m
k8s.io_cluster-autoscaler_node-template_resources_memory: 11Gi

NOTE: GPU autoscaling consideration on VMSS : In case of scale set of GPU nodes, kubelet node label accelerator have to be added to node provisionned to make GPU scaling works.

Some autoscaling options can be defined per VM Scale Set, with tags. Those tags values have the format as the respective cluster-autoscaler flags they override: floats or durations encoded as strings.

Supported options tags (with example values) are:

# overrides --scale-down-utilization-threshold global value for that specific VM Scale Set
k8s.io_cluster-autoscaler_node-template_autoscaling-options_scaledownutilizationthreshold: "0.5"

# overrides --scale-down-gpu-utilization-threshold global value for that specific VM Scale Set
k8s.io_cluster-autoscaler_node-template_autoscaling-options_scaledowngpuutilizationthreshold: "0.5"

# overrides --scale-down-unneeded-time global value for that specific VM Scale Set
k8s.io_cluster-autoscaler_node-template_autoscaling-options_scaledownunneededtime: "10m0s"

# overrides --scale-down-unready-time global value for that specific VM Scale Set
k8s.io_cluster-autoscaler_node-template_autoscaling-options_scaledownunreadytime: "20m0s"

Cluster autoscaler supports four Kubernetes cluster options on Azure:

• vmss: Autoscale VMSS instances by setting the Azure cloud provider's vmType parameter to vmss or to an empty string. This supports clusters deployed with aks-engine.
• standard: Autoscale VMAS (Virtual Machine Availability Set) VMs by setting the Azure cloud provider's vmType parameter to standard. This supports clusters deployed with aks-engine.

NOTE: only the vmss option supports scaling down to zero nodes.

NOTE: The subscriptionID parameter is optional. When skipped, the subscription will be fetched from the instance metadata.

Prerequisites:

• Get Azure credentials from the Permissions step above.
• Get the name of the VM scale set associated with the cluster's node pool. You can find this in the Azure Portal or with the az vmss list command.

Make a copy of cluster-autoscaler-vmss.yaml. Fill in the placeholder values for the cluster-autoscaler-azure secret data by base64-encoding each of your Azure credential fields.

• ClientID: <base64-encoded-client-id>
• ClientSecret: <base64-encoded-client-secret>
• ResourceGroup: <base64-encoded-resource-group>
• SubscriptionID: <base64-encoded-subscription-id>
• TenantID: <base64-encoded-tenant-id>

NOTE: Use a command such as echo $CLIENT_ID | base64 to encode each of the fields above.

In the cluster-autoscaler spec, find the image: field and replace {{ ca_version }} with a specific cluster autoscaler release.

To run a cluster-autoscaler which auto-discovers VMSSs with nodes use the --node-group-auto-discovery flag. For example, --node-group-auto-discovery=label:cluster-autoscaler-enabled=true,cluster-autoscaler-name=<YOUR CLUSTER NAME> will find the VMSSs tagged with those tags containing those values.

Note that:

• It is recommended to use a second tag like cluster-autoscaler-name=<YOUR CLUSTER NAME> when cluster-autoscaler-enabled=true is used across many clusters to prevent VMSSs from different clusters recognized as the node groups
• There are no --nodes flags passed to cluster-autoscaler because the node groups are automatically discovered by tags
• No min/max values are provided when using Auto-Discovery, cluster-autoscaler will detect the "min" and "max" tags on the VMSS resource in Azure, adjusting the desired number of nodes within these limits.

kubectl apply -f examples/cluster-autoscaler-autodiscover.yaml

Below that, in the command: section, update the --nodes= arguments to reference your node limits and VMSS name. For example, if node pool "k8s-nodepool-1-vmss" should scale from 1 to 10 nodes:

        - --nodes=1:10:k8s-nodepool-1-vmss

or to autoscale multiple VM scale sets:

        - --nodes=1:10:k8s-nodepool-1-vmss
        - --nodes=1:10:k8s-nodepool-2-vmss

Note that it doesn't mean the number of nodes in nodepool is restricted in the range from 1 to 10. It means when ca is downscaling (upscaling) the nodepool, it will never break the limit of 1 (10). If the current node pool size is lower than the specified minimum or greater than the specified maximum when you enable autoscaling, the autoscaler waits to take effect until a new node is needed in the node pool or until a node can be safely deleted from the node pool.

To allow scaling similar node pools simultaneously, or when using separate node groups per zone and to keep nodes balanced across zones, use the --balance-similar-node-groups flag (default false). Add it to the command section to enable it:

        - --balance-similar-node-groups=true

See the FAQ for more details.

Save the updated deployment manifest, then deploy cluster-autoscaler by running:

kubectl create -f cluster-autoscaler-vmss.yaml

To run a cluster autoscaler pod on a control plane (previously referred to as master) node, the deployment should tolerate the master taint, and nodeSelector should be used to schedule pods. Use cluster-autoscaler-vmss-control-plane.yaml in this case.

To run a cluster autoscaler pod with Azure managed service identity (MSI), use cluster-autoscaler-vmss-msi.yaml instead.

Azure has hard limits on the number of read and write requests against Azure APIs per subscription, per region. Running lots of clusters in a single subscription, or running a single large, dynamic cluster in a subscription can produce side effects that exceed the number of calls permitted within a given time window for a particular category of requests. See the following documents for more detail on Azure API throttling in general:

• https://docs.microsoft.com/en-us/azure/azure-resource-manager/management/request-limits-and-throttling
• https://docs.microsoft.com/en-us/azure/virtual-machines/troubleshooting/troubleshooting-throttling-errors

Given the dynamic nature of cluster autoscaler, it can be a trigger for hitting those rate limits on the subscriptions. This in turn can affect other components running in the cluster that depend on Azure APIs such as kube-controller-manager.

When using K8s versions older than v1.18, we recommend using at least v.1.17.5, v1.16.9, v1.15.12 which include various improvements on the cloud-provider side that have an impact on the number of API calls during scale down operations.

As for CA versions older than 1.18, we recommend using at least v.1.17.2, v1.16.5, v1.15.6.

In addition, cluster-autoscaler exposes a AZURE_VMSS_CACHE_TTL environment variable which controls the rate of GetVMScaleSet being made. By default, this is 15 seconds but setting this to a higher value such as 60 seconds can protect against API throttling. The caches used are proactively incremented and decremented with the scale up and down operations and this higher value doesn't have any noticeable impact on performance. Note that the value is in seconds

The AZURE_VMSS_VMS_CACHE_TTL environment variable affects the GetScaleSetVms (VMSS VM List) calls rate. The default value is 300 seconds. A configurable jitter (AZURE_VMSS_VMS_CACHE_JITTER environment variable, default 0) expresses the maximum number of second that will be subtracted from that initial VMSS cache TTL after a new VMSS is discovered by the cluster-autoscaler: this can prevent a dogpile effect on clusters having many VMSS.

The AZURE_ENABLE_DYNAMIC_INSTANCE_LIST environment variable enables workflow that fetched SKU information dynamically using SKU API calls. By default, it uses static list of SKUs.

The AZURE_ENABLE_VMSS_FLEX environment variable enables VMSS Flex support. By default, support is disabled.

When using K8s 1.18 or higher, it is also recommended to configure backoff and retries on the client as described here

Prerequisites:

• Get Azure credentials from the Permissions step above.
• Get the name of the initial Azure deployment resource for the cluster. You can find this in the Azure Portal or with the az deployment list command. If there are multiple deployments, get the name of the first one.

Make a copy of cluster-autoscaler-standard-control-plane.yaml. Fill in the placeholder values for the cluster-autoscaler-azure secret data by base64-encoding each of your Azure credential fields.

• ClientID: <base64-encoded-client-id>
• ClientSecret: <base64-encoded-client-secret>
• ResourceGroup: <base64-encoded-resource-group>
• SubscriptionID: <base64-encoded-subscription-id>
• TenantID: <base64-encoded-tenant-id>
• Deployment: <base64-encoded-azure-initial-deployment-name>

NOTE: Use a command such as echo $CLIENT_ID | base64 to encode each of the fields above.

In the cluster-autoscaler spec, find the image: field and replace {{ ca_version }} with a specific cluster autoscaler release.

Below that, in the command: section, update the --nodes= arguments to reference your node limits and node pool name (tips: node pool name is NOT availability set name, e.g., the corresponding node pool name of the availability set agentpool1-availabilitySet-xxxxxxxx would be agentpool1). For example, if node pool "k8s-nodepool-1" should scale from 1 to 10 nodes:

        - --nodes=1:10:k8s-nodepool-1

or to autoscale multiple VM scale sets:

        - --nodes=1:10:k8s-nodepool-1
        - --nodes=1:10:k8s-nodepool-2

Create the Azure deploy parameters secret cluster-autoscaler-azure-deploy-parameters by running:

kubectl -n kube-system create secret generic cluster-autoscaler-azure-deploy-parameters --from-file=deploy-parameters=./_output/<your-output-path>/azuredeploy.parameters.json

Then deploy cluster-autoscaler by running:

kubectl create -f cluster-autoscaler-standard-control-plane.yaml

To run a cluster autoscaler pod with Azure managed service identity (MSI), use cluster-autoscaler-standard-msi.yaml instead.

WARNING: Cluster autoscaler depends on user-provided deployment parameters to provision new nodes. After upgrading your Kubernetes cluster, cluster autoscaler must also be redeployed with new parameters to prevent provisioning nodes with an old version.

Node Pool Autoscaling is a first class feature of your AKS cluster. The option to enable cluster autoscaler is available in the Azure Portal or with the Azure CLI:

az aks create \
  --resource-group myResourceGroup \
  --name myAKSCluster \
  --kubernetes-version 1.25.11 \
  --node-count 1 \
  --enable-cluster-autoscaler \
  --min-count 1 \
  --max-count 3

Please see the AKS autoscaler documentation for details.

The new version of Azure client supports rate limit and back-off retries when the cluster hits the throttling issue. These can be set by either environment variables, or cloud config file. With config file, defaults values are false or 0.

NOTE: * These rate limit configs can be set per-client. Customizing QPS and Bucket through environment variables per client is not supported.