Configure multiple node pools by using AKS spot node pools with the cluster autoscaler

Azure provides Azure Virtual Machine instances that offer scalability while reducing costs, and are ideal for workloads that can be interrupted. However, these virtual machines (VMs) access unused Azure compute capacity at lower prices but still support high-performance computing scenarios.

Your company's drone-tracking solution is deployed on Azure Kubernetes Service (AKS) as many containerized applications and services. One of these services is a batch-processing service that schedules drone flight paths. With a sudden growth in your customer base, the batch-processing service gets inundated with requests and builds up a backlog of deliveries. This situation is causing delays and customer frustration.

Automatically scaling the number of batch-processing service replicas provides for timely order processing. However, it also requires you to deploy more nodes to keep up with computing resource needs. Analyzing usage trends in Azure Monitor, you notice that these nodes are used only at specific times and not in a cost-effective way. The batch-processing service is stateless and doesn't save any client-session data. You realize that you can save money by:

• Using lower-cost node instances.
• Automatically scaling the node count in the node pool that's configured for batch processing.

Let's look at the infrastructure that underlies this cost-saving solution in AKS.

What is a spot virtual machine (spot VM) in Azure?

A spot virtual machine is a VM that gives you access to unused Azure compute capacity at deep discounts. Spot VMs replace the existing, low-priority VMs in Azure. You can use spot VMs to run workloads that include:

• High-performance computing scenarios, batch processing, or visual-rendering applications.

• Large-scale, stateless applications.

• Developer/test environments, including continuous integration (CI) and continuous delivery (CD) workloads.

Spot VM availability

Spot VM availability depends on factors such as capacity, size, region, and time of day. Azure allocates VMs only if capacity is available. As a result, there's no service-level agreement (SLA) for these types of VMs, and they offer no high-availability guarantees.

Spot VM eviction policy

The default eviction policy for spot VMs is Deallocate. Azure evicts spot VMs with 30 seconds of notice when capacity in a region becomes limited. A VM that's set with the Deallocate policy moves to the stopped-deallocated state when evicted. You can redeploy an evicted VM when spot capacity becomes available again. A deallocated VM is still counted toward your spot virtual CPU (vCPU) quota, and charges for the underlying allocated disks still apply.

What is a spot virtual machine scale set?

A spot virtual machine scale set is a virtual machine scale set that supports Azure spot VMs. These VMs behave the same way as normal spot VMs, but with one difference: when you use virtual machine scale set support for spot VMs in Azure, you choose between two eviction policies:

• Deallocate: The Deallocate policy functions exactly as described earlier.

• Delete: The Delete policy allows you to avoid the cost of disks and hitting quota limits. With the Delete eviction policy, evicted VMs are deleted together with their underlying disks. The scale set's autoscaling feature can now automatically try to compensate for VM eviction by creating new VMs. Although the creation of VMs isn't guaranteed, the evicted VMs don't count toward your vCPU quota or incur costs for underlying disks.

  A best practice is to use the autoscale feature only when you set the eviction policy to Delete on the scale set.

What is a spot node pool in Azure Kubernetes Service (AKS)?

A spot node pool is a user node pool that uses a spot virtual machine scale set. AKS supports spot VMs when you:

• Need to create user node pools.
• Want the cost benefits offered by virtual machine scale set support for Azure spot VMs.

Use spot node pools to:

• Take advantage of unused capacity in Azure.
• Use scale set features with the Delete eviction policy.
• Define the maximum price you want to pay per hour.
• Enable the recommended AKS Kubernetes cluster autoscaler when using spot node pools.

For example, to support the drone-tracking application's batch-processing service, you can create a spot user node pool and enable the cluster autoscaler. You can then configure the horizontal pod scaler to deploy more batch-processing services to match resource demands.

As the demand for nodes increases, the cluster autoscaler can scale the number of nodes up and down in the spot node pool. If node evictions happen, the cluster autoscaler keeps trying to scale up the node count if extra nodes are still needed.

Spot node pool limitations

Before you decide to add a spot user node pool to your AKS cluster, consider the following limitations:

• The underlying spot scale set is deployed only to a single fault domain and offers no high-availability guarantees.
• The AKS cluster needs multiple node-pool support to be enabled.
• You can use spot node pools only as user node pools.
• You can't upgrade spot node pools.
• The creation of spot VMs isn't guaranteed. The creation of spot nodes depends on capacity and quota availability in the cluster's deployed Azure region.

Remember that spot node pools should be used only for workloads that can be interrupted.

Add a spot node pool to an AKS cluster

A spot node pool can't be a system node pool for an AKS cluster. First, you need to create your cluster and then use the az aks nodepool add command to add a new user node pool.

You set several parameters for a new node pool to configure it as a spot node pool.

Priority

The --priority parameter is set to Regular by default for a new node pool. Set the value to Spot to indicate that the new pool you're creating is a spot node pool. This value can't be changed after creation.

Eviction policy

A spot node pool must use a virtual machine scale set. Recall from earlier that the spot node pool uses a spot scale set. Set --eviction-policy to Delete to allow the scale set to remove both the node and the underlying, allocated disk the node uses. You can't change this value after creation.

You can set the eviction policy to Deallocate, but when these nodes are evicted, they still count against your compute quota towards scaling or upgrading the cluster.

Maximum price for spot node

Spot node pools optimize costs by capping the maximum amount that you're willing to pay per spot node per hour. To set your safe amount, use the --spot-max-price parameter. Newly created spot nodes are evicted when this value is reached.

You can set this value to any positive amount up to five decimal places, or set it to -1. Setting the --spot-max-price value to -1 affects your node pool in the following ways:

• Nodes aren't be evicted based on the node's price.
• The cost for new nodes is based on the current price for spot nodes, or the price for a standard node, using whichever is lower.

For example, if you set the value to 0.98765, the maximum price for a node in USD is 0.98765 per hour. When the node's consumption exceeds this amount, it's evicted.

Enable the cluster autoscaler

We recommend that you enable the cluster autoscaler by using the --enable-cluster-autoscaler parameter. If you don't use the cluster autoscaler, you risk the node count dropping to zero in the node pool as nodes are evicted because of Azure capacity constraints.

Minimum node count

Set the minimum node count to a value between 1 and 100 by using the --min-count parameter. A minimum node count is required when you enable the cluster autoscaler.

Maximum node count

Set the maximum node count to a value between 1 and 100 by using the --max-count parameter. A maximum node count is required when you enable the cluster autoscaler.

Sample Configuration

Here's an example az aks nodepool add command that adds a spot node pool with a max count of 3 and min count of 1. Notice the use of --enable-cluster-autoscaler to enable the spot node features.

az aks nodepool add \
    --resource-group resourceGroup \
    --cluster-name aksCluster \
    --name spotpool01 \
    --enable-cluster-autoscaler \
    --max-count 3 \
    --min-count 1 \
    --priority Spot \
    --eviction-policy Delete \
    --spot-max-price -1 \
    --no-wait

Deploy pods to spot node pools

When deploying workloads in Kubernetes, you can provide information to the scheduler to specify which nodes the workloads can or can't run. You control workload scheduling by configuring taints, toleration, or node affinity. Spot nodes are configured with a specific label and taint.

What is a taint?

A taint is applied to a node to indicate that only specific pods can be scheduled on it. Spot nodes are configured with a label set to kubernetes.azure.com/scalesetpriority:spot.

What is toleration?

Toleration is a specification applied to a pod to allow, but not require, a pod to be scheduled on a node with the corresponding taint. Spot nodes are configured with a node taint set to kubernetes.azure.com/scalesetpriority=spot:NoSchedule.

What is node affinity?

You use node affinity to describe which pods are scheduled on a node. Affinity is specified by using labels defined on the node. For example, in AKS, system pods are configured with anti-affinity towards spot nodes to prevent the pods from being scheduled on these nodes.

Define toleration in a pod manifest file

You specify node-taint toleration by creating a tolerations dictionary entry in your workload manifest file. In this dictionary, you set the following properties for each node taint the workload has to tolerate in this section:

Property	Description
key	Identifies a node taint key-value pair specified on the node. For example, on a spot node pool, the key-value pair is kubernetes.azure.com/scalesetpriority:spot. The key is kubernetes.azure.com/scalesetpriority.
operator	Allows the toleration to match a taint. The default operator is Equal. You can also specify Exists to match toleration. However, when you use Exists, you don't specify the following property (value).
value	Represents the value part of the node-taint key-value pair that's specified on the node. For example, on a spot node pool with a key-value pair of kubernetes.azure.com/scalesetpriority:spot, the value is spot.
effect	Indicates how the scheduling of a pod is handled in the system. There are three options: NoSchedule, PreferNoSchedule, and NoExecute. NoSchedule ensures that the system can't schedule the pod. PreferNoSchedule allows the system to try not to schedule the pod. NoExecute either evicts pods that are already running on the tainted node or doesn't schedule the pod at all.

Define node affinity in a pod manifest file

You specify affinity by creating an affinity entry in your workload manifest file. In this entry, you set the following properties for each node label that a workload must match:

Property	Description
nodeAffinity	Describes node affinity scheduling rules for the pod.
requiredDuringSchedulingIgnoredDuringExecution	If the affinity requirements specified by this field aren't met at scheduling time, the pod can't be scheduled onto the node. If the affinity requirements specified by this field cease to be met at some point during pod execution (for example, due to an update), the system can choose to attempt to evict the pod from its node.
nodeSelectorTerms	A list of node selector terms. Terms returned match any of the filters, rather than all of the filters.
matchExpressions	A list of node selector requirements by node's labels.
key	The label key that the selector applies to. The key is kubernetes.azure.com/scalesetpriority
operator	Represents a key's relationship to a set of values. Valid operators are In, NotIn, Exists, DoesNotExist Gt, and Lt
values	Represents the value part of the node label key-value pair that is specified on the node. On a spot node pool with a key-value pair of kubernetes.azure.com/scalesetpriority:spot, the value is spot.

Here's an example of a workload that has toleration and affinity added for spot node pools.

apiVersion: v1
kind: Pod
metadata:
  name: nginx
  labels:
    env: test
spec:
  containers:
  - name: nginx
    image: nginx
    imagePullPolicy: IfNotPresent
  tolerations:
  - key: "kubernetes.azure.com/scalesetpriority"
    operator: "Equal"
    value: "spot"
    effect: "NoSchedule"
  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: "kubernetes.azure.com/scalesetpriority"
            operator: In
            values:
            - "spot"