Cluster Autoscaler Demo¶

Now that we have the Cluster Autoscaler deployed in our EKS cluster, let's see it in action.

Docker Images¶

Here is the Docker Image used in this tutorial: reyanshkharga/nodeapp:v1

Objective¶

Currently, we have 2 t3.medium nodes in private-nodegroup. We also have min(2) and max(4) nodes set for the nodegroup.

# List nodes
kubectl get nodes

We'll follow these steps to test the Cluster Autoscaler:

1. We'll create a deployment with 2 replicas, each requesting 1 core CPU.
2. We'll scale the deployment to add more replicas.
3. We'll observe Cluster Autoscaler taking autoscaling actions to accommodate pending pods that couldn't be scheduled due to insufficient resources.

Let's see this in action!

Step 1: Create a Deployment¶

First, let's create a deployment as follows:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

apiVersion: apps/v1
kind: Deployment
metadata:
  name: my-deployment
spec:
  replicas: 2
  selector:
    matchLabels:
      app: nginx
  template:
    metadata:
      labels:
        app: nginx
    spec:
      containers:
      - name: nginx
        image: nginx
        resources:
          requests:
            cpu: "1"

Apply the manifest to create the deployment:

kubectl apply -f my-deployment.yml

Verify deployment and pods:

# List deployments
kubectl get deployments

# List pods
kubectl get pods

Note that each pod requests 1 core CPU. Also, we already have some workload running in the kube-system namespace. In the best case, the two nodes can accommodate two pods created by my-deployment.

You will observe that both pods are scheduled and running without any issues.

Step 2: Update the Deployment¶

Now, let's update the deployment by setting the replicas to 4.

# Apply the modified deployment
kubectl apply -f deployment.yml

# List pods
kubectl get pods

You might notice the new pods getting stuck in the Pending state due to the lack of CPU resources.

Describe the pod to determine why it is still in a Pending state:

kubectl describe pod <pod-name>

At this point, the Cluster Autoscaler will assess the autoscaling needs and launch additional instances in the node group.

Once the new nodes are up and running, the new pods can be scheduled.

List nodes and pods in watch mode:

# List nodes
kubectl get nodes -w

# List pods
kubectl get pods -w

You can view the autoscaling events in the cluster-autoscaler log:

# View autoscaler logs
kubectl logs -f deployment.apps/cluster-autoscaler -n kube-system

You can also use the following command to see kubernetes events:

kubectl events

Step 3: Delete the Deployment¶

Let's delete the deployment:

kubectl delete -f deployment.yml

When the resources are freed you will notice that the autoscaling scales down the instances in the autoscaling group.

You can view the autoscaling events in the cluster-autoscaler log:

kubectl logs -f deployment.apps/cluster-autoscaler -n kube-system