Kubernetes scaling strategies

This project helps you create a cluster with idle nodes ready to deploy new workloads.

Getting started

You need to create a Linode token to access the API:

linode-cli profile token-create
export LINODE_TOKEN=<insert the token here>

# Create the cluster
terraform -chdir=01-clusters init
terraform -chdir=01-clusters apply -auto-approve

# Label the node, install KEDA + Prometheus
terraform -chdir=02-setup init
terraform -chdir=02-setup apply -auto-approve

# Scale to 2 nodes
linode-cli lke pool-update <cluster id> <pool id> --count 2 # cluster 1

# # Disable scaling down (you can't label nodes in LKE)
# kubectl annotate node <node1> cluster-autoscaler.kubernetes.io/scale-down-disabled=true
# kubectl annotate node <node1> cluster-autoscaler.kubernetes.io/scale-down-disabled=true

# Cleanup
terraform -chdir=02-setup destroy -auto-approve
terraform -chdir=01-clusters destroy -auto-approve

Why scale to 3 nodes after the first step?

This is to ensure that all LKE's controllers end up in the same node, and you can uniquely tag that node.

Demo the cron scaling

Make sure that your kubectl is configured with the current kubeconfig file:

export KUBECONFIG="${PWD}/kubeconfig"

The execute:

kubectl apply -f 03-demo-cron/01-podinfo.yaml

Amend the time to be 1 minute in the future.

Submit the ScaledObject:

kubectl apply -f 03-demo-cron/02-scaled-object.yaml

Dashboard

kubectl proxy --www=./dashboard

Placeholder demo

Make sure that your kubectl is configured with the current kubeconfig file:

export KUBECONFIG="${PWD}/kubeconfig"

Then execute:

kubectl apply -f 04-demo-proactive/01-podinfo.yaml

When ready, observe the node scaling up with:

kubectl scale deployment/podinfo --replicas=5

Or use the dashboard.

The total scaling time for the 5th pod should take ~2m.

Scale back to 1 and submit the placeholder pod with:

kubectl scale deployment/podinfo --replicas=1
kubectl apply -f 04-demo-proactive/02-placeholder.yaml

Click on the scale button to scale up to 5. Then show the new node is being provisioned.

kubectl proxy --www=./dashboard --www-prefix=/static &
open http://localhost:8001/static

Repeat the experiment. The total scaling time should go down to ~10s.

HPA demo

Make sure that your kubectl is configured with the current kubeconfig file:

export KUBECONFIG="${PWD}/kubeconfig-hpa"

Then execute:

kubectl apply -f 05-demo-hpa/01-rate-limiter.yaml
kubectl apply -f 05-demo-hpa/03-scaled-object.yaml

At this point, you should have at least two nodes. One has four podinfo pods.

Scale your workers to three instances so that you have an empty node.

Retrieve the IP address of the load balancer for Locust with:

kubectl get service locust -o jsonpath='{.status.loadBalancer.ingress[0].ip}'

Drive traffic to the instance:

• Number of users: 300
• Concurrent users: 0.4
• Url: http://podinfo

Repeat the experiment.

kubectl delete -f 05-demo-hpa/03-scaled-object.yaml
kubectl scale deployment/podinfo --replicas=1

Wait for the Cluster autoscaler to drain the nodes. By the end, you should have only 2.

Submit the placeholder pod:

kubectl apply -f 05-demo-hpa/placeholder.yaml

The pod stays Pending until the cluster autoscaler creates the third node.

Open Locust and drive the traffic to the deployment with:

• Number of users: 300
• Concurrent users: 0.4
• Url: http://podinfo

Building the rate limiter

cd app
docker build -t learnk8s/rate-limiter:1.0.0 .
docker push learnk8s/rate-limiter:1.0.0