Cluster Autoscaling

Scales the number of nodes in our cluster based off usage metrics Documentation

Understanding Resources

In this example, I'll be focusing on CPU for scaling.
We need to ensure we have an understanding of the compute resources we have.

How many cores do we have
How many cores do our application use

I go into more details about pod resource utilisation in the Horizontal Pod Autoscaler guide.

We need a Kubernetes cluster with Cluster Autoscaler

# azure example

NAME=aks-getting-started
RESOURCEGROUP=aks-getting-started
SERVICE_PRINCIPAL=
SERVICE_PRINCIPAL_SECRET=

az aks create -n $NAME \
--resource-group $RESOURCEGROUP \
--location australiaeast \
--kubernetes-version 1.16.10 \
--nodepool-name default \
--node-count 1 \
--node-vm-size Standard_F4s_v2  \
--node-osdisk-size 250 \
--service-principal $SERVICE_PRINCIPAL \
--client-secret $SERVICE_PRINCIPAL_SECRET \
--output none \
--enable-cluster-autoscaler \
--min-count 1 \
--max-count 5

Deploy Metric Server

Metric Server provides container resource metrics for use in autoscaling pipelines

We will need to deploy Metric Server 0.3.7
I used components.yamlfrom the release page link above.

Note: For Demo clusters (like kind), you will need to disable TLS
You can disable TLS by adding the following to the metrics-server container args

- --kubelet-insecure-tls
- --kubelet-preferred-address-types="InternalIP"

Deploy it:

cd kubernetes\autoscaling
kubectl -n kube-system apply -f .\metric-server\metricserver-0.3.7.yaml

#test 
kubectl -n kube-system get pods

#wait for metrics to populate
kubectl top nodes

Example App

We have an app that simulates CPU usage

# build

cd kubernetes\autoscaling\application-cpu
docker build . -t aimvector/application-cpu:v1.0.0

# push
docker push aimvector/application-cpu:v1.0.0

# resource requirements
resources:
  requests:
    memory: "50Mi"
    cpu: "500m"
  limits:
    memory: "500Mi"
    cpu: "2000m"

# deploy 
kubectl apply -f deployment.yaml

# metrics
kubectl top pods

Generate some CPU load

# Deploy a tester to run traffic from

cd kubernetes/autoscaling
kubectl apply -f ./autoscaler-cluster/tester.yaml

# get a terminal
kubectl exec -it tester sh
# install wrk
apk add --no-cache wrk curl

# simulate some load
wrk -c 5 -t 5 -d 99999 -H "Connection: Close" http://application-cpu

# scale and keep checking `kubectl top`
# every time we add a pod, CPU load per pod should drop dramatically.
# roughly 8 pods will have each pod use +- 400m

kubectl scale deploy/application-cpu --replicas 2

Deploy an autoscaler

# scale the deployment back down to 2
kubectl scale deploy/application-cpu --replicas 2

# deploy the autoscaler
kubectl autoscale deploy/application-cpu --cpu-percent=95 --min=1 --max=10

# pods should scale to roughly 7-8 to match criteria

kubectl describe hpa/application-cpu 
kubectl get hpa/application-cpu  -owide

3.3 KiB Raw Blame History