Consolidation

Scaling out infrastructure is only one side of the equation for operating compute infrastructure in a cost-effective manner. We also need to be able to optimize on an on-going basis such that, for example, workloads running on under-utilized compute instances are compacted to fewer instances. This improves the overall efficiency of how we run workloads on the compute, resulting in less overhead and lower costs.

Karpenter offers two main ways this can be accomplished:

1. Leverage the ttlSecondsUntilExpired property of Provisioners so that instances are regularly recycled, which will result in compaction of workloads indirectly
2. Since v0.15 its possible to use the Consolidation feature, which will actively attempt to compact under-utilized workloads

We'll be focusing on option 2 in this lab, and to demonstrate we'll be performing these steps:

1. Adjust the Provisioner created in the previous section to enable consolidation
2. Scale the inflate workload from 5 to 12 replicas, triggering Karpenter to provision additional capacity
3. Scale down the workload back down to 5 replicas
4. Observe Karpenter consolidating the compute

Now, let's update the Provisioner to enable consolidation:

Kustomize Patch

~/environment/eks-workshop/modules/autoscaling/compute/karpenter/consolidation/provisioner.yaml

apiVersion: karpenter.sh/v1alpha5
kind: Provisioner
metadata:
  name: default
spec:
  consolidation:
    enabled: true

Provisioner/default

apiVersion: karpenter.sh/v1alpha5
kind: Provisioner
metadata:
  name: default
spec:
  consolidation:
    enabled: true
  labels:
    type: karpenter
  limits:
    resources:
      cpu: 1000
      memory: 1000Gi
  providerRef:
    name: default
  requirements:
    - key: karpenter.sh/capacity-type
      operator: In
      values:
        - on-demand
    - key: node.kubernetes.io/instance-type
      operator: In
      values:
        - c5.large
        - m5.large
        - m5.xlarge

Diff

 kind: Provisioner
 metadata:
   name: default
 spec:
+  consolidation:
+    enabled: true
   labels:
     type: karpenter
   limits:
     resources:

Let's apply this update:

~$kubectl apply -k ~/environment/eks-workshop/modules/autoscaling/compute/karpenter/consolidation

Now, let's scale our inflate workload again to consume more resources:

~$kubectl scale -n other deployment/inflate --replicas 12

~$kubectl rollout status -n other deployment/inflate --timeout=180s

This changes the total memory request for this deployment to around 12Gi, which when adjusted to account for the roughly 600Mi reserved for the kubelet on each node means that this will fit on 2 instances of type m5.large:

~$kubectl get nodes -l type=karpenter

Next, scale the number of replicas back down to 5:

~$kubectl scale -n other deployment/inflate --replicas 5

We can check the Karpenter logs to get an idea of what actions it took in response to our scaling in the deployment:

~$kubectl -n karpenter logs deployment/karpenter -c controller | grep 'deprovisioning via consolidation delete' -A 2

The output will show Karpenter identifying specific nodes to cordon, drain and then terminate:

2023-07-20T22:06:33.926Z        INFO    controller.deprovisioning       deprovisioning via consolidation delete, terminating 1 nodes ip-10-42-159-233.us-west-2.compute.internal/m5.large/on-demand  {"commit": "5a7faa0-dirty"}
2023-07-20T22:06:33.984Z        INFO    controller.termination  cordoned node   {"commit": "5a7faa0-dirty", "node": "ip-10-42-159-233.us-west-2.compute.internal"}
2023-07-20T22:06:34.263Z        INFO    controller.termination  deleted node    {"commit": "5a7faa0-dirty", "node": "ip-10-42-159-233.us-west-2.compute.internal"}

This will result in the Kubernetes scheduler placing any Pods on those nodes on the remaining capacity, and now we can see that Karpenter is managing a total of 1 node:

~$kubectl get nodes -l type=karpenter

Karpenter can also further consolidate if a node can be replaced with a cheaper variant in response to workload changes. This can be demonstrated by scaling the inflate deployment replicas down to 1, with a total memory request of around 1Gi:

~$kubectl scale -n other deployment/inflate --replicas 1

We can check the Karpenter logs and see what actions the controller took in response:

~$kubectl -n karpenter logs deployment/karpenter -c controller | grep 'deprovisioning via consolidation replace' -A 2

The output will show Karpenter consolidating via replace, replacing the m5.large node with the cheaper c5.large instance type defined in the Provisioner:

2023-07-20T22:08:54.965Z        INFO    controller.deprovisioning       deprovisioning via consolidation replace, terminating 1 nodes ip-10-42-83-198.us-west-2.compute.internal/m5.large/on-demand and replacing with on-demand node from types c5.large   {"commit": "5a7faa0-dirty"}
2023-07-20T22:08:54.980Z        INFO    controller.deprovisioning       launching node with 1 pods requesting {"cpu":"125m","memory":"1Gi","pods":"3"} from types c5.large  {"commit": "5a7faa0-dirty", "provisioner": "default"}
2023-07-20T22:08:55.229Z        DEBUG   controller.deprovisioning.cloudprovider discovered launch template      {"commit": "5a7faa0-dirty", "provisioner": "default", "launch-template-name": "Karpenter-eks-workshop-16555401392435391284"}

Since the total memory request with 1 replica is much lower around 1Gi, it would be more efficient to run it on the cheaper c5.large instance type with 4GB of memory. Once the node is replaced, we can check the metadata on the new node and confirm the instance type is the c5.large:

~$kubectl get nodes -l type=karpenter -o jsonpath="{range .items[*]}{.metadata.labels.node\.kubernetes\.io/instance-type}{'\n'}{end}"

c5.large