Karpenter Nodes

• Nodes (VMs) created by Karpenter are Self Managed (managed by the Karpenter Controller).
• Differently from conventional nodes which are usually managed by a Node Group

Conventional Nodes

• Karpenter eliminates the need of Node Groups for the main workloads
• However, you should not run Karpenter controller in the nodes managed by Karpenter itself!
• Therefore you should always have a conventional Nodes to accommodate your controllers (along with other critical components like coredns and metrics server)
• Either by a small Node Group with 2 nodes in different AZs or a Fargate Profile
• You can use node affinity to force Karpenter controller into a specific NodeGroup

affinity:
  nodeAffinity:
    requiredDuringSchedulingIgnoredDuringExecution:
      nodeSelectorTerms:
      - matchExpressions:
        - key: karpenter.sh/nodepool
          operator: DoesNotExist
        - key: eks.amazonaws.com/nodegroup
          operator: In
          values:
          - my-node-group
  podAntiAffinity:
    requiredDuringSchedulingIgnoredDuringExecution:
      - topologyKey: kubernetes.io/hostname

Node Labels

• topology.kubernetes.io/zone: E.g., us-east-2a
• node.kubernetes.io/instance-type: E.g., g4dn.8xlarge
• kubernetes.io/os: E.g., linux
• kubernetes.io/arch: E.g., amd64
• karpenter.sh/capacity-type: E.g., spot
• karpenter.k8s.aws/instance-hypervisor: E.g., nitro
• karpenter.k8s.aws/instance-encryption-in-transit-supported: E.g., true
• karpenter.k8s.aws/instance-category: E.g., g