EKS Karpenter Spot Instances Autoscaling: Complete 70% Cost Savings Guide (2025)

Step 4: Create NodePool Configuration (The Core of Karpenter)

NodePool is a Karpenter custom resource that defines what nodes to provision and when to consolidate them.

4.1 Create Basic Production NodePool

# nodepool-default.yaml – Production-grade configuration
apiVersion: karpenter.sh/v1beta1
kind: NodePool
metadata:
name: default
spec:
template:
metadata:
labels:
workload-type: general
spec:
nodeClassRef:
name: default
requirements:
# Capacity type: Spot (cheap) + On-Demand (fallback)
– key: karpenter.sh/capacity-type
operator: In
values: [“spot”, “on-demand”]
# Instance families: c (compute), m (memory balanced), r (memory)
– key: karpenter.k8s.aws/instance-family
operator: In
values: [“c6i”, “c6a”, “m6i”, “m6a”, “t3”]
# Instance sizes: small to 2xlarge
– key: karpenter.k8s.aws/instance-size
operator: In
values: [“medium”, “large”, “xlarge”, “2xlarge”]
# CPU: 2-16 cores
– key: kubernetes.io/arch
operator: In
values: [“amd64”]
# Zone: Spread across multiple AZs
– key: topology.kubernetes.io/zone
operator: In
values: [“us-west-2a”, “us-west-2b”, “us-west-2c”]
# Exclude graviton (for compatibility)
– key: karpenter.k8s.aws/instance-cpu
operator: Gt
values: [“0”]
taints:
– key: karpenter.sh/lifecycle
value: spot
effect: NoSchedule
limits:
cpu: 1000
memory: 1000Gi
disruption:
# Consolidation removes underutilized nodes
consolidationPolicy: WhenUnderutilized
consolidateAfter: 30s
# Expire nodes after 30 days (force refresh)
expireAfter: 720h
# Disruption budgets prevent too many simultaneous node terminations
budgets:
– nodes: “20%”
reasons:
– “Underutilized”
– “Empty”
duration: “5m”
– nodes: “10%”
reasons:
– “Drifted”
duration: “10m”
– nodes: “15%”
schedule: “0 9 * * mon-fri”
reasons:
– “Drifted”
—
# EC2NodeClass: AWS-specific node configuration
apiVersion: karpenter.k8s.aws/v1beta1
kind: EC2NodeClass
metadata:
name: default
spec:
amiFamily: AL2
role: “KarpenterNodeRole-${CLUSTER_NAME}”
subnetSelectorTerms:
– tags:
karpenter.sh/discovery: “${CLUSTER_NAME}”
securityGroupSelectorTerms:
– tags:
karpenter.sh/discovery: “${CLUSTER_NAME}”
blockDeviceMappings:
– deviceName: /dev/xvda
ebs:
volumeSize: 100Gi
volumeType: gp3
deleteOnTermination: true
encrypted: true
userData: |
#!/bin/bash
echo “Karpenter node launched”
tags:
ManagedBy: Karpenter
Environment: Production
—
# PriorityClass: Ensure system pods aren’t evicted
apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
name: system-cluster-critical
value: 2000000000
globalDefault: false
description: “System critical pods”

4.2 Deploy NodePool

# Replace cluster name in YAML
sed -i “s/\${CLUSTER_NAME}/$CLUSTER_NAME/g” nodepool-default.yaml

# Apply configuration
kubectl apply -f nodepool-default.yaml

# Verify NodePool created
kubectl get nodepools
# Output: default ec2 true

# Check EC2NodeClass
kubectl get ec2nodeclass
# Output: default true

4.3 Create Dedicated GPU NodePool (For ML/AI Workloads)

# nodepool-gpu.yaml – GPU nodes for machine learning
apiVersion: karpenter.sh/v1beta1
kind: NodePool
metadata:
name: gpu
spec:
template:
metadata:
labels:
workload-type: gpu
spec:
nodeClassRef:
name: gpu
requirements:
– key: karpenter.sh/capacity-type
operator: In
values: [“spot”] # GPU Spot is 70% cheaper
– key: karpenter.k8s.aws/instance-family
operator: In
values: [“g4dn”, “g5”, “g6”] # NVIDIA GPU instances
– key: karpenter.k8s.aws/instance-gpu-count
operator: Gt
values: [“0”] # At least 1 GPU
– key: karpenter.k8s.aws/instance-size
operator: In
values: [“xlarge”, “2xlarge”]
limits:
cpu: 200
memory: 200Gi
disruption:
consolidationPolicy: WhenUnderutilized
budgets:
– nodes: “5%” # Strict budget for GPU (expensive)
—
apiVersion: karpenter.k8s.aws/v1beta1
kind: EC2NodeClass
metadata:
name: gpu
spec:
amiFamily: AL2_NVIDIA
role: “KarpenterNodeRole-${CLUSTER_NAME}”
subnetSelectorTerms:
– tags:
karpenter.sh/discovery: “${CLUSTER_NAME}”
userData: |
#!/bin/bash
# GPU driver installation handled by AL2_NVIDIA AMI
echo “GPU node ready”

# Apply GPU NodePool
kubectl apply -f nodepool-gpu.yaml