Karpenter Provision

/karpenter-provision

Purpose

Generate production-ready Karpenter configurations (NodePool + EC2NodeClass) optimized for cost, reliability, and the specific workload characteristics.

Requirement Gathering

Ask the user for these details (if not already provided):

1. Workload type: stateless API, stateful (database/cache), batch/CI, GPU/ML, mixed
2. Resources per pod: typical vCPU and memory requests
3. Capacity type: on-demand only, spot only, or mixed (specify ratio)
4. Monthly budget target: helps size limits appropriately
5. Region and AZs: for subnet and zone configuration
6. EKS version and Karpenter version: for API compatibility
7. Special requirements: GPU, ARM64, high-memory, Windows, etc.

Generation Rules

Instance Type Selection

• Stateless APIs: c6i, c7g, m6i, m7g (compute/general purpose, good price-performance)
• Memory-intensive: r6i, r7g, r6a (memory optimized)
• Batch/CI: c6i, c6a, c7g, c7a, m6i, m6a, m7g (maximize diversity for Spot)
• GPU: p4d, p5, g5, g6 (match GPU type to workload: training vs inference)
• ARM64 preferred: Include Graviton variants (c7g, m7g, r7g) — 20-40% cheaper

Instance Size Selection

• Match to pod resource requests: if pods need 2 vCPU/4GB, use large to 2xlarge
• Avoid oversized instances (4xlarge+) unless pods are large or bin-packing is critical
• Include at least 3 sizes for flexibility

Spot Strategy

• Minimum 15 instance types for Spot-to-Spot consolidation
• Always include on-demand fallback for mixed strategies
• Use weight field: Spot NodePool weight=100, On-Demand weight=10 (fallback)
• Configure interruption handling (SQS + EventBridge)

Disruption Policy

• Production: consolidateAfter: 60s, budgets: nodes: "20%"
• Batch: consolidateAfter: 30s, budgets: nodes: "40%"
• Add peak-hours protection budget for production
• Use WhenEmptyOrUnderutilized (default, recommended for most cases)

Limits

• Set based on budget (rough formula: $0.04/vCPU-hour on-demand, $0.015 spot)
• Include both cpu and memory limits
• Leave 20% headroom above expected peak

Output

Generate complete YAML files:

1. NodePool(s) with:

   • Instance family and size requirements
   • Capacity type (on-demand/spot)
   • AZ requirements
   • Labels and taints (for workload isolation)
   • Disruption policy with budgets
   • Resource limits
   • Weight (priority)
   • expireAfter (TTL for node rotation)

2. EC2NodeClass with:

   • amiSelectorTerms (alias for EKS-optimized AMI)
   • subnetSelectorTerms (tag-based discovery)
   • securityGroupSelectorTerms (tag-based discovery)
   • instanceProfile
   • blockDeviceMappings (root volume sizing)

3. Supporting resources:

   • PodDisruptionBudget for the primary workload
   • Recommended Prometheus ServiceMonitor config for Karpenter

Validation Checklist

Before presenting the configuration, verify:

• Instance type diversity: ≥15 types for Spot, ≥6 for On-Demand
• All 3 AZs included for HA (if multi-AZ requested)
• Resource limits align with budget target
• consolidateAfter is appropriate (not 0s in production)
• Peak-hours budget protection included for production workloads
• Taints configured for workload isolation (if multiple NodePools)
• expireAfter set for node rotation (720h prod, 168h batch recommended)
• amiSelectorTerms uses alias (not custom AMI) unless specifically required
• No launch template references (not supported in v1 API)

Examples

Common configurations the skill should handle:

• "Provision for a microservices platform with 50 stateless APIs"
• "Set up Karpenter for ML training workloads with p5.48xlarge GPUs"
• "Configure Spot-only NodePool for CI/CD runners"
• "Mixed strategy: critical payment service on On-Demand, everything else on Spot"
• "ARM64-first strategy with x86 fallback for cost optimization"