CAST AI Production Checklist

Overview

Complete checklist for enabling CAST AI cost optimization on a production Kubernetes cluster. Covers Phase 1 (monitoring) through Phase 2 (full automation) with validation steps at each stage.

Prerequisites

• CAST AI tested on a staging cluster first
• Production API key (Full Access)
• Change management approval for node lifecycle changes

Instructions

Complete the phases in order and preserve evidence for every checked item. Start in monitoring-only mode, verify the production cluster identity and baseline metrics, then request the approved change window before enabling automation. Use a two-person review for capacity limits, disruption budgets, and the emergency-disable procedure; do not copy staging keys, policies, or test evidence into the production record without revalidation.

Phase 1: Monitoring Only

• [ ] Agent installed with read-only key
• [ ] Agent pod healthy: kubectl get pods -n castai-agent
• [ ] Console shows cluster as "Connected"
• [ ] Savings report generating (wait 24h for full data)
• [ ] Review savings estimate before enabling automation

Phase 2: Autoscaling Enabled

• [ ] Full Access API key provisioned and stored in secrets manager
• [ ] Cluster controller installed
• [ ] Evictor installed with conservative settings (non-aggressive)
• [ ] Spot handler installed for graceful interruption handling
• [ ] Autoscaler policies configured with appropriate limits:
• [ ] clusterLimits.cpu.maxCores set to safe ceiling
• [ ] unschedulablePods.headroom configured (10-15%)
• [ ] nodeDownscaler.emptyNodes.delaySeconds >= 300 for production
• [ ] spotInstances.spotDiversityEnabled = true
• [ ] Node templates created for workload-specific needs (GPU, high-memory)
• [ ] PodDisruptionBudgets set on all critical workloads

Workload Autoscaler

• [ ] Workload autoscaler installed
• [ ] Critical deployments annotated with min/max resource bounds
• [ ] Anti-shrink cooldown set (300s minimum)
• [ ] Memory headroom >= 20% for production workloads

Security

• [ ] API key in secrets manager (not Helm values files)
• [ ] Kvisor security agent installed
• [ ] Network policies applied to castai-agent namespace
• [ ] RBAC reviewed and minimized
• [ ] Key rotation scheduled (90-day interval)

Monitoring and Alerting

• [ ] Alert on agent pod restarts: kube_pod_container_status_restarts_total{namespace="castai-agent"}
• [ ] Alert on API errors in agent logs
• [ ] CAST AI console email notifications enabled
• [ ] Savings report reviewed weekly
• [ ] Dashboard tracking spot vs on-demand node ratio

Rollback Procedure

# Disable autoscaling immediately (keeps agent monitoring)
curl -X PUT -H "X-API-Key: ${CASTAI_API_KEY}" \
  -H "Content-Type: application/json" \
  "https://api.cast.ai/v1/kubernetes/clusters/${CASTAI_CLUSTER_ID}/policies" \
  -d '{"enabled": false}'

# Or remove all CAST AI components
helm uninstall castai-evictor -n castai-agent
helm uninstall cluster-controller -n castai-agent
# Keep the agent for monitoring if desired

Error Handling

Validation Commands

# Final pre-go-live verification
echo "=== CAST AI Production Validation ==="

# Agent healthy
kubectl get pods -n castai-agent -o wide

# All components running
helm list -n castai-agent

# Policies correct
curl -s -H "X-API-Key: ${CASTAI_API_KEY}" \
  "https://api.cast.ai/v1/kubernetes/clusters/${CASTAI_CLUSTER_ID}/policies" \
  | jq '{enabled, unschedulablePods: .unschedulablePods.enabled, downscaler: .nodeDownscaler.enabled, spot: .spotInstances.enabled}'

# Savings estimate
curl -s -H "X-API-Key: ${CASTAI_API_KEY}" \
  "https://api.cast.ai/v1/kubernetes/clusters/${CASTAI_CLUSTER_ID}/savings" \
  | jq '{monthly: .monthlySavings, percent: .savingsPercentage}'

Output

Create a production-readiness record that ties each checklist item to its evidence, accountable owner, approval, validation timestamp, and tested rollback. A healthy agent alone does not authorize go-live: policy limits, disruption controls, monitoring, and the emergency-disable path must all be confirmed against the intended production cluster.

Examples

For a production launch, capture a redacted policy response, current Helm release state, agent health, alert test, and savings-baseline review in the change record. If an approval or rollback test is missing, leave autoscaling disabled and resolve that gap before enabling it for production workloads.

Resources

• CAST AI Autoscaler Checklist
• CAST AI Status

Next Steps

For version upgrades, see castai-upgrade-migration.