K8s Clean Cluster

Force-clean all Kurtosis resources from a Kubernetes cluster when the normal kurtosis clean -a command hangs or fails.

When to use

• kurtosis clean -a hangs for more than a few minutes
• Orphaned kurtosis namespaces remain after a failed clean
• remove-dir-pod-* pods are stuck in Pending state
• Engine start fails because old resources exist

Steps

1. Kill any running kurtosis processes

pkill -f "kurtosis gateway" 2>/dev/null
pkill -f "kurtosis clean" 2>/dev/null

2. Stop the engine gracefully (if possible)

kurtosis engine stop || true

3. Delete all kurtosis namespaces

# List them first
kubectl get ns | grep kurtosis

# Delete all kurtosis namespaces (engine, enclaves, logs)
kubectl get ns | grep kurtosis | awk '{print $1}' | xargs -r kubectl delete ns --force --grace-period=0

4. Clean up cluster-scoped resources

# Delete kurtosis cluster roles
kubectl get clusterrole | grep kurtosis | awk '{print $1}' | xargs -r kubectl delete clusterrole

# Delete kurtosis cluster role bindings
kubectl get clusterrolebinding | grep kurtosis | awk '{print $1}' | xargs -r kubectl delete clusterrolebinding

5. Clean up stuck pods

# Force-delete any remaining kurtosis pods
kubectl get pods -A | grep kurtosis | awk '{print $2 " -n " $1}' | xargs -L1 kubectl delete pod --force --grace-period=0

# Clean up evicted pods
kubectl get pods -A | grep Evicted | awk '{print $2 " -n " $1}' | xargs -L1 kubectl delete pod --force --grace-period=0

6. Verify clean state

kubectl get ns | grep kurtosis
kubectl get pods -A | grep kurtosis
kubectl get ds -A | grep kurtosis

All three commands should return empty results.

7. Restart

kurtosis engine start
kurtosis gateway &

Why clean hangs

The most common cause is the fluentbit logs collector Clean method which:

1. Evicts all DaemonSet pods by adding a non-existent node selector
2. Waits for each pod to terminate (up to 5 min per pod, sequentially)
3. Creates remove-dir-pod cleanup pods targeted at each node
4. Cleanup pods on tainted/unhealthy nodes get stuck in Pending

The fix in the codebase makes these operations best-effort with timeouts and detects unschedulable pods early, but if running an unfixed version, manual cleanup is needed.