Upgrade Cluster

Upgrade Kube-Hetzner Cluster

Safely upgrade an existing kube-hetzner-managed cluster with Terraform/OpenTofu and Kubernetes runtime proof.

Use When

• A user asks to upgrade a live kube-hetzner cluster.
• A module version bump must be applied to an existing cluster.
• k3s or RKE2 should move to a newer channel or explicit version.
• Provider/module state must be reconciled without recreating live nodes.
• The user asks whether the cluster is HA before or during an upgrade.
• The user wants to replace nodes because of suspected compromise, malware, bad host state, server type deprecation, or architecture/capacity migration.

Hard Rules

• Do not print or commit secrets. Never commit kubeconfigs, *.tfstate, *.tfvars, local env files, plan files, or .terraform/.
• Use the IaC runner already used by the target root. Examples below use terraform; substitute tofu only if the target root already uses OpenTofu.
• Never let a module refactor recreate live servers, networks, load balancers, volumes, or primary IPs by accident. If a plan shows replacement/destruction, stop and root-cause it.
• Upgrade module convergence and Kubernetes runtime versions as separate phases.
• Upgrade k3s/RKE2 one minor at a time unless the operator has explicit upstream proof that skipping minors is safe for that exact version span.
• If host compromise or malware is in scope, prefer immutable node replacement over in-place package updates or reboots. Updates patch software; fresh nodes remove old-host persistence.
• Respect kube-hetzner nodepool lifecycle rules: only add/remove nodepools at the end of each list, set old pools to count = 0 before retiring them, do not rename a non-zero pool, and keep the first control-plane nodepool at count >= 1 unless the module/root has deliberately modeled a safe replacement topology.
• Do not use k3s server --cluster-reset during healthy quorum replacement. It is a lost-quorum disaster recovery tool, not a normal control-plane migration primitive.
• Treat operator firewall access as temporary. If access is opened for maintenance, close it and prove closure before reporting done.
• Keep final proof concrete: Terraform convergence, node versions, upgrade plans complete, workloads healthy, API ready, ingress/LB healthy, and app readiness healthy.

Inputs To Establish

• Terraform root path and backend type.
• Current module source/version and target module version/commit.
• Current kubernetes_distribution, node versions, and target channel/version.
• Kubeconfig/API access path.
• Cluster topology: control-plane count, etcd membership, agent pools, autoscaler pools, ingress/load balancer.
• HA risks: singleton StatefulSets, attached volumes, PodDisruptionBudgets, critical workloads without replicas.
• Stateful data safety: CSI/Longhorn/PV attachment, backup/snapshot path, app-specific flush commands, and post-move data integrity checks.
• Threat model: ordinary upgrade/patching vs distrust of existing hosts. This changes the correct workflow.
• Health checks: Kubernetes API readiness, ingress/load-balancer health, and at least one application-level readiness URL or command.
• Maintenance constraints and rollback expectations.
• Temporary access model: SSH/Kubernetes API firewall source rules, VPN/bastion access, or control-plane LB exposure.

Preflight

cd <terraform-root>
git status --short --branch
git pull origin <default-branch>

# Confirm local secret/state files are ignored or outside git.
git ls-files | rg '(^|/)(.*kubeconfig.*|.*\.tfstate(\.backup)?|\.terraform/|.*\.tfvars|.*\.auto\.tfvars|\.terraform\.local\.env)$' || true

# Backup state without printing it. Prefer a directory outside the repo.
RUN_DIR="${KUBE_HETZNER_UPGRADE_RUN_DIR:-${TMPDIR:-/tmp}/kube-hetzner-upgrade-$(date -u +%Y%m%dT%H%M%SZ)}"
mkdir -p "$RUN_DIR"
terraform state pull > "$RUN_DIR/terraform.tfstate.before.json"

kubectl --kubeconfig <kubeconfig> get nodes -o wide
kubectl --kubeconfig <kubeconfig> get deploy,sts,pdb -A
kubectl --kubeconfig <kubeconfig> get pods -A -o wide
kubectl --kubeconfig <kubeconfig> get --raw='/readyz?verbose'

SSH Access Sanity

A live node must be reachable operationally by SSH during replacement. Check firewall reachability separately from SSH authentication:

nc -vz -w 5 <node-ip> 22
ssh root@<node-ip> -i <path-to-private-key-from-kube.tf> -o StrictHostKeyChecking=no "hostname; id; systemctl is-active sshd || systemctl is-active ssh || true"

If nc reaches port 22 and SSH returns Permission denied (publickey,keyboard-interactive), the firewall path is open and the problem is identity selection or authorized keys. Do not loosen firewall rules to fix an auth failure.

Use the exact private key configured in the target root's kube.tf/vars. This follows the upstream kube-hetzner README pattern: ssh root@<control-plane-ip> -i /path/to/private_key -o StrictHostKeyChecking=no.

If local SSH config pins another identity, plain ssh root@<node-ip> can fail even though the node and firewall are healthy. Use the key from the target root's ssh_private_key, for example:

ssh root@<node-ip> -i /path/to/private_key -o StrictHostKeyChecking=no "hostname"

If there are singleton stateful workloads or strict PDBs, decide upgrade behavior before any apply:

• system_upgrade_use_drain = true: safer for replicated stateless workloads, but can stall or move singleton stateful workloads.
• system_upgrade_use_drain = false: cordons instead of draining; useful when a drain would cause worse downtime or volume attach churn. Existing pods remain until the node service restarts.
• system_upgrade_enable_eviction = false: only relevant when draining; can unstick upgrades blocked by PDBs, but may delete pods directly.

If operator access is normally closed, open it through the target root's local, ignored variables only. Keep source CIDRs narrow, usually /32 or /128, and plan/apply the open/close firewall deltas separately from node changes.

Phase 1: Module Convergence

Update only module/provider settings first. Do not change the K3s channel in the same plan unless unavoidable.

cd <terraform-root>
terraform init -upgrade -input=false
terraform fmt -check -diff
terraform validate
terraform plan -input=false -parallelism=1 -out=module-upgrade.tfplan

Review the plan before applying.

Proceed only if the plan is explainable and does not unexpectedly destroy or replace live infrastructure.

For production in-place upgrades, MIGRATION.md is the safety contract. Save the plan JSON and run the protected hcloud no-destroy gate before applying:

terraform show -json module-upgrade.tfplan \
  | jq -r '
      .resource_changes[]?
      | select(.type as $type | [
          "hcloud_server",
          "hcloud_network",
          "hcloud_network_subnet",
          "hcloud_load_balancer",
          "hcloud_volume",
          "hcloud_primary_ip",
          "hcloud_placement_group",
          "hcloud_firewall"
        ] | index($type))
      | select(.change.actions | index("delete"))
      | "\(.address): \(.type) \(.change.actions | join(","))"
    '

Any output is a stop condition. Root-cause it before apply.

terraform apply -input=false -parallelism=1 module-upgrade.tfplan
terraform plan -input=false -parallelism=1 -detailed-exitcode

If the module changed resource ownership or addresses:

• Back up state again.
• Prefer terraform import, terraform state mv, or module-supported moved blocks over recreation.
• Migrate one live resource at a time.
• Re-run terraform plan after every state operation.
• Do not apply broad plans while state addresses are still ambiguous.

Existing Node Interface Check

Kube-hetzner v2.20 expects the private network interface to be named eth1. Fresh nodes get this from cloud-init, but long-lived pre-v2.20 nodes may still expose the private interface as a predictable kernel name such as enp7s0. If K3s restarts and /readyz stalls with logs like unable to find interface eth1, fix the existing nodes before continuing.

Verify every existing node before treating module convergence as healthy:

ssh root@<node-ip> "hostname; ip -br addr; systemctl is-active k3s k3s-agent 2>/dev/null || true"
ssh root@<node-ip> "journalctl -u k3s -u k3s-agent -n 80 --no-pager | rg 'unable to find interface eth1|server is not ready' || true"

If /etc/cloud/rename_interface.sh exists on the node, use the module-provided script so the repair matches kube-hetzner's invariant:

ssh root@<node-ip> "if ! ip link show eth1 >/dev/null 2>&1; then /etc/cloud/rename_interface.sh; fi; ip -br addr"
ssh root@<node-ip> "systemctl restart k3s || systemctl restart k3s-agent"

After repair, require the normal proof again:

kubectl --kubeconfig <kubeconfig> get nodes -o wide
kubectl --kubeconfig <kubeconfig> get --raw='/readyz?verbose'

Phase 2: Kubernetes Runtime Upgrade

Determine current and target minors:

kubectl --kubeconfig <kubeconfig> get nodes -o wide

For each minor step:

1. Keep kubernetes_distribution stable unless the operator is explicitly rebuilding/migrating runtime. For k3s, update k3s_channel or k3s_version; for RKE2, update rke2_channel or rke2_version.
2. Keep upgrade drain/eviction settings aligned with the HA risk assessment.
3. Plan and apply serially.
4. Wait for system-upgrade plans and nodes.
5. Run runtime checks before proceeding to the next minor.

terraform fmt -check -diff
terraform validate
terraform plan -input=false -parallelism=1 -out=kubernetes-<target-minor>.tfplan
terraform apply -input=false -parallelism=1 kubernetes-<target-minor>.tfplan

kubectl --kubeconfig <kubeconfig> -n system-upgrade get plans,jobs,pods -o wide
kubectl --kubeconfig <kubeconfig> get nodes -o wide
kubectl --kubeconfig <kubeconfig> get deploy,sts -A
kubectl --kubeconfig <kubeconfig> get --raw='/readyz?verbose'

Treat these as stop conditions:

• A node is not Ready.
• Server or agent upgrade plans are not complete after a reasonable wait.
• Terraform shows unexpected drift.
• Public ingress or application readiness fails repeatedly.
• A singleton stateful workload is stuck on volume attach/detach or unavailable.

RKE2 no longer implies an 8GB control-plane floor. The v3 size-aware kubelet reservation defaults make 4GB cx23 control planes viable for RKE2, and that shape is live-proven in CI. Still size real production pools for workload and etcd headroom, not just module minimums.

Phase 3: Immutable Node Replacement

Use this phase instead of in-place patching when old hosts are distrusted or when server types/architectures must be replaced cleanly.

For the full field-proven workflow, read references/immutable-node-replacement.md. It covers nodepool plan shape, Hetzner capacity checks, load-balancer target proof, etcd snapshots/membership, attached-volume StatefulSet moves, Redis/Dragonfly-compatible verification, and final firewall closure.

Capacity And Plan Shape

Before editing nodepools, verify replacement capacity and avoid deprecated/unavailable types:

hcloud server-type describe <candidate-type>
hcloud datacenter describe <candidate-datacenter-or-location>

Prefer appending a new nodepool with a new name and setting the old pool to count = 0 after evacuation. Do not remove middle-of-list pools. Kube-hetzner allocates nodepool subnets/IPs by list position, so FILO discipline prevents accidental address churn.

Stateful Workloads And Volumes

For each singleton StatefulSet or attached-volume workload:

1. Identify PVCs, PVs, storage class, volume affinity/location, and current node.
2. Take an application-consistent backup or snapshot before moving the pod.
3. Run the app-specific flush command if one exists, for example a Redis-compatible SAVE.
4. Scale the StatefulSet down, wait for CSI/Longhorn detach, then scale it up on a fresh node.
5. Verify with application-level integrity checks, not only Kubernetes readiness.

Example skeleton:

kubectl --kubeconfig <kubeconfig> -n <ns> get sts,pvc,pv -o wide
kubectl --kubeconfig <kubeconfig> -n <ns> exec <pod> -- <app-flush-or-backup-command>
kubectl --kubeconfig <kubeconfig> -n <ns> scale sts/<name> --replicas=0
kubectl --kubeconfig <kubeconfig> get volumeattachments
kubectl --kubeconfig <kubeconfig> -n <ns> scale sts/<name> --replicas=1
kubectl --kubeconfig <kubeconfig> -n <ns> exec <pod> -- <app-integrity-check>

Worker Replacement

1. Append a fresh agent_nodepools entry with available server types, enough capacity, and the same labels/taints semantics required by workloads.
2. Plan/apply with only the new workers added.
3. Wait for new workers to be Ready.
4. Verify ingress/load-balancer targets include healthy new workers.
5. Move or reschedule stateful workloads deliberately.
6. Cordon and drain old workers with flags appropriate to the workloads.
7. Scale the old worker pool to count = 0, apply, and verify the cloud servers are gone.
8. Verify final load-balancer targets reference only fresh workers.

Do not rely on Kubernetes rescheduling alone for attached-volume singletons; explicitly manage detach/attach and verify the data plane.

Control-Plane Replacement

Maintain etcd quorum at every step.

1. Confirm an odd healthy etcd/control-plane count before starting.
2. Take an etcd snapshot from a healthy control-plane node and copy or verify it outside the node being replaced.
3. Add fresh control-plane voters before removing old voters.
4. Remove only one old control-plane server at a time.
5. Delete the matching Kubernetes Node object after server removal if it remains.
6. Verify etcd membership after each removal.

Example skeleton:

ssh root@<healthy-control-plane> "k3s etcd-snapshot save --name pre-cp-replace-$(date -u +%Y%m%dT%H%M%SZ)"
kubectl --kubeconfig <kubeconfig> get nodes -l node-role.kubernetes.io/etcd=true -o wide
kubectl --kubeconfig <kubeconfig> get --raw='/db/info' || true

If the old first control-plane server type cannot be recreated, do not keep hammering Terraform against exhausted capacity. Append a new control-plane pool with available capacity, join clean voters, then retire the old pool according to the root/module constraints. If the root cannot safely model first-pool count = 0, keep a small non-serving placeholder or use state surgery only after proving the exact resource-address impact.

Secret And Incident Boundary

Immutable replacement reduces old-host persistence risk. It does not prove secrets were not exfiltrated before replacement. If the incident is confirmed rather than precautionary, report that credential rotation, image provenance review, and workload redeploy from trusted artifacts remain required.

Live Health Checks

Use checks appropriate for the cluster. Examples:

# API readiness
kubectl --kubeconfig <kubeconfig> get --raw='/readyz?verbose'

# Workload readiness
kubectl --kubeconfig <kubeconfig> get deploy,sts,pods -A -o wide

# Ingress/LB readiness, if hcloud is configured
hcloud load-balancer describe <load-balancer-name>

# Application readiness, if available
for i in 1 2 3 4 5; do
  curl -sS -o /dev/null -w "$i %{http_code}\n" <readiness-url>
  sleep 2
done

Old system-upgrade pods may show Unknown after node restarts while their jobs are Complete. Treat that as cleanup noise only if the plans are complete, all nodes are ready, and workloads are healthy.

SELinux Denials

If workloads fail under enforcing SELinux, do not globally disable SELinux as a first response. Follow docs/selinux.md: collect AVC lines from audit logs, confirm the workload name/version and k3s/RKE2 distribution, try a udica-derived workload policy first, and propose upstream module policy only with reproducible AVC evidence. Use per-pool selinux = false only as the last resort for a specific nodepool that cannot run under policy.

Firewall Closure Proof

If SSH/API access was opened for the run, close it in its own final Terraform plan after cluster health is proven.

# Remove local ignored access tfvars/env overrides first, then:
terraform plan -input=false -parallelism=1 -out=operator-close.tfplan
terraform apply -input=false -parallelism=1 operator-close.tfplan
rm -f operator-close.tfplan
terraform plan -input=false -parallelism=1 -detailed-exitcode

Then verify:

• Cloud firewall rules no longer include temporary SSH/API source CIDRs.
• Public kubectl from the closed source fails or times out, unless the cluster intentionally exposes the API through a control-plane LB.
• Public ingress/application readiness still succeeds.

Destroy Or Abort Cleanup

If an upgrade exercise is on a disposable cluster and must be torn down, run scripts/destroy.sh from the Terraform root. It auto-retries only the known benign ingress-LB detach race and then prints a read-only orphan report. Use scripts/cleanup.sh only as the forceful fallback when state is already broken or the report identifies leftovers. It treats the token's entire HCloud project as cluster-dedicated; review the dry run and include persistent data only deliberately.

Autoscaler-created servers are outside Terraform state. If they pin network/subnet deletion, delete them only after the control plane is dead, or first scale the autoscaler pool to min_nodes = 0; deleting them earlier while Cluster Autoscaler is alive can recreate them.

Final Report

Report:

• Module version/source before and after.
• Kubernetes distribution and version/channel before and after.
• HA assessment: control plane/etcd count, agent pools, ingress replicas, application replicas, and singleton stateful risks.
• Terraform proof: fmt, validate, final plan -detailed-exitcode result.
• Kubernetes proof: node versions, upgrade plans complete, API readyz, deployments/statefulsets ready.
• Ingress/app proof: load balancer target health and application readiness.
• Immutable replacement proof, if used: old servers gone, replacement nodes Ready, etcd membership contains only intended voters, stateful data integrity checks passed, and firewall access is closed again.
• Any incident during the upgrade and the corrective setting or action.
• Remaining risks and the exact condition that would make them safe.

Git Hygiene

Before committing:

git status --short
git ls-files | rg '(^|/)(.*kubeconfig.*|.*\.tfstate(\.backup)?|\.terraform/|.*\.tfvars|.*\.auto\.tfvars|\.terraform\.local\.env)$' || true
git diff --cached --check

If a secret-bearing local file is tracked, remove it from the index with git rm --cached <path>, add an ignore rule, and keep the local file only if it is still operationally needed.