K3S-Deploy: harden HA bring-up and expand README caveats

Add a post-join control-plane readiness gate so a half-succeeded server join on a degraded etcd can no longer pass as success. Retry the MetalLB IPAddressPool/L2Advertisement applies to ride out the validating-webhook startup race. Narrow the k3sup cleanup glob, quiet the final k3sup ready tip, and clarify two comments. README: add networking caveats (VIP/LB outside the DHCP pool, same-subnet ARP/L2, interface naming), certName key placement, a post-run verification section, machine-count/admin-box notes, and fix the RAW_BASE description.
2026-07-26 18:38:14 +00:00 · 2026-07-15 22:26:45 -06:00
parent 49313239f4
commit c436d29182
2 changed files with 89 additions and 13 deletions
--- a/Kubernetes/K3S-Deploy/README.md
+++ b/Kubernetes/K3S-Deploy/README.md
@@ -38,26 +38,71 @@ duplicate LoadBalancer controllers.

 - An **SSH key pair** you can use to reach the nodes (the script distributes
  the public key and never clobbers your `~/.ssh/config`).
- At least **3 control-plane nodes** for real HA (etcd needs a quorum).
- Run the script from an **Ubuntu/Debian admin box** (it installs `k3sup`
-  and `kubectl` locally if missing).
+- **5 nodes by default** (3 servers + 2 agents). The minimum for HA is 3
+  servers (etcd needs a quorum); workers are optional. Adjust the node list in
+  the config block if you have fewer machines.
+- Run the script from a **separate Ubuntu/Debian admin machine** (your
+  laptop/workstation) that can SSH to every node — not on one of the nodes
+  themselves. It installs `k3sup` and `kubectl` locally if missing.
+
+## Important caveats
+
+- **The VIP and the MetalLB range must be outside your DHCP pool**, and must
+  not overlap each other or any node IP. If your router hands out `vip` or an
+  address in `lbrange` via DHCP, you'll get intermittent, hard-to-debug
+  failures — reserve them on your router/DHCP server first.
+- **Same subnet / L2 only.** kube-vip (ARP mode) and MetalLB (L2) both
+  advertise via ARP, so the VIP and LoadBalancer IPs must be on the **same
+  subnet/VLAN** as the nodes and the clients reaching them — they are not
+  routed across subnets.
+- **`interface` must match your nodes' real NIC.** It's `eth0` on many
+  systems but often `ens18`/`enp0s3` on Proxmox and cloud images. Check with
+  `ip -o -4 route show default`.
+- **Host keys are trusted on first use** (`accept-new`). If you re-image a
+  node, clear its stale entry first: `ssh-keygen -R <node-ip>`.
+- This is a **homelab tutorial, not production-hardened** as-is (aggressive
+  leader election, no etcd backups, single L2 domain).

 ## Usage

-1. Snapshot your VMs.
-2. Copy your SSH key into your home directory (or into `~/.ssh`).
+1. Snapshot your VMs (so you can roll back — the script modifies every node).
+2. Place your SSH **private** key at `~/<certName>` or `~/.ssh/<certName>`,
+   with the matching `.pub` beside it. `certName` is the key's *filename only*
+   — no path, no `.pub` suffix. Modern OpenSSH defaults to `id_ed25519`, so set
+   `certName` to match the key you actually have.
 3. Edit the **"YOU SHOULD ONLY NEED TO EDIT THIS SECTION"** block in
   `k3s.sh` — node IPs, `user`, `interface`, `vip`, `lbrange`, `certName`.
+   The example IPs (`192.168.3.x`, VIP `.50`, range `.60-.80`) and
+   `interface=eth0` are placeholders — change them for your network.
 4. `chmod +x k3s.sh && ./k3s.sh`
 5. Review the pre-flight summary and confirm. Grab a coffee.

+### After it finishes
+
+The kubeconfig is **merged** into `~/.kube/config` as context `k3s-ha` (set by
+`context` in the config block), pointed at the VIP. Select it and check the
+cluster:
+
+```bash
+kubectl config use-context k3s-ha
+kubectl get nodes -o wide
+curl http://<load-balancer-ip>   # the script prints the assigned IP
+```
+
+Worker nodes are labeled `worker=true` and `longhorn=true` (the latter for the
+companion Longhorn storage tutorial — harmless if you don't use it).
+
+If a node fails to join, reset just that node and re-run the script (it's
+idempotent): `k3s-uninstall.sh` on a server, `k3s-agent-uninstall.sh` on a
+worker.
+
 ### Options (environment variables)

 | Variable         | Purpose                                               |
 | ---------------- | ----------------------------------------------------- |
 | `ASSUME_YES=1`   | Skip the pre-flight `[y/N]` prompt (unattended runs). |
 | `NO_COLOR=1`     | Disable colored output.                               |
-| `RAW_BASE=<url>` | Base URL for the sibling manifests (default: `main`). |
+| `RAW_BASE=<url>` | Where to fetch the sibling manifests (kube-vip, ipAddressPool, l2Advertisement); default: upstream `main`. Override for a fork/branch/local copy. |

 ### Tracking the latest k3s instead of a pinned version

@@ -75,7 +120,6 @@ docker run --rm ghcr.io/kube-vip/kube-vip:<version> manifest daemonset \
  --controlplane --arp --leaderElection --taint --inCluster > kube-vip
 sed -i 's/value: eth0/value: REPLACE_INTERFACE/' kube-vip
 sed -i 's/value: 10.0.0.254/value: REPLACE_VIP/' kube-vip
-
 ```

 `--inCluster` is required on k3s: it makes kube-vip use the `kube-vip`
--- a/Kubernetes/K3S-Deploy/k3s.sh
+++ b/Kubernetes/K3S-Deploy/k3s.sh
@@ -120,6 +120,21 @@ server_extra_args() {  # $1 = node ip
 # never node-token (whose stdout is captured for the token).
 k3sup_quiet() { command k3sup "$@" 2>&1 | sed '/Create clusters on Mac/d'; }

+# Apply a manifest, retrying on transient failure. MetalLB installs a
+# validating webhook served by its controller; there is a brief window where
+# the controller Deployment reports Available but the webhook endpoints/TLS
+# are not serving yet, so the first apply can fail with "no endpoints
+# available for service metallb-webhook-service". Retry to ride that out.
+kubectl_apply_retry() {  # $1 = manifest path or URL
+  local i
+  for i in $(seq 1 6); do
+    kubectl apply -f "$1" && return 0
+    warn "kubectl apply $1 failed (attempt $i/6) — retrying in 5s"
+    sleep 5
+  done
+  die "kubectl apply $1 failed after 6 attempts"
+}
+
 # ── Pre-flight: show config, confirm before touching any node ──────────
 step "Pre-flight · review configuration"
 printf '  %-16s %s\n' \
@@ -153,7 +168,7 @@ if ! command -v k3sup &>/dev/null; then
  done
  [[ -n "$k3sup_bin" ]] || die "k3sup installer produced no binary"
  sudo install "$k3sup_bin" /usr/local/bin/k3sup
-  rm -f k3sup k3sup-*
+  rm -f "$k3sup_bin"
 else
  ok "k3sup present"
 fi
@@ -269,8 +284,10 @@ REMOTE
 done
 ok "kube-vip deployed to ${#masters_all[@]} control-plane nodes"
 info "Pointing local kubeconfig at the VIP ($vip)"
-# Rewrite only the server URL (anchored) to avoid the regex dots in the IP
-# matching anything else in the kubeconfig (e.g. base64 cert data).
+# Rewrite the server URL from master1 to the VIP. The dots in the IP are
+# unescaped regex metacharacters, but a false match would need the literal
+# "https://<master1-ish>:6443" string elsewhere in the kubeconfig, which does
+# not occur (cert data is base64 — no scheme or port), so this is safe here.
 sed -i "s#https://$master1:6443#https://$vip:6443#" "$HOME/.kube/config"

 # Confirm the VIP is actually answering before any VIP-routed kubectl below.
@@ -296,6 +313,9 @@ ok "Join token fetched"

 # ── Step 7/9 · Join control-plane and worker nodes ─────────────────────
 step "Step 7/9 · Join control-plane and worker nodes"
+# The prefetched --node-token is what keeps these joins from ever SSHing to
+# the server: without it, k3sup would try --server-user (default: root) against
+# --server-ip (the VIP, not an SSH target) and fail confusingly. Keep the token.
 for node in "${masters[@]}"; do
  info "Joining control-plane node $node"
  k3sup_quiet join \
@@ -324,6 +344,18 @@ for node in "${workers[@]}"; do
  ok "Worker node $node joined"
 done

+# Assert the full HA control plane is up before continuing. The VIP gate in
+# Step 5 only proved master1 (a single-member etcd); a control-plane join that
+# half-succeeds would otherwise sail through as success on a degraded cluster.
+info "Waiting for all ${#masters_all[@]} control-plane nodes to be Ready"
+kubectl wait --for=condition=Ready node \
+  -l node-role.kubernetes.io/control-plane --timeout=180s || \
+  die "Control-plane nodes did not all become Ready — check etcd/kube-vip on the masters: kubectl get nodes; kubectl -n kube-system logs -l app.kubernetes.io/name=kube-vip-ds"
+cp_ready="$(kubectl get nodes -l node-role.kubernetes.io/control-plane --no-headers 2>/dev/null | awk '$2=="Ready"{c++} END{print c+0}')"
+[[ "$cp_ready" -eq "${#masters_all[@]}" ]] || \
+  die "Expected ${#masters_all[@]} control-plane nodes Ready, found $cp_ready — etcd quorum may be degraded"
+ok "All ${#masters_all[@]} control-plane nodes Ready"
+
 # ── Step 8/9 · Install MetalLB (service LoadBalancer) ──────────────────
 step "Step 8/9 · Install MetalLB (service LoadBalancer)"
 kubectl apply -f "https://raw.githubusercontent.com/metallb/metallb/$METALLB_VERSION/config/manifests/metallb-native.yaml"
@@ -335,15 +367,15 @@ kubectl -n metallb-system rollout status deploy/controller --timeout=120s
 info "Configuring address pool ($lbrange)"
 curl -sfL "$manifest_base/ipAddressPool" -o "$HOME/ipAddressPool.src"
 sed "s|REPLACE_LBRANGE|$lbrange|g" "$HOME/ipAddressPool.src" > "$HOME/ipAddressPool.yaml"
-kubectl apply -f "$HOME/ipAddressPool.yaml"
+kubectl_apply_retry "$HOME/ipAddressPool.yaml"
 curl -sfL "$manifest_base/l2Advertisement.yaml" -o "$HOME/l2Advertisement.yaml"
-kubectl apply -f "$HOME/l2Advertisement.yaml"
+kubectl_apply_retry "$HOME/l2Advertisement.yaml"
 ok "MetalLB configured"

 # ── Step 9/9 · Verify cluster and LoadBalancer ─────────────────────────
 step "Step 9/9 · Verify cluster and LoadBalancer"
 info "Waiting for the cluster to be ready"
-k3sup ready --context "$context" --kubeconfig "$HOME/.kube/config"
+k3sup_quiet ready --context "$context" --kubeconfig "$HOME/.kube/config"
 info "Deploying nginx sample"
 kubectl apply -n default -f https://raw.githubusercontent.com/inlets/inlets-operator/master/contrib/nginx-sample-deployment.yaml
 # `|| true` — expose is imperative and errors with AlreadyExists on a re-run,