viktor/infra
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

docs/plans: HA control plane — design + plan + deferral

Browse source 
Investigated, designed, and planned the 3-master HA control plane
migration triggered by 2026-05-21's autonomous k8s upgrade cascade.

Locked 14 design decisions across two passes:
- 10 initial decisions (LB strategy, IPs, sizing, etcd, kured gate, etc)
- 4 challenger-pass amendments (cloud-init template bump, rbac stack
  multi-master refactor, HTTPS /readyz health check, expanded blast
  radius to include /home/wizard/code/infra/config root kubeconfig,
  config.tfvars, k8s-portal user kubeconfigs, etcd-backup nodeSelector,
  k8s-version-upgrade chain extension as Phase 7)

Plan covers 11 phases end-to-end including panic-mode rollback.

DEFERRED before execution. PVE host is 98% RAM-committed
(262 GB allocated / 267 GB physical, 1.5 GB swap active); the
planned 3 x 32 GB masters would push allocation to 326 GB and OOM
the host. k8s-master currently uses only 4.6 GB of its 32 GB
allocation (5-6x oversized).

Revisit triggers documented in design doc:
1. Second PVE host added → hardware HA becomes possible.
2. Right-sizing pass OR planning masters at 16 GB each.
3. Cumulative manual upgrade nursing > ~10h.

Standalone candidate worth lifting independently: Phase 1.5's
rbac stack refactor (apiserver-oidc + audit-policy + etcd-tuning
to loop over k8s_master_hosts list) — future-proofs the cluster
without committing to the HA migration.

Refs: code-n0ow (open, deferred via bd note).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
Viktor Barzin 2026-05-23 08:32:15 +00:00
parent 70a334e431
commit 7f63d35d0a
2 changed files with 387 additions and 19 deletions
Download patch file Download diff file Expand all files Collapse all files

81
docs/plans/2026-05-21-ha-control-plane-design.md
View file

@ -1,9 +1,32 @@
# HA Control Plane (3 masters) — Design
**Date**: 2026-05-21
**Status**: Drafted, NOT scheduled
**Beads**: code-n0ow
**Trigger**: today's k8s 1.34.7→1.34.8 autonomous-upgrade session repeatedly hit a storm cascade rooted in single-master apiserver outages
**Date**: 2026-05-21 (decisions locked 2026-05-22; **deferred 2026-05-23**)
**Status**: **DEFERRED** — design + plan complete, NOT scheduled. Awaiting either PVE host capacity expansion OR a separate right-sizing pass on the existing master before this becomes affordable. Paired plan: `2026-05-21-ha-control-plane-plan.md`.
**Beads**: code-n0ow (open, deferred — see `bd show code-n0ow`)
**Trigger**: 2026-05-21 k8s 1.34.7→1.34.8 autonomous-upgrade session repeatedly hit a storm cascade rooted in single-master apiserver outages
## Why deferred (2026-05-23)
Measured during the locking pass:
- **k8s-master uses 4.6 GB of 32 GB allocated** (kube-apiserver 2.6 GB + etcd 660 MB + cm 360 MB + ~1 GB everything else). The 32 GB sizing is ~5-6× oversized vs working set.
- **PVE host is already 98% RAM-committed** — 262 GB allocated to VMs against 267 GB physical, with 1.5 GB of active swap. The planned 3 × 32 GB control plane (+64 GB net) would push allocation to 326 GB → OOM on the host.
- **Software-only HA on a single PVE host has bounded value** — a hypervisor crash still loses all 3 masters. The big resilience wins (kubeadm upgrades, cert rotation, planned reboots) are real but the disaster-recovery angle is limited until a second PVE host exists.
### Revisit triggers — any of:
1. **Second PVE host added** to the lab. Hardware HA becomes possible; HA control plane becomes the natural follow-up. Spread the 3 masters across 2 hosts (2+1).
2. **Cluster-wide right-sizing pass** that frees enough headroom for the original 3 × 32 GB plan, OR pre-agreed amendment to provision 16 GB masters (right-sized to actual usage; 3-4× current working-set headroom).
3. **Storm cascade burns enough hours** that the operational cost outweighs the memory cost — track minutes spent manually nursing kubeadm upgrades; if cumulative > ~10h over a few months, revisit.
### What's still good
The design + plan in this directory remain authoritative. When we revisit:
- All 14 locked decisions stand.
- Challenger amendments (cloud-init template bump, rbac multi-master refactor, HTTPS `/readyz` health check, expanded blast radius, etcd-backup nodeSelector, chain extension as Phase 7) are baked in.
- Only the sizing decision needs revisiting — likely 16 GB per master instead of 32 GB.
- Adding `k8s_master_hosts` list-based refactor to the rbac stack (Phase 1.5) is a **standalone win** that could be done independently of HA — it would future-proof the cluster against the day HA lands. Consider lifting that as its own task.
## Problem statement
@ -50,18 +73,24 @@ The k8s upgrade chain doesn't need to be aware of *any* of this — the
underlying availability of apiserver makes the chain's gates
naturally pass on each iteration.
## Decisions (proposed — to be confirmed)
## Decisions (locked 2026-05-22)
| # | Decision | Notes |
|---|----------|-------|
| 1 | **3 masters** (not 5) | Quorum tolerates 1 failure, sufficient for home-lab. 5 would tolerate 2 but doubles etcd write amplification. |
| 2 | **Sizing**: match current `k8s-master` (8 vCPU, 32GB RAM, ~64 GB disk) for all 3 | Symmetric. New VMs `k8s-master-2`, `k8s-master-3` on Proxmox. |
| 3 | **Apiserver LB**: **pfSense HAProxy** (existing pattern, see mailserver-pfsense-haproxy.md) over keepalived+haproxy-on-each-master | Pros: no per-node moving parts, mirrors the mailserver layout already in production. Cons: pfSense becomes more SPoF — but it's already SPoF for everything else (DNS, gateway, ingress). |
| 4 | **VIP**: pick an unused IP on the cluster VLAN, e.g. `10.0.20.99`, point all kubeconfigs + kubelet `--server` at it | Internal-only VIP; external API access stays via Cloudflared. |
| 5 | **etcd**: kubeadm-managed (existing); just `kubeadm join --control-plane` brings new members into the etcd cluster automatically | Avoids running etcd separately. |
| 6 | **kured-sentinel-gate**: extend "quorum-safe" check to verify ≥2 control-plane nodes Ready before allowing a reboot | Otherwise kured could reboot 2 masters at once and break quorum. |
| 7 | **etcd backup**: today's `etcd-backup` CronJob already takes a snapshot from one member; that's still sufficient (etcd snapshot is a consistent point-in-time). No new work needed. | |
| 8 | **Migration order**: add masters one at a time, run smoke (kubectl from each), then cut over kubeconfigs | Each `kubeadm join --control-plane` is reversible (just `kubeadm reset` + remove from etcd member list). |
| 2 | **Sizing**: match current `k8s-master` (8 vCPU, 32GB RAM, ~64 GB disk) for all 3 | Symmetric. New VMs `k8s-master-2` (VMID 205, 10.0.20.110), `k8s-master-3` (VMID 206, 10.0.20.111). |
| 3 | **Apiserver LB**: **pfSense HAProxy** — new TCP frontend on `10.0.20.99:6443` mirroring the mailserver pattern. Idempotent via `scripts/pfsense-haproxy-bootstrap.php`. | Pros: no per-node moving parts, mirrors existing mailserver layout. Cons: pfSense becomes more SPoF — but it's already SPoF for everything else (gateway/DNS/ingress). |
| 4 | **VIP**: `10.0.20.99` (one below current master `.100`, well clear of MetalLB pool `.200-.220`). Internal-only — external API access stays via Cloudflared. | All kubeconfigs + kubelet.conf entries flip from `10.0.20.100:6443` → `10.0.20.99:6443`. |
| 5 | **etcd**: kubeadm-managed stacked; `kubeadm join --control-plane` brings new members into the etcd cluster automatically | Avoids running etcd separately. |
| 6 | **kured-sentinel-gate**: extend the bash loop in `stacks/kured/main.tf` with a "≥2 control-plane nodes Ready" check between the existing all-nodes-Ready and calico-Ready checks | Otherwise kured could reboot 2 masters at once and break quorum. |
| 7 | **etcd backup**: `etcdctl snapshot save` from any member is a consistent point-in-time of the full quorum state — but the existing CronJob is pinned `node_name = "k8s-master"`. Phase 4.5 flips this to a control-plane label + toleration so backups don't silently skip when master-1 is drained. | Snapshot CORRECTNESS unchanged; SCHEDULING needs fixing. |
| 8 | **Migration order**: Phase 0 (retrofit existing cluster) → Phase 1 (LB up, single backend, HTTPS health check) → Phase 1.5 (rbac stack refactor) → Phase 2 (cloud-init bump + master-2 join + add to LB) → Phase 3 (master-3 join + add to LB) → Phase 4 (flip clients + workers to VIP) → Phase 4.5 (etcd-backup CronJob fix) → Phase 5 (kured-sentinel-gate quorum check) → Phase 6 (E2E validation) → Phase 7 (k8s-version-upgrade chain extension) | Each kubeadm join is reversible (`kubeadm reset` + `etcdctl member remove`). |
| 9 | **VM provisioning**: cloud-init via `create-template-vm` module, **but the template needs an apt-source bump first** (v1.32 → v1.34) and a control-plane gate on `k8s_join_command` so master VMs don't auto-join as workers. Existing master stays as the legacy manual VM (not rebuilt). | The repo has zero VMs using cloud-init for provisioning today — we're the first user. Update template first, then use it. |
| 10 | **Cert SAN + controlPlaneEndpoint retrofit**: Phase 0, before any new master joins. Patch `kubeadm-config` via `kubeadm init phase upload-config kubeadm --config <file>` (kubeadm-owned write, future-proof against `kubeadm upgrade apply`), regen `apiserver.crt` via `kubeadm init phase certs apiserver`, restart the kube-apiserver pod (~30s outage on the existing master only). | Standard kubeadm retrofit path; `kubeadm join --control-plane` requires controlPlaneEndpoint to be set. |
| 11 | **Multi-master config propagation (Phase 1.5)**: refactor `stacks/rbac/modules/rbac/{apiserver-oidc,audit-policy,etcd-tuning}.tf` to loop over a list of master hosts. Apply BEFORE master-2/3 join so they boot with OIDC, audit policy, and etcd tuning already in place. | Today these stacks SSH into a single master and sed into `kube-apiserver.yaml` — if not propagated, Authentik login flaps depending on which master the LB lands on. |
| 12 | **k8s-version-upgrade chain extension (Phase 7)**: extend `stacks/k8s-version-upgrade/scripts/upgrade-step.sh` to discover and iterate over all control-plane nodes (drain → upgrade → uncordon, gated by quorum check). | Without this, chain only upgrades master-1; masters 2/3 drift behind one version per upgrade. Original autonomous-upgrades goal unmet. |
| 13 | **LB health check**: HTTPS `GET /readyz` (with `verify none` for self-signed apiserver cert), NOT plain TCP. | Plain TCP misses apiserver-NotReady states (etcd unreachable, controller-manager flapping). |
| 14 | **VIP DNS name**: add `k8s-apiserver IN A 10.0.20.99` to `config.tfvars` BEFORE Phase 4. Delete stale `kubernetes IN A 10.0.20.100`. Consumers reference the FQDN, not the bare IP — future renumbering is then a single record change. | |
## Out of scope
@ -74,12 +103,15 @@ naturally pass on each iteration.
| Risk | Mitigation |
|---|---|
| Phase 0 cert regen on existing master triggers a brief apiserver outage (~30s) | Already a known cluster behaviour during static-pod restart. Schedule during a low-activity window. Tigera/operators will crash-loop briefly but recover — same blast radius as today's k8s upgrade. **Once HA is up, future restarts won't have this surface at all.** |
| etcd quorum split-brain during member join | kubeadm join is atomic; if it fails, the new member doesn't join the quorum. Existing etcd stays healthy. |
| LB misconfiguration → all kubectl breaks | Smoke-test from each master before flipping clients. Keep a kubeconfig pointing directly at one master as fallback. |
| Existing kubeconfigs (dev VM, agents, woodpecker) need updating | List all consumers, update in a single TF apply. |
| New masters get scheduled some workload pods unintentionally | Verify control-plane taint is applied at join time. |
| Cluster-wide cert rotation might be needed | kubeadm join handles certs automatically using the `--certificate-key` from `kubeadm init phase upload-certs`. |
| 32GB per master × 3 = 96GB RAM used for control plane alone | Proxmox host has headroom; not blocking. |
| LB misconfiguration → all kubectl breaks | Smoke-test from each master directly (bypass LB) before flipping clients. Keep a kubeconfig pointing at `10.0.20.100:6443` as fallback. |
| Existing kubeconfigs (Woodpecker pipelines, agents, dev VM, in-cluster RBAC default) need updating | Single Terraform apply touches `stacks/rbac/modules/rbac/apiserver-oidc.tf` (default), `.woodpecker/*.yml` (committed kubeconfigs). Worker `kubelet.conf` files patched in Phase 4 via ssh loop. |
| New masters get scheduled workload pods unintentionally | Verify `node-role.kubernetes.io/control-plane:NoSchedule` taint is applied at join time (default with `--control-plane`). |
| Cert rotation propagation | kubeadm join uses the `--certificate-key` from `kubeadm init phase upload-certs` to fetch existing CA materials. Single short-lived secret in `kube-system/kubeadm-certs` (**2h TTL** — Phases 2 + 3 must complete within the window, or re-upload between them). |
| 32GB per master × 3 = 96GB RAM used for control plane alone | PVE host has 272GB total, 176GB allocated to cluster pre-HA. Post-HA: 240GB allocated, 32GB headroom. Sufficient. |
| Pre-existing kubeadm-config does NOT have `controlPlaneEndpoint` set | Phase 0 patches it. Verify: `kubectl -n kube-system get cm kubeadm-config -o yaml \| grep controlPlaneEndpoint` (absent → `10.0.20.99:6443` post-Phase 0). |
| Existing master cert SANs are `[k8s-master, 10.96.0.1, 10.0.20.100]` only — missing VIP | Phase 0 regens with `--apiserver-cert-extra-sans 10.0.20.99` after patching kubeadm-config. |
## Verification
@ -91,12 +123,23 @@ kubectl get nodes -l node-role.kubernetes.io/control-plane=
# etcd quorum healthy
kubectl -n kube-system exec etcd-k8s-master -- etcdctl \
--endpoints=https://10.0.20.100:2379,https://10.0.20.X:2379,https://10.0.20.Y:2379 \
--endpoints=https://10.0.20.100:2379,https://10.0.20.110:2379,https://10.0.20.111:2379 \
--cacert=/etc/kubernetes/pki/etcd/ca.crt \
--cert=/etc/kubernetes/pki/etcd/server.crt \
--key=/etc/kubernetes/pki/etcd/server.key \
endpoint health --cluster
# Kubeconfig points at VIP
kubectl --kubeconfig ~/.kube/config config view --minify -o jsonpath='{.clusters[0].cluster.server}'
# Expect: https://10.0.20.99:6443
# Worker kubelet.conf points at VIP
for n in k8s-node{1,2,3,4}; do
ssh wizard@$n.viktorbarzin.lan "sudo grep -E '^\s+server:' /etc/kubernetes/kubelet.conf"
done
# Expect: server: https://10.0.20.99:6443 on every node
# Failover test: cordon master-1, reboot it, observe kubectl still works through LB
kubectl drain k8s-master --delete-emptydir-data --ignore-daemonsets
ssh wizard@k8s-master.viktorbarzin.lan sudo reboot
@ -110,7 +153,7 @@ kubectl -n k8s-upgrade create job --from=cronjob/k8s-version-check ha-validation
- 2× VMs at 8 vCPU + 32GB RAM each = +64GB RAM on Proxmox host
- ~+128GB disk usage (2× 64GB master disks)
- ~2-4 hours of operator time end-to-end (VM provisioning + kubeadm join + LB config + smoke)
- **~5-7 hours of operator time end-to-end** (cloud-init template bump + Phase 0 retrofit + LB + Phase 1.5 rbac refactor + 2× kubeadm join + Phase 4 cutover + Phase 4.5 etcd-backup fix + Phase 5 kured-gate + Phase 6 validation + Phase 7 chain extension). Phases 0–6 can land in one session; Phase 7 can be deferred a few days if needed.
## What's already in place from today's work