Harvester Upgrade

harvester-upgrade

Plan, sequence, and de-risk a community-edition Harvester HCI upgrade across a real fleet — typically off an EOL line (1.5.x) up to the latest settled stable. The hard part is almost never the ISO/UI upgrade click. It is the coordination: Harvester is one bundle whose embedded RKE2, KubeVirt, Longhorn, SLE Micro, and Rancher pairing all move together, the upgrade controller picks its own node order, and the workloads that matter most — VM-hosted Kubernetes control planes — must stay quorate while each host reboots underneath them.

Community editions only. SUSE sells "SUSE Virtualization" (Prime) as a paid support subscription on the same bits — it is not a different artifact set, and patch releases are community-downloadable (House Rule #1). Flag Prime-only support claims; never build a plan on them.

Companion to k8s-components-checker and rancher-upgrade — who owns what

This skill is the Harvester-side upgrade-methodology companion in the same trio (rancher-upgrade SKILL.md explicitly defers Harvester host→guest coordination here). Respect the boundary to avoid drift:

The load-bearing facts — read these before anything else

1. No minor skipping, ever. The only supported path is one minor at a time: 1.5.x → 1.6.x → 1.7.x → 1.8.x (each Harvester minor bumps embedded RKE2 exactly one k8s minor; skipping a k8s minor is unsupported upstream). Intermediate patches within a jump may be skipped (1.5.2 → 1.6.1 directly) — land on each minor's latest patch. Never hand-edit embedded RKE2 — it is locked to the Harvester version.
2. External Rancher leads every hop. When Harvester is imported into an external Rancher, each hop is a three-step sequence in order: upgrade Rancher → upgrade the Harvester UI extension → upgrade Harvester. The pairing is required (1.6↔Rancher 2.12, 1.7↔2.13, 1.8↔2.14); a mismatch yields "VM tab missing" / unmanageable cluster. The external Rancher's own multi-minor walk is its own gated project — defer to rancher-upgrade. references/external-rancher-coupling.md.
3. The node-upgrade order is not operator-choosable. Order is delegated to RKE2/Rancher and forced strictly serial (concurrency=1, one-host-at-a-time interlock). The only "pause between nodes" knob (nodeUpgradeOption: manual) exists only from v1.7.0+ — unavailable for the 1.5→1.6 and 1.6→1.7 hops. So protect VM-hosted control planes structurally, not with a button. references/controlled-flow-and-node-order.md.
4. Spreading the control-plane VMs is the real safety mechanism. Anti-affinity so the etcd VMs land on distinct hosts → a one-host-at-a-time upgrade only ever touches one member, and quorum holds whether that member migrates or is cleanly stopped. Do not assume a busy etcd VM will live-migrate — write-heavy VMs often never converge in pre-copy (Harvester ships auto-converge off and a self-throttling bandwidthPerMigration: 0), so for production etcd plan to take one member down cleanly per host and let quorum cover it, rather than relying on live migration. references/guest-rke2-survivability.md.
5. There is no downgrade. A bricked host = rebuild + restore. Back up VMs to NFS/S3 (not in-cluster snapshots) and take a guest-cluster etcd snapshot before each hop. Volume health (single-replica / degraded Longhorn) is the #1 thing that blocks a node drain — heal it first. references/landmines-and-rollback.md.

Workflow

1. Establish the change set (detect, don't assume)

Read the running versions on the (air-gapped) cluster, then map onto the ladder. Never assume the starting patch. Commands + version→minor map: references/version-ladder-and-detection.md.

2. Pick the target — latest settled stable, not bleeding-edge

"Latest stable" usually means the newest minor that has a patch released and a settled Rancher pair — not a fresh .0. Avoid a first-of-minor release for a controlled upgrade off EOL hardware; landing one minor lower often also avoids the newest Rancher minor's churn. Ground the actual latest patch of each minor via gh (House Rule #2) and apply look-ahead (House Rule #5). references/version-ladder-and-detection.md.

Do not use releases/latest to find the newest version here. Harvester patches several minors in parallel, so GitHub's date-ranked "latest" regularly points at a lower minor. Verified 2026-07-21: releases/latest resolved to v1.7.3 (2026-08-07) while v1.8.2 (2026-08-06) was the highest GA. Enumerate the tag list, drop prereleases, and derive the top patch per minor line:

gh release list -R harvester/harvester --limit 40 \
  --json tagName,publishedAt,isPrerelease \
  --jq '.[] | select(.isPrerelease==false) | "\(.tagName)\t\(.publishedAt[0:10])"'

3. For each hop: external Rancher → UI-extension → pre-flight → Harvester → verify guests

Each hop is self-contained: upgrade the external Rancher + UI extension to the pair, run the enforced pre-flight gates, then the Harvester upgrade, then verify the guest control planes before the next hop. Per-hop breaking changes + ordered runbook: references/per-hop-runbook.md. Pre-flight gate list: § below.

4. Control the flow and keep guests alive

Set up the structural protection (anti-affinity spread, N+1, MAC-pinned IPs, migratable VMs) BEFORE starting; use the v1.7.0+ pause-map only where available. references/controlled-flow-and-node-order.md + references/guest-rke2-survivability.md.

Pre-flight gates (run before EVERY hop)

Harvester blocks an upgrade that fails these. Run the official pre-check (bash pre-check/v1.x/check.sh -v from harvester/upgrade-helpers) on a control-plane node and clear every failure. The high-value gates:

• Free system space ≥ 30 GiB on /usr/local and projected post-image-load usage < 85% (kubelet imageGC). Fix with crictl rmi --prune.
• All Longhorn volumes healthy — no degraded/faulted, no single-replica volumes (they block drain), no stale attachments.
• All nodes Ready (none cordoned); CAPI local Provisioned, machine-count == node-count, all Running.
• Certs > 7 days from expiry; NTP in sync; pods Ready (rancher-logging healthy, Fleet bundles ok).
• Back up first (VM Backup to NFS/S3 + guest etcd snapshot); disable recurring Longhorn jobs.

Full gate table with thresholds + sources: references/landmines-and-rollback.md § Pre-flight.

House rules (encode into every plan)

1. Community editions only. Patch releases (1.6.1, 1.7.1, …) ARE community — verify against the real artifact (public ISO + the official 2-month patch cadence), not a "x.y.0=community / x.y.1+=Prime" myth. "Prime" = paid support on identical bits. references/version-ladder-and-detection.md § Editions.
2. Never invent versions; ground or abstain. The ladder/pairing mechanics are this skill's methodology; specific latest-patch numbers, GA dates, and "fixed in vX" claims are volatile — state them only if cluster-reported, freshly gh-grounded (anti-confirmation: enumerate tags, derive per-minor latest, never ask "does vX exist"), or marked UNVERIFIED. Run gh with valid auth (enumeration exhausts anonymous).
3. Rancher leads, no minor skips (both axes; fact 2). Cite rancher-upgrade for the Rancher chain — don't re-derive it.
4. Structural safety over the pause button (facts 3–4): the pause knob is v1.7.0+ and field-flaky, so design in anti-affinity spread + serial interlock + MAC-pinned IPs (and N+1 where possible) before the first host.
5. Look-ahead version targeting. Recommend the target that also covers the next planned hop, not the bare immediate minimum — and don't land on a fresh .0/a minor at its support ceiling. (Same rule as k8s-components-checker House Rule #9 / rancher-upgrade House Rule #4.)
6. No downgrade — back up before every hop (fact 5): VM Backup to NFS/S3 (filesystem-consistent via qemu-guest-agent) + a guest etcd snapshot; in-cluster VM snapshots are NOT DR.
7. Cite the real source per claim. Pairing/Node-Driver numbers → compat/harvester.md; Rancher-side → rancher-upgrade; everything else → the reference file that carries the doc/issue evidence.

References

• references/version-ladder-and-detection.md — the no-skip ladder + component table, the community-vs-Prime edition reality (corrected), choosing the target, and air-gapped version-detection commands (Harvester / RKE2 / KubeVirt / Longhorn / UI-extension).
• references/external-rancher-coupling.md — the Rancher→UI-extension→Harvester order, the required pairing table, air-gapped UI-extension install (ui-plugin-catalog image map), mismatch failure modes, and how the external Rancher's own chain gates the campaign (pointer to rancher-upgrade).
• references/controlled-flow-and-node-order.md — node order isn't operator-choosable; what the upgrade gates on between hosts (Longhorn yes, guest-etcd no); the 3-host physics; the per-version control table (restoreVM truth, pause-map is 1.7.0+); the guaranteed no-outage pause-map procedure; the native PDB + VMI-readinessProbe gate for 1.5/1.6 (with the pinning-bypass + must-test caveats); why the 1.8 upgrade-manager is off-limits.
• references/guest-rke2-survivability.md — making self-managed RKE2 guests Harvester-aware and survivable: anti-affinity spread (the #1 lever), N+1 capacity, live-migration prerequisites, qemu-guest-agent, the Harvester cloud-provider (CCM) + CSI standalone install, etcd-during-migration reality, MAC-pinned IPs.
• references/per-hop-runbook.md — per-hop (1.5→1.6, 1.6→1.7, 1.7→1.8) breaking changes, manual pre-steps, and the ordered runbook (wicked→NetworkManager, Intel NIC rename, DHCP client-ID, secondary-VLAN strip, etc.).
• references/landmines-and-rollback.md — the enforced pre-flight gate table, the landmine quick-reference (with issue numbers), Longhorn-as-#1-blocker, and the no-downgrade backup/rollback reality.
• references/sources.md — primary sources (Harvester docs, release notes, controller source, upgrade-helpers, CSI/CCM) with a freshness ledger.