Adopt Longhorn for heterogeneous node storage

[jorisjonkers-dev-agents]

Context. Nodes differ in storage size/speed; not all should host replicas. Plan + roll out Longhorn with tiers, using the existing per-node inventory (incl. GPU-memory scheduling) to place replicas.

Steps

• Define storage tiers / node labels from the inventory (fast-ssd vs bulk vs no-storage nodes).
• Install Longhorn via a platform-blueprints pack; set replica count + node selectors per tier.
• Migrate one non-critical PVC as a pilot; verify replica placement honours the tiers.

Acceptance

• Longhorn healthy; replicas only on storage-capable nodes; pilot volume survives a node drain.

[ExtraToast]

Still open, and genuinely unstarted — status as of 2026-08-27

Checked rather than assumed. Longhorn was never installed:

$ kubectl get sc
NAME                   PROVISIONER             ...
local-path (default)   rancher.io/local-path

No longhorn namespace, one storage class, local-path. So none of the three steps here have happened.

The groundwork does exist

nix-config/inventory/nodes/*.yml already carries per-node Longhorn declarations — storage.longhorn.eligible, nodeTags like site-enschede / longhorn-storage, and per-disk entries tagged fast/nvme and bulk/hdd. Step 1 of this issue is effectively done in the inventory; it is the install and migration that never followed.

Two things that have changed since this was written

1. It has a concrete cost now. The absence of Longhorn is why PVC-level snapshots are impossible: no VolumeSnapshot CRDs, and local-path has no CSI snapshot support. That forced the backup design into logical dumps on a single node-pinned PVC, and a job that pretended snapshots worked had to be deleted. Recorded in JorisJonkers-dev/workspace#32.
2. Step 2 names a retiring repo. "Install Longhorn via a platform-blueprints pack" — ExtraToast/platform-blueprints is scheduled for a confirm retirement decision on 2026-09-26, and may be superseded by nixos-modules + flux-modules. Worth resolving that before starting, or the first step is against a repo that is going away.

Leaving open: this is real work someone chose not to do, not a stale entry.