0004. local-path rather than TopoLVM or Longhorn

Status: Accepted, 2026-08-15

Context

The cluster needs persistent volumes for Grafana, Prometheus and whatever comes later. There is one node and a dedicated second SSD.

Most of the storage options in this space assume more than one node, because their value is replication.

Decision

Rancher's local-path-provisioner, backed by the second SSD, which Talos hands over as a UserVolumeConfig mounted at /var/mnt/data.

Correction, 2026-08-16. The UserVolumeConfig described here never provisioned. /var/mnt/data was a directory on the ephemeral partition from the day the cluster was built until the fix in ADR 0012. The decision recorded here stands; the claim that it was in effect did not.

Consequences

Good: almost nothing to run. No extra CRDs, no operator, no system extension, no hugepages. It is the smallest thing that provides a working StorageClass.

Bad, and this is the real cost: local-path does not implement VolumeSnapshot. VolSync is therefore limited to its Direct copy method, which reads a live volume while the application is still writing to it. For Grafana that is fine. For a database it is not, and the first database that arrives forces this decision to be revisited.

Also bad: moving off local-path later means migrating data, not just changing a StorageClass name.

Alternatives

TopoLVM. Thin provisioning, real CSI snapshots, online resize. The upgrade path when snapshots start mattering. Rejected for now because it needs an LVM2 system extension and a workaround for Talos's read-only /etc/lvm (lvm.hostWritePath=/var/lib/lvm), and there is an open compatibility issue against Talos 1.11. Complexity ahead of need.

Longhorn. Recommends 4 vCPU and 8GB per storage node and its V2 engine reserves a core plus 2GiB of hugepages. On one node it delivers a snapshot API and a web UI in exchange for that, and no redundancy at all, because there is nowhere to put a second replica.

Rook-Ceph. Documents a three node minimum. Talos's own storage guide notes Ceph "can be rather slow for small clusters". Not a real option here.