Storage

All data lives on TrueNAS SCALE (a VM on Proxmox) in one pool, Firelink: three disks in RAIDZ1, about 7.2 TiB usable. The layout was cleaned up in October 2026, before backups were set up.

Rule: one top-level dataset per kind of data

Snapshots, retention, quotas and replication are set per dataset in ZFS. So each top-level dataset holds one kind of data with one backup policy, and anything below it inherits that policy. Dark Souls names at the top level, plain names below.

DatasetContentsBackup policy
AnorLondoPersonal files: photos (Immich), documents (Nextcloud), SMB shareHighest, offsite. Irreplaceable.
DukesArchivesMedia library: movies/, series/, audiobooks/, downloads/Low. Can be downloaded again.
BlightTownKubernetes app data over NFS, one child dataset per appMedium
SensFortressThe only Proxmox storage (ISOs, LXC templates, VM disks)Low
UndeadParishData for TrueNAS apps (Omada)Medium
CatacombsBackup target only(is the backup)
ix-appsTrueNAS system datasetDo not touch
Firelink
├── AnorLondo
│   ├── immich            Immich originals, profile/ and backups/ (DB dumps)
│   └── nextcloud         Nextcloud user files (planned)
├── DukesArchives         movies/ series/ audiobooks/ downloads/  (folders, not datasets)
├── BlightTown
│   ├── home-assistant
│   ├── immich            Immich thumbs/ and encoded-video/ (can be regenerated)
│   └── uptime-kuma       Old Uptime Kuma data, unused since the move to LXC 101; delete in phase 14
├── SensFortress          template/iso, template/cache, images/
├── UndeadParish
│   └── omada             Omada data (planned; today still in Catacombs/data)
└── Catacombs
    └── backups
        ├── restic        Laptop backups (restic repositories, one per user, encrypted), see docs/operations/backups.md
        ├── k8s           Copies of Kubernetes app data (planned)
        └── truenas       TrueNAS config exports (root only)

Who uses what

ConsumerHowPath
Kubernetes (static PVs)NFS, one share per app dataset, only 192.168.30.0/24/mnt/Firelink/BlightTown/<app>, /mnt/Firelink/AnorLondo/immich, /mnt/Firelink/DukesArchives, /mnt/Firelink/Catacombs/backups/restic
Proxmox storage SensFortressSMB (CIFS), mounted by Proxmox at /mnt/pve/SensFortressSensFortress
Nothing (was Plex and the docker LXC, both removed 2026-11-19)SMB, /etc/fstab on pve at /mnt/dukesarchives. Unused: remove the fstab line and unmountDukesArchives
Personal devicesSMBAnorLondo
Omada (TrueNAS app)host pathCatacombs/data (moving to UndeadParish/omada)

Kubernetes: NFS or node-local disk

Two kinds of storage, chosen by the kind of data:

DataWhereStorageClass
Files: photos, documents, media, app config foldersNFS from TrueNASnone (static PVs)
Databases: Postgres (CloudNativePG), SQLite, PrometheusThe node’s own disklocal-path

Why databases are not on NFS:

  • SQLite needs file locking that NFS does not provide reliably. Uptime Kuma’s README says NFS is not supported, and the *arr apps are known to corrupt their databases on NFS. Postgres is slow and fragile on NFS (every fsync goes over the network), and Immich’s docs require local storage for its database.
  • Local disk does not stop when TrueNAS or Proxmox restarts.

What it costs, and how that is handled:

  • A local-path volume pins its pod to one node. If that node is down, the app is down. Postgres gets its redundancy from CloudNativePG instead (more instances on other nodes), not from the storage.
  • Local volumes are outside ZFS snapshots, so each database needs its own backup to the NAS or S3 (see the backup plan).
  • The PVC size is not enforced. Volumes share the node’s root disk:
NodeRoot diskFree (October 2026)
lenovo198 GB (LVM)80 GB
lenovo2196 GB (LVM)155 GB
pi5459 GB (NVMe)417 GB

Immich is split across two datasets, as Immich’s “custom locations” guide allows: AnorLondo/immich is mounted at /data (originals, profile pictures, database dumps) and BlightTown/immich at /data/thumbs and /data/encoded-video. The generated files can be 10–20% of the library and do not need the offsite backup. Immich runs as UID/GID 3000 (larma on TrueNAS), which owns both datasets, so the originals can also be read over SMB with the same account. NFS maps root to nobody, so running as root would not work.

Rejected: iSCSI from TrueNAS (democratic-csi) and replicated storage (Longhorn, Rook-Ceph). Both would let a stateful pod move between nodes, but add a lot of moving parts, and Ceph is too heavy for 8 GB nodes.

Snapshots

Every dataset has periodic ZFS snapshots. The schedules, and how to restore from them, are in backups.

Things to know

  • An NFS share does not reach into child datasets. A share on BlightTown shows each child dataset as an empty, root-owned folder. Every app dataset needs its own share.
  • Kubernetes PVs are static, and spec.nfs.path cannot be changed. Keep the dataset at the path the PV points to.
  • Do not put SQLite on NFS (the *arr apps, Jellyfin, Audiobookshelf, Homarr). Use local disk in the cluster instead.
  • Media stays one dataset. Radarr and Sonarr hardlink from downloads/ to movies/ and series/, and hardlinks do not work across datasets or across separate mounts. Mount the whole dataset once (as /data) in the apps that move files.
  • Proxmox’s folder layout is fixed: ISOs go in template/iso/, LXC templates in template/cache/, disks in images/<vmid>/. Removing a Proxmox storage does not unmount it, so also run umount /mnt/pve/<name>.
  • Boot order: TrueNAS is a VM on the same host as the CIFS mounts, so at boot the mounts and the guests that use them can start before TrueNAS is up. This is not fixed yet.
  • Fixed for good: the pool layout (RAIDZ1), ashift, encryption and case sensitivity. Dataset names and properties like recordsize can be changed later.