Homelab / Architecture

Storage & Backup

Ceph for guest disks, separated file tiers, backup instead of RAID everywhere — plus an offsite target for site resilience.

Compute disks and file data are different problems. Resilience comes from backup and measurable recovery — not from mirroring every drive.

Four roles

VM and container disks on the cluster — separate from file shares. Guest disks Ceph on hypervisor SSDs Documents, pictures, and structured media — the everyday hot path. Hot ZFS TrueNAS shares Large, swappable volumes — media, ISOs, archive, Time Machine. Warm / cold USB OMV · external disks Proxmox Backup Server backs up guest disks — not the file datasets in the NAS. PBS primary backs up guests Selective second backup target at another site — not a full mirror. PBS offsite second site Regular file copies of important shares onto a 1-bay NAS on the LAN. Second medium file copies · Synology Compute / Files Backup

Hover or focus a node for details.

VM and container disks on the cluster — separate from file shares.

Documents, pictures, and structured media — the everyday hot path.

Large, swappable volumes — media, ISOs, archive, Time Machine.

Proxmox Backup Server backs up guest disks — not the file datasets in the NAS.

Selective second backup target at another site — not a full mirror.

Regular file copies of important shares onto a 1-bay NAS on the LAN.

Role For Deliberately not
Ceph on hypervisor SSDs VM and container disks File shares and archives
ZFS hot (TrueNAS) Documents, pictures, structured media shares USB bulk and offsite substitute
USB warm/cold (OMV) Large, swappable volumes — media, ISOs, archive, Time Machine The only home for personal hot data
Second medium (Synology, 1-bay) File copies of important shares and cold data on the LAN Primary NAS, guest backup server, offsite

Physical disks and raw capacity: Storage & Disks. Appliance: Synology DS118.

Why virtual?

TrueNAS and OMV deliberately run as guests on the hypervisor — not as separate bare-metal NAS boxes.

  • One platform: storage roles live in the same compute fleet as everything else (infrastructure as code, backup of the appliance guests) — instead of second physical appliance silos.
  • Role split: hot ZFS stays lean and fast; large USB volumes stay swappable and attach only to the warm/cold guest (passthrough), not to the hot pool.
  • Physical only where required: USB bulk needs hypervisor passthrough; the hot path does not.

Synology remains the exception: a physical second medium on the LAN for file copies — deliberately simple, not the primary NAS.

Backup: guests, files, offsite

Three separate paths — mixing them is the failure mode:

  1. Guests (VM/LXC) → Proxmox Backup Server as the primary target. That backs up the guest disk, not the file datasets inside the NAS.
  2. File shares → snapshots on the hot pool plus a regular copy to a second medium on the LAN (Synology). You still have a copy if the primary NAS is gone.
  3. Offsite → a second backup target at another site, selective by importance and capacity — not a full mirror of everything.

In short: a 3-2-1 target — production, a local backup medium, and a copy outside the site. Service tiles: Services (PBS, PBS Offsite, TrueNAS, OMV, Synology).

Why not RAID everywhere

RAID protects against drive failure inside a running array. It does not protect against deletion, ransomware, or losing the site — and it doubles capacity cost.

So in this lab:

  • Backup and time-to-recovery beat “mirror everything”.
  • Hot paths (ZFS, Ceph replica) carry sensible redundancy; cold/archive and 1-bay volumes may stay simple as long as a second copy exists.
  • Restore targets are pragmatic (hours to days), not an enterprise SLA — validated by restore tests, not marketing RAID levels.

Budget goes to capacity + a second medium + offsite, not to doubling primary storage on every USB disk.

Performance — qualitative

On a fast client link, roughly:

  • ZFS hot — fast enough for daily documents and pictures.
  • USB HDD warm — typical HDD sequential rates; SMB can serve from server RAM after a fresh write and fake disk speed.
  • Backup/cold NAS — deliberately HDD and slower; archive and second copy, not a hot path.

Concrete MB/s and hostnames stay out of the portfolio — only the pattern matters.

What this page deliberately omits

No internal hostnames, IPs, VLAN/VMID plans, backup target paths, or “here’s how you mount it” guides. The portfolio shows the design — operations stay private.

Layer context: Architecture. Site connectivity: Connectivity.