Snapshots Aren't Backups: Building the Lab's Safety Net
The short version
In October 2026 my homelab became load-bearing. It held the domain controller, the DNS for the whole house, and (after I walked away from the cloud drive) the only copy of my family's files, on a self-hosted Nextcloud. It had no VM-level backups; snapshots, optimism, and a nice vSAN were the entire disaster-recovery plan.
The fix was a backup application on a dedicated backup VM, local backup storage on ReFS, restic shipping everything off-site to Backblaze B2 every night, and a separate scheduled backup for vCenter itself. It cost zero dollars in licensing. It has been restore-tested, and its first fully unattended weekend passed every check I set for it.
What "done" meant before I started
Homelab projects die of vague goals, so the requirements went on paper first:
- Cover what matters. The domain controller, Nextcloud, both Pi-hole DNS servers, and the operations appliance. Five VMs.
- vCenter covers itself. The VCSA isn't covered by the backup schedule the way a normal VM is, so it needed its own answer in the plan from the start.
- An off-site copy, encrypted. A backup that lives in the same rack as the thing it protects is only a copy.
- Restore-tested. The build wasn't done until something had been brought back from it.
- Clamped. Dedicated accounts, scoped access, nothing with domain-admin-for-convenience. A backup system is the most powerful box in the building; it gets treated that way.
The design
The backup application runs on its own Windows Server 2025 VM whose only job is backups. Backups land on local storage formatted ReFS with 64K blocks. Block cloning makes the synthetic fulls cheap: the weekly "full" is assembled from blocks already on disk, without re-reading every VM. The Saturday synthetic full for the fleet came to 50.9 GB and finished in 28 minutes with zero warnings and zero errors.
Off-site is a second tool doing one job. A scheduled restic run at 02:00 ships the repository to a private Backblaze B2 bucket, encrypted, every night. Local copy for fast restores, cloud copy for the house-is-gone scenario, different failure domains, no single product trusted with both halves.
vCenter got its own lane: scheduled file-based backups from the appliance itself, landing on an SMB share on the backup server. That share has exactly one account with access, a dedicated local account created for nothing else. The share retains seven sets, and the backup itself is encrypted. Nobody's daily credentials touch it.
The exclusions, which are also design
Two things are deliberately not backed up:
- The VCF license server is excluded because it's rebuildable. If it ever dies, it gets redeployed and relicensed; spending backup budget on it protects nothing.
- Plex media is excluded because it's re-acquirable. What is backed up is Plex's data directory: the configuration, metadata, and watch state that can't be re-downloaded. That backup is restic too, and it has been restore-tested like everything else.
The proofs
The build ran October 15–16, 2026. The first full backup and the first restore test happened the same day. A VM came back from the repository and booted. Only then did the schedule go live.
The real exam was the first fully unattended weekend, October 17–18:
- The Saturday synthetic full completed 5-for-5 VMs, zero warnings, zero errors.
- The 02:00 restic shipment to B2 returned success.
- The first scheduled vCenter backup set landed on its share, on time, without a hand on it.
- The rest of the lab's nightly automation ran through the same window untouched.
That's when I wrote "now it's a properly maintained lab" in my notes, and I meant it as a technical statement.
The wrinkles
Two, for the record. The backup server runs a Windows Server evaluation license, because homelab; its rearm date is on my calendar like any other maintenance item, months ahead. And the whole design rests on the free edition's limits: five VMs is a comfortable fit, but a bigger fleet would mean rethinking the licensing.
The lesson
Everything in this lab had redundancy except recovery. vSAN tolerates a host dying; it does nothing for a fat-fingered delete, a corrupted VM, or a bad patch; those replicate faithfully. The fix was a standard: no system is "done" until it has survived being restored.
Updated Nov 17, 2026: edited the prose for clarity.