Backup and recovery: how I'd back up my homelab

Created 2026-10-12

Backup and recovery: how I’d back up my homelab

What I use today, and where it falls short

I already back up the data I could not recreate, with restic among other tools. I wanted to know which tools best cover three needs of a homelab backup setup:

I reviewed the whole category and then tested the tools that fit these needs, because a comparison table can’t tell me whether a restore works.

The tools

I used the Community Scripts Backup & Recovery category as the candidate list on 2026-10-10: sixteen scripts, three of them marked as testing, plus Proxmox Backup Server, which the site lists under host management and which is the obvious tool for a Proxmox host. Category membership says what I reviewed, not that every entry does the same job. Only eight of the seventeen back up files or guests; three back up databases, and six are archivers, mirrors or sync servers that happen to be filed here.

Resource figures are the installer’s default Debian container allocation, not measured use. Project snapshots come from each repository’s public page on 2026-10-10. “Not described” means I did not find the answer in the project’s own README or features page, not that the answer is no.

Backing up files and guests

Backing up databases

Archivers, mirrors and sync

Which tools fit which job

What I tested

I picked the top two for these needs: Proxmox Backup Server 4.0.10 (G2) and Backrest v1.14.1 driving restic (G1 and G3). On 2026-10-10 I installed each from its pinned community script into a temporary container on the Proxmox host, and backed up the same synthetic data: 1.08 GB of generated files (empty, small and medium files, compressible text, incompressible binary and an exact duplicate), identical for both tools and byte-for-byte reproducible. I never used my real files or guests, and every backup and restore stayed inside the throwaway test.

For each tool I took an initial backup, an unchanged repeat, and a restore to a fresh path, then compared the restored files to the source by file set, SHA-256 and direct bytes. PBS also got a whole-guest test: a throwaway container with the same data and no network device, backed up twice to a temporary PBS storage entry, restored to a new container ID, booted with no network and checked from inside and outside.

File backups, same data, fresh store each

Sizes are decimal megabytes (1 MB = 1,000,000 bytes), measured after flushing and a six-second wait, and “data stored” means the store’s size minus an empty store of the same tool; allocated is what the disk holds, apparent is the sum of file sizes. Both tools deduplicate the unchanged repeat to almost nothing. On this fixture mix restic stored about half what PBS stored. I did not isolate why (restic’s compression and its dedup of the duplicate file are my guesses), and the data includes compressible text, so I would not read “half” as a general ratio. PBS pays a fixed 42 MB for its empty chunk store, which is 65,536 directories.

Whole-guest backup and restore (PBS only)

Each tool took roughly two to three minutes to install unattended, with no manual steps beyond my config. That is a coarse estimate from watching the containers appear, not a timed measurement, and setting up a repository, a schedule and credentials by hand would take longer.

All numbers are single runs on one quiet test machine with synthetic data, so they show the shape of each tool, not a ranking. They say nothing about a network target, a large photo library, encryption cost or restore from an object store.

What I learned

• PBS, vzdump and temporary directories: The first whole-guest restore “failed” because my checksum list lived in a temporary directory. Proxmox’s guest backup to PBS excludes the temporary directories on purpose, so anything kept there never gets backed up. The restore was fine; my test was wrong. It cost one run.
• PBS, the enterprise repository: The community install script leaves the paid enterprise package source switched on. Without a subscription it answers 401 and apt-get update exits non-zero, which breaks any follow-up package install until I disable that source in the container.
• PBS, the client: The server package does not include the backup client. For file backups I had to install the client package separately.
• PBS, an empty store is not empty: The first 42 MB of “storage used” is the chunk directory skeleton, so I measure the growth over an empty store, not the total.
• Backrest, setting it up through its API: In v1.14.1 adding a repository fails until an instance name is set, and the autoInitialize option creates the repository and then fails its own validation. I initialised the repository with the restic that Backrest manages, then added it without that option. A scripted install needs this.
• Measuring size on ZFS: Reading disk use right after a write showed the repeat backup bigger than the first one, because ZFS delays allocating blocks. I now flush, wait six seconds, flush again, and record both allocated and apparent bytes.
• Proxmox storage commands touch the config file: Adding and removing a temporary storage entry rewrote the storage configuration with the same settings in a different order, left a client cache file, and created an empty credentials directory. None was a real change, but a cleanup check that compares bytes will fail on a clean run.

Seven runs were needed, because each exposed one gap in my own test harness. Every aborted run left no containers or volumes behind. The final run’s automated cleanup passed its own checks. In the earlier runs a few small leftovers (a reordered storage configuration, a client cache file and an empty credentials directory) needed manual removal, which I checked afterwards. I never touched anything outside the test.

What this means for my setup: what I’d do

These are proposals, not a done deal.

Cost of the offsite part. On 2026-10-10 two object-storage providers’ public price pages listed standard storage at about $0.015 per GB-month (roughly $15 per TB-month) with free downloads on one, and about $6.95 per TB per 30 days on the other, with free downloads up to three times the amount stored and $0.01 per GB beyond. Both give 10 GB free. PBS can also use an S3-compatible bucket as its datastore, and its documentation warns that request charges can add cost; I did not measure those. Pricing changes, so I would re-read the pages before I decide.

What would change this. If I want one interface for every backup, Backrest or Zerobyte beats raw restic. If the guests multiply, the PBS machine matters more. If PostgreSQL restores become the worry, Databasus is the one tool here that tests its own restores. Adopting any of this is a separate project with its own plan.

Limits

I reviewed the pages and repositories listed above on 2026-10-10 and read each community installer at the level of “what does it write on the host”, not line by line. I tested two tools on one data set; I did not test encryption, a remote target, a large library, an object-store restore or Zerobyte, Duplicati, Borg-UI or the database tools. Fit statements for tools I did not install are reasoning from their documentation, not a demonstrated deployment. The temporary containers, storage entry and scratch space were removed after the test.