Overview of how to plan backups for a homelab or small server estate: the 3-2-1 rule, recovery objectives, what to back up, and a restore-testing checklist. The tool-specific sheets hold the commands; this one holds the thinking. A backup you have never restored is a hope, not a backup.
Documentation: Backblaze: the 3-2-1 backup strategy · restic docs · BorgBackup docs · borgmatic
Verification: This sheet is conceptual. The tool commands live in the linked sheets, each of which has its own verification note. No command was executed.
See also: restic · BorgBackup · borgmatic · rclone · rsync · ZFS Snapshots and Replication · Docker Volume Backup · Database Dumps · Proxmox Backup and Restore · TrueNAS Community Edition · Linux · Docker
| RULE | MEANING |
|---|---|
| 3 copies of the data | The live data plus at least two backups |
| 2 different media or storage systems | A second disk, NAS or a different filesystem, so one failure mode does not take both |
| 1 copy off site | Another building, another provider, or cloud object storage |
Common extensions:
- 3-2-1-1-0: one copy offline or immutable (append-only, object lock, air-gapped), and 0 errors after verification.
- A snapshot on the same pool is not a backup. It protects against deletion and ransomware on a share, not against pool loss.
- RAID is availability, not backup. It replicates mistakes instantly.
| TERM | MEANING | DRIVES |
|---|---|---|
| RPO (Recovery Point Objective) | How much data you can afford to lose, expressed as time | Backup frequency |
| RTO (Recovery Time Objective) | How long recovery may take before it hurts | Tool and storage choice, restore speed, documentation |
Examples: a nightly backup means an RPO of up to 24 hours. A 2 TB restore over a 100 Mbit/s link takes about 44 hours (2 TB / 12.5 MB/s), so an RTO of "a day" needs a local copy as well as the cloud one.
| INCLUDE | NOTES |
|---|---|
| Application data and volumes | Bind mounts and named volumes, see Docker Volume Backup |
| Databases | Logical dumps or a stopped/snapshotted data directory, see Database Dumps |
Compose files, .env files, configs |
Keep in git where possible, but secrets files need a backup too |
/etc and service configs |
Reverse proxy, DNS, firewall, systemd units |
| Encryption keys, passphrases, repo keys | See "Keys and credentials" below |
| VM and container images | Hypervisor-level, see Proxmox Backup and Restore |
| Package lists and bootstrap notes | dpkg --get-selections, Ansible playbooks, so a rebuild is quick |
| EXCLUDE | WHY |
|---|---|
Caches, node_modules, .cache, thumbnails |
Rebuildable, large |
| Media you can re-download | Judgement call, decide per library |
/proc, /sys, /dev, /run, /tmp |
Virtual or transient |
| Docker image layers | Pull them again; keep the Dockerfile and compose file instead |
Live database files copied with cp |
Inconsistent, use a dump or snapshot |
- The backup passphrase and any key file must exist somewhere other than the backed-up host: a password manager (for example Bitwarden Cheat-Sheet) plus a printed or offline copy.
- Borg: export the key with
borg key export. restic: the password is the only key. Lose it and the repository is unreadable. - Do not keep the only copy of the repo password inside the repo it protects.
| NEED | TOOL |
|---|---|
| Deduplicated, encrypted, versioned file backups to local, SFTP or cloud | restic or BorgBackup (+ borgmatic) |
| Mirror or sync a directory to cloud storage | rclone |
| Simple file mirror or hardlinked snapshots over SSH | rsync |
| Filesystem-level snapshots and replication | ZFS Snapshots and Replication |
| Docker volumes | Docker Volume Backup |
| Databases | Database Dumps, or borgmatic's native database hooks |
| Whole VMs | Proxmox Backup and Restore |
rclone sync and rsync --delete mirror deletions. They are copies, not versioned backups, unless you add --backup-dir or --link-dest.
| LAYER | FREQUENCY | RETENTION |
|---|---|---|
| Local snapshots | Hourly | 24 hourly, 7 daily |
| Local backup repo | Daily | 7 daily, 4 weekly, 6 monthly |
| Off-site repo | Daily or weekly | 7 daily, 4 weekly, 12 monthly, 2 yearly |
| Database dumps | Before the file backup runs | 7 to 14 days on disk, the file backup keeps the history |
- Alert on failure and on silence. A dead timer sends no failure email. Use a dead-man's-switch such as Healthchecks.io (
curl -fsS -m 10 --retry 5 https://hc-ping.com/<UUID>), which borgmatic supports natively, see borgmatic. - Log backup size and duration. A sudden size drop usually means something is not being backed up.
- Run integrity checks on a schedule, not just once (
restic check,borg check,zpool scrub).
Do this at least quarterly and after every major change to the backup setup.
- Can you reach the backup from a clean machine with only the documented credentials (not the original host's config or keys)?
- List snapshots or archives and confirm the newest one is as recent as expected.
- Restore a single file to a scratch directory and compare it:
diff -ror checksums against the live copy. - Restore a whole application (volume plus database) into a throwaway Compose project and confirm it starts with the right data.
- Restore a database dump into a scratch instance and run a row count or
PRAGMA integrity_check, see Database Dumps. - Time the restore. Compare against your RTO.
- Run the repository integrity check (
restic check --read-data-subset=5%,borg check --verify-data). - Confirm the off-site copy is restorable too, not just the local one.
- Confirm alerts fire: temporarily break a backup and check you are told.
- Write down what you learned in the runbook.
# Quick sanity restore pattern (replace the tool line with restic/borg/rclone/rsync)
SCRATCH="$(mktemp -d)"
restic restore latest --target "$SCRATCH" --include <PATH>
diff -r "$SCRATCH/<PATH>" /<PATH> && echo "restore matches"
rm -rf "$SCRATCH"- Where the backups are (repository URLs, bucket names, hosts).
- Where the credentials are (password manager entry, not the secret itself in a git repo).
- The exact commands to list and restore, tested.
- Order of rebuild: network and DNS, storage, hypervisor or Docker, then applications, then data.
- Who to contact and what the acceptable downtime is (RTO).