The VPS now takes a daily logical dump, verifies it, encrypts it, and hands
it to a workstation; a monthly drill proves the archive restores.
- deploy/backup: pg_dump -Fc plus pg_dumpall --globals-only, pg_restore -l
and globals-role checks before anything is promoted, row counts for the
drill, sha256, age encryption into offsite/, retention (3d plaintext,
14d encrypted, 5 pre-deploy sets), JSON logging and a healthchecks.io
heartbeat
- systemd service and timer (03:15 UTC, Persistent, jittered) plus an
OnFailure unit that logs a distinct event
- deploy/activate takes a pre-deploy dump after extracting and before
flipping current, because migrations run in ExecStartPre; a failed dump
aborts the activation, with SKIP_PRE_DEPLOY_BACKUP=1 as the escape hatch
- deploy/workstation/backup-pull: rsync the encrypted archives over SSH,
verify by decrypting and checking the inner sha256, retain 7d/4w/6m, and
fail when the newest archive is older than 48h
- deploy/workstation/backup-drill: restore the newest archive into a
throwaway postgres:18.6, replay the globals dump as the same superuser,
assert both roles, compare users/endgames/attempts, and append the result
to drills.log
- user units for the daily pull and monthly drill, with a shared OnFailure
alert template
- docs: install, key handling, retention, restore, disaster recovery, a run
by hand section, monitoring and troubleshooting
Verified against production: the first backup completed and pinged the
heartbeat, the pull verified the archive, and the drill restored
users=1 endgames=39 attempts=0, matching the live database.