Backups are not a checkbox

Why we stop the server to take a backup, and the three ways a "backup" turns out to be useless when someone actually needs it.

Every host advertises daily backups. Very few of them have tried to restore one under pressure. We have, both for customers and for our own worlds, and the failures are consistent enough to be worth writing down.

Failure one: the backup is a copy of a corrupted world

If a server dies badly, the world files on disk may already be inconsistent. A backup tool that copies the directory after the crash preserves the damage. Stopping the server first, waiting for the save to complete, and only then copying is slower — it costs about thirty seconds of downtime — but it is the difference between a restore and a second copy of the same problem.

Failure two: the restore is missing half the files

Valheim writes two files, .fwl and .db, and a restore that brings back only one of them produces a world that loads empty. Project Zomboid keeps the player database outside the map directory. ARK players live in a separate folder from the map. Any backup that was defined as “the world folder” rather than “the paths this game actually needs” will be wrong for at least one game.

Failure three: nobody checked

A backup that has never been restored is a hypothesis. We run a restore of a random instance every week, on purpose, and check that the server starts and the world appears. It is a boring job that has caught two real problems in two years — both of them in the backup path, not in the game.

What we do now

  • Stop, save, then copy. Never copy a live world directory.
  • Back up the game-specific path list, not “the folder”.
  • Checksum the copy and verify it before the snapshot is listed as valid.
  • Restore a random instance every week.
  • Keep the retention window per plan, so the number in the pricing table matches the number in the panel.

None of this is clever. It is just the difference between having backups and being able to use them.

← All posts