Raise the start_period to high duration because of the way the container
starts up (pretty much runs the entier setup from scratch with Chef).
Now with the new Compose project health check script, it should not fail
as long as the state is starting instead of failing.

- The command silently failed for ages (the container name was missing).
  When I had less information in GitLab it wasn't noticeable, but that's
no longer the case.
- Delete backups that haven't been accessed instead of modified. In case
  I did get something from that backup, keep it around for a while
longer.
- Keep backups for 10 days, in case it's been a busy week or over the
  weekend or something.