This just confirms what I suspected previously. The recurring logics (the part that keeps track) got out of sync with delayed plans (the actual runtime unit representing something to happen in the future). Usually, this manifests itself with the trigger_repeat error, but it is possible that if you stop the workers at the right time, it will just fail silently.
I know this is a rather bad workaround, but would it be possible for you not to take offline backups while actively using the system?