[jlokier]

That works if the backup uses a snapshot of the filesystem or a point in time. Then the backup state is equivalent to what you'd get if the server suddenly lost power, which all good ACID databases handle.

The GP is talking about when the backup software reads database files gradually from the live filesystem at the same time as the database is writing the same files. This can result in an inconsistent "sliced" state in the backup, which is different from anything you get if the database crashes or the system crashes or loses power.

The effect is a bit like when "fsync" and write barriers are not used before a server crash, and an inconsistent mix of things end up in the file. Even databases that claim to be append-only and resistant to this form of corruption usually have time windows where they cannot maintain that guarantee, e.g. when recycling old log space if the backup process is too slow.