You basically had to engineer around the corruption risks for anything actually in production. Data redundancy, old-school 'Save' buttons (your work isn't properly saved until you click it, because it sends copies of that work to multiple backends) and isolating users as much as possible to their own 'shards' was part of how I saw it kludged through in the real world.

There was still a lot of weird behavior, though, and while you could reduce data loss you couldn't eliminate it.