It’s a tricky problem that goes beyond the fs as you know. Since it’s cloud and since it’s distributed, a manager wouldn’t be that far fetched that could issue CRDT like messages across the cluster of nodes to issue a “write” or utilize the sync mechanism you have to propagate a “master” db that you write to (aggregate or designate, either way). I did some work on this on a go based database graph and ended up doing a gossip sync crdt message bus.

If latency isn't important. Something like this could work. For applications where latency or linearizability is important, it probably won't work.