their snapshots are marked as copy on write, so my assumption is that for every write, there is replication somewhere.

I might be wrong here, but my understanding is that their snapshots are the kind we find in modern filesystems (ZFS/BTRFS/…) : that is a point-in-time kind of functionnality, where a file will be duplicated (and the original version then will only belong to the snapshot) only when it is written to. This is just the way snapshots are implemented here - and a rather common way of doing it efficiently - not a reliability feature.