I'm part of a team responsible for cloud-based video editing software. We use multi-master replication (perhaps also known as optimistic replication) with our own tools throughout, but it does require careful design - keep as much data as immutable as possible, every piece of data that might be updated by different machines at the same time should have its own row, GUIDs on each row.
Each machine can generate its own local IDs, which look a lot like a timestamp with some unique stuff on the end. Each row gets a GUID and a 'version' ID column, and we only update relayed database updates if the incoming version is newer. This is largely last-timestamp-wins for the case of conflicts (rare because of design decisions), but there is some Lamport timestamp behaviour in there too for updating a existing row.
The main downside is still that all machines need to handle every write, but with batching up incoming processing into larger transactions, we've had no problems with quite a number of database updates on a dozen commodity machines. Obviously filtering into different shards would be an easy solution.
I'm looking forward to seeing what other people are doing with multi-master replication.