A single kafka partition should be able to handle the load - it only needs to write 333 ops/second, and I could coalesce several edits together over half second windows if I wanted to make that more efficient. And I'll only have as many kafka consumers as I have frontend servers. So we're talking about 333 ops/second and like, 5 consumers of those operations. Somehow I think a single kafka partition should be able to manage the load :)

But there's no reason I couldn't use multiple partitions if I had to. To make it fully consistent I'd just need a consistent way to resolve conflicting concurrent writes on different partitions. Either adding a server timestamp to the messages, or using the partition number to pick a winner would work fine for this. It would just make the server logic a little more complicated.