It's been a while since I've checked this but a few years ago we tried to limit test kine on a large-ish cluster and it performed pretty poorly. It's fine for small clusters but the way they have to implement the watch semantics makes it perform poorly (at least this was the case a few years ago).