A couple of posts were deleted on this thread, perhaps because the authors realized they were mistaken. The claim was that long-polling and WebSockets are bi-directional in the same sense, and mine was that one type of bi-directionality is simulated and has real limitations.

It doesn't matter if the overhead is low: with long-polling, the server can't send a waiting message until the client requests them. With true sockets, the server can send messages as soon as they are available.