The Long Poll: AJAX Push(like) Chat with Comet

Recently I’ve been working on an AJAX based chat application (in development..). The obvious way to do it is send an XMLHttpRequest every few seconds to check for new messages. Unless it’s a particularly animated conversation most requests won’t return any new content, so I added a simple Conditional-GET like system based on the chat’s text size. Here’s the client side implementation:

function refresh_chat() {
	$.ajax({
	  	url: "/chat",
	   	data: "format=xhr&chat_id={{chat_id}}&cur_len=" + chat_content.length,
		  complete: function(xhr){					
				if (xhr.status == 200) render_chat(xhr.responseText);
				setTimeout("refresh_chat()", 5000)
		  }
	 });	
}

And the server code that handles it:

cur_len = self.request.get("cur_len", 0)
if len(chat.content) == int(cur_len):
	self.error(304) # return 304 Not Modified
else:
	self.response.out.write(chat.content) # return new content

That’s basically the standard approach. Pretty simple, works ok (could be optimized a bit, for example return only the actual new content etc). It’s not exactly an elegant design, though. Trying to use HTTP, designed as a Pull protocol, for an application that requires Push results creates this system of frequent server requests with empty responses, kind of like the “Are we there yet?” conversations with kids on long road trips.

Jack Moffitt’s JSConf talk introduced me to the concept of Long Polling, aka Comet or (with a lot added) BOSH, as a way to simulate HTTP Push. Rather than have the client sending a lot of short, frequent requests and the server responding to each as fast as possible, long polling turns it around: the server holds the requests as long as it can, returning a response only when it has new data or a timeout limit was hit. So, instead of sending request every 3 seconds, for example, you can send one every 30 seconds.

Client side code remains almost the same:

function refresh_chat() {
	$.ajax({
	  	url: "/chat",
	   	data: "format=xhr&chat_id={{chat_id}}&cur_len=" + chat_content.length,
		  complete: function(xhr){					
				if (xhr.status == 200) render_chat(xhr.responseText);
				setTimeout("refresh_chat()", 1000);
		  }
	 });
}

But on the server side, there’s a bit of new logic to keep checking for new content while the server holds the response:

cur_len = self.request.get(“cur_len”, 0)
end_by = int(time.time()) + 30

while int(time.time()) < end_by: if len(chat.content) != int(cur_len): return self.response.out.write(chat.content) # return new content time.sleep(1) self.error(304) # return 304 Not Modified [/sourcecode] If you have any experience building web applications, you've spent a lot of effort making sure servers respond quickly to requests. Delaying the response is counter-intuitive, which in itself makes Comet useful to know, if only for its new perspective. However, this also makes production use a bit complicated, since most web server stacks are optimized for maximum requests/second rather than long concurrent requests. Content-rich sites often use separate servers for big media content for this reason, and Comet also has its own server (er "HTTP-based event routing bus") in Cometd.