So the new GTalk interface in GMail is pretty rad. Congrats to Dan and the rest of the team that made it “go”.
The talk feature is cool not just from a UI perspective as the code is also chock full of little gems. I’m kind of a dork about low-latency data transport to the browser. HTTP wasn’t meant to be used this way…so of course I’m interested! Ever since Joyce got me involved in the rewrite of mod_pubsub I’ve had my eye on the various ways that servers can push data to browsers and the kinds of technology that will prevent a server that’s doing this from melting down (hellooooooooo Twisted). Using just what’s available to the browser, it’s possible to have the server push data encapsulated in <script> blocks and rely on a progressive rendering behavior that every modern browser implements to dispatch events in near real-time (compared to full page refresh or polling delay). There are a mountain of browser quirks that of course play into this process. The least desirable of these to the user are the “phantom click” and the “throbber of doom” that afflict IE users.
When a page (or an iframe it hosts) is loading content, your browser usually shows some sort of “I’m working” indicator. In the bottom “taskbar” there is usually some sort of progress meter. In the upper right (on IE) the “throbber” will continue to animate until the work is done. Of course in the scenario I’m describing the sent page is never done. The whole point is that the server keeps the connection open. Combine this with the IE behavior of producing a “click” like sound when an iframe is navigated to a different URL, and you’ve got a pretty poor user experience.
But couldn’t you do something with XMLHTTP? Short answer: yes, but not as portably and it won’t get you around IE’s 2-connection limit either so there’s not much of a win. For the long answer, see my talk at ETech or wait for me to post the slides. At the end of the day, the hidden <iframe> hack scales best and is the most portable. Especially if you can lick the UX problems.
Which Google has.
How? By cleverly abusing another safe-for-scripting ActiveX control in IE. Here’s the basic structure of the hack:
// we were served from child.example.com but // have already set document.domain to example.com var currentDomain = "http://exmaple.com/"; var dataStreamUrl = currentDomain+"path/to/server.cgi"; var transferDoc = new ActiveXObject("htmlfile"); // !?! // make sure it's really scriptable transferDoc.open(); transferDoc.write("<html>"); transferDoc.write("<script>document.domain='"+currentDomain+"';</script>"); transferDoc.write("</html>"); transferDoc.close(); // set the iframe up to call the server for data var ifrDiv = transferDoc.createElement("div"); transferDoc.appendChild(ifrDiv); // start communicating ifrDiv.innerHTML = "<iframe src='"+dataStreamUrl+"'></iframe>";
This is the kind of fundamental technique that is critical to making the next generation of interactive experiences a reality. Server tools like mod_pubsub and LivePage (and perhaps even JMS buses) are starting to come into their own and the benefits of event-driven IO are starting to become well understood by server-side devs. It’s only a matter of time before server-push data hits an inflection point in the same way that background single-request/single-response data transfer did with Ajax. Dojo will, of course, have infrastructure to support this kind of thing when the borader developer community is ready (most if it is already in place).
From long and painful experience and amazingly deep respect, I take my hat off and bow to whoever it was on the GMail/GTalk team that figured this out. It’s a hell of a hack. It’s no wonder that Google has been able to attract and develop the best DHTML hackers in the world.
Update: so just to be *very* clear, I worked on the rewrite of the mod_pubsub *client*. The server rewrite was handled by some folks who are much smarter than I am.