How does AJAX work with non-fixed-length HTTP responses?

Preemptively nthing Comet!

i don't know a lot about comet, but if you need to work with a raw http://livedocs.adobe.com/flex/3/langref/flash/net/Socket.html on the client, this can be done with a small flash movie that communicates with JS via http://livedocs.adobe.com/flex/3/langref/flash/external/ExternalInterface.html. worth considering if comet doesn't do it for you. (btw: though flash's http methods work using the browser's socket, the socket class itself makes direct connections using the plugins own socket. that means any headers that would be appended by the browser for, say, a request by urlloader won't be if you make a request from socket.)

You're describing what the current Java application does, not what the application does. The previous posters are correct that a Comet-style approach is what you want if you need to directly replicate the type of communication currently occurring Alternatively you could just wrap communication of thread #2 in a separate server-side script that holds the "messages" for the client, which could then use normal Ajax to poll this script at a given interval and fetch any pending messages Comet is useful for situations in which you need to send data with greater frequency than it's feasible to poll. If the client can wait 5-10 seconds for new data to come in, normal Ajax polling should work just fine, and creating an intermediary server-side script for Ajax polling to talk to may be be easier than using a Comet approach

Sorry, should clarify: describing the way the java app is implemented, not what the user requirements of the application are

well... it's not really a user requirement that the client behaves the way it does, but it's a functional requirement that any new client uses this sort of two-socket mechanism. And upon re-reading of my post, I realize that the sentence "changing the communication scheme isn't possible" is a pretty vague one. Apologies for not being clearer. I'm looking to do this with as little backend work as possible (the guys in charge of that have enough things to worry about at the moment).

Absolutely this can be done, it is sometimes referred to as "long polling". I have worked on a web application with exactly this architecture and it works great. The web app makes an HTTP request saying "are there any events pending?" If there are, they are sent immediately, and the web app immediately makes the same request again. If there are not, the server doesn't respond at all, leaving the client hanging. When an event occurs, the server sends it right down the open connection if there is one, or else, queues it. Apply a few sensible timeouts and retries here and there, and you end up with a near real-time "push" capability without frequent polling or using anything more exotic that XmlHttpRequest. And as you said, user input from the client-side can go through a separate connection with regular old GETs and POSTs or whatever without causing any problems.