Itai Tavor wrote:
Tres Seaver wrote:
By default, browsers aren't servers: they don't know from bind()/listen()/accept(), which is what push technology requires (unless you have a long-running, bi-directional connection between browser and server, which HTTP doesn't allow). Polling is the only "native" technique for refresh, and suffers from _horrible_ scalability problems.
Options in order of simplicity:
1. Pull instead of push (ick!)
My sentiments exactly.
2. As Hung Jung suggests, build the "push model" stuff separately from the browser, using Python or Java to implement HTTP (and perhaps XML-RPC?) Integration with the browser is difficult: clients will have to have some mechanism for registering their mini-server with your server; if you want the broser's display to update when the push event comes in,you may be stuck.
Right. Mixing two GUI's - a browser for most of the interface and a separate app for pushed data - would be real ugly. So I either have to work inside the browser, as in option 3 below, or code my own GUI.
Registering the mini-servers would not be a problem. This project will be running on a small number of workstations, all located in one location. It's not designed for public or wide-area use and won't even be connected to the internet.
If you control the deployment platform, and the pipe between client and server, then some of my tradeoffs don't apply, as you note: trusted applets are reasonable, for instance, and polling is not nearly as terrible a strategy over a fast network with a small number of clients.
So, the problem with this solution is how to get the pushed data into the GUI.
3. Code a lightweight socket server applet in Java, persuade everyone in the world to trust you to run it in their "applet sandbox", and then "register" it with your server.
Again, because of the physical layout of the project, trust won't be an issue, which makes this solution rather attractive.
4. Write a CORBA callback object in Java, and register it with your server; the server then pushes events to your object (scalability is again an issue, as the server blocks for each client being pushed). Note that adding CORBA into Zope is a non-trivial exercise, for the moment at least; see my "notes":http://www.zope.org/Members/tseaver/CosNaming on the issues.
I read your notes, and I think I'd like to stay away from COBRA - I'd use it if I decided to go totally overboard and code both the server and client in Java. But with Zope on the server end, XML-RPC seems like the best way to communicate.
CORBA is probably overkill, in your case; the reason I introduced it into the discussion is to lead to #5, which gives you scalable event delivery in a heterogenous, wide-area environment.
5. Write a CORBA event-channel applet (CosEventPushConsumer) in Java, set up a CosEvent/CosNotification channel, subscribe the applet to it, and push events to the channel from your server (decouples your server from event-delivery hassles/blockage, scales MUCH better than direct callbacks).
Scalability is not much of a problem for me. I wouldn't like the server to block too much but that shouldn't be much of a problem either - with 10 clients on a 100baseT network, and 4 Zope threads on a fast server, I can't imagine running into any performance problems. But I do want to keep the number of technologies and separate apps involved to a minimum, and this seems fairly complicated.
Out of the various options suggested it seems to me that the main options to consider are:
- Use a browser for the main GUI, and either compromise the requirements to eliminate the need for push or code an applet to display push data in the browser.
- Write my own GUI.
The second option raises some new questions:
- Would it still be a good idea to use Zope for the server, or should I code the server as well (a full Java solution comes to mind using COBRA for two-way communication). How would I handle push in Zope? Assuming I code a push client that speaks XML-RPC, how do I get Zope to send data outside the context of an http transaction? Actually, this same question is relevant for the solution of using a Java applet within the browser.
You might look at Loren Stafford's recently-announced ZScheduler product: Loren is working on a way to have "internally-generated" events fire inside Zope.
- Two good candidates for writing the GUI in are Java and wxPython. Using wxPython would have the advantage of keeping everything in Python. On the other hand, there's an XML-RPC server for Java, but not for Python - I guess I would use HTTP if I use wxPython. Can anyone suggest pros/cons for these solutions?
I think I missed something here: Python *does* have an XML-RPC server implementation (xmlrpclib.py), which is how Zope can do XML-RPC.
I have one other question related to the use of Zope in this project: The server has to monitor a serial line and respond to received data by pushing messages to the clients. What would be the best way to achieve this? An external python process reading the serial line and calling ZPublisher.Client to triger a DTML method that will do the rest of the work?
Right, but ot necessarily a DTML Method. The ZClient request might, for example, trigger a method of a Python product instance, which could enqueue the event for processing by a notification thread, for instance.
P.S. I know this is getting way off topic for the Zope group. I really appreciate all of you contributing your experience.
I don't think this is off-topic at all -- we are pushing the *normal* Zope envelope a bit, but I think Zope is up to the challenge. Tres. -- ========================================================= Tres Seaver tseaver@palladion.com 713-523-6582 Palladion Software http://www.palladion.com