The Twilight Sun

Aspect

Cooking up Dishes

by Jonathan on Aug.08, 2010, under Aspect, Web programming

I’ve just begun building the messaging model for Aspect, and I’m developing it as a separate, reusable Rack-based framework. It’s called Dishes, which is an extraordinarily lame pun derived from “asynchronous” sounding like it has a “sink” in it. So far things are looking good, but there’s a lot of work still ahead!

As it happens, I read up a little on BOSH, and it uses exactly the same two-connection model I described in my last post. It’s interesting reading and I’m sure to take a lot of inspiration from it. I don’t really want to adopt BOSH in its entirety because I think it’s way more than I need; Dishes is probably going to be fairly small.

Something cool I’m trying in Dishes is putting each request in its own Fiber. Since Aspect is going to be extremely I/O bound, it makes sense to re-invert the evented model using fibers to make handlers feel more synchronous. That will make it easier to use Dishes.

Anyways, that’s enough late-night rambling for tonight. When I have a simple working Dishes example, I’ll post again and explain how everything works. Meanwhile, you can always take a look at the current code. I even have a basic goal application to aim for.

Comments always welcome!
~Jonathan

7 Comments :, , , more...

Aspect’s messaging model

by Jonathan on Jul.28, 2010, under Aspect, Computers, Programming

(Before I begin, I should warn you that I haven’t had much time to actively work on this yet.)

I found an article about Starbuck’s method of handling coffee orders on Hacker News today. I was surprised, because I had explained this exact concept to some friends already, but using Pat & Oscar’s instead! This kind of query-response architecture strikes me as the perfect model for Aspect, because everything you do constitutes a request, and everything Aspect tells you is some form of response.

The Pat & Oscar’s analogy is a little different from the Starbucks one, though, and it highlights a few key points. If you’ve never been to a Pat & Oscar’s, it works like this: You go to the counter, make your order, receive a number, sit down at a table, and place the number card in the little holder on the table so they can find you. When the order’s ready, they come to you with the food. Importantly, you can make multiple orders and receive multiple numbers, and the orders will be served whenever they’re ready.

Now, how do you apply this to web communication? The classic HTTP protocol has a strictly blocking request/response format, meaning that every request must wait until the response is sent before you can reuse the connection. Most browsers have a cap on how many active connections you can have, and the bare minimum is IE’s two connections. So we need to make this work by using only two connections.

The solution is to use one connection for making requests, and the other for receiving results! You keep the results connection open constantly, and use the other whenever you make a request. The request connection must be closed as quickly as possible, so you just return an “order number”, a job ID. When the job is done, it’s pushed down the response connection along with its job ID. Then you re-connect the response connection so it’s constantly open. The approach the response connection is using is called long polling. Instead of polling periodically, the server hangs on to the connection until it has something to send.

I believe that this is a powerful approach to Comet-like communication. Unlike pure long polling – where the server simply holds the connection instead of responding immediately, and the client doesn’t need to care – it does require some infrastructure on both sides of the network gap. But if it can be abstracted properly, it should really be very easy to use. It’s just a layer you can work on top of.

I’ll be working on this as time permits, and of course I’ll open-source my work on this (though probably not Aspect itself) after a certain point. I’m a firm believer in open-sourcing platforms so everyone can benefit. If anyone else is interested in helping though, I’d be glad to make it public sooner! I think this is a useful model that definitely has applications beyond Aspect.

~Jonathan Castello

1 Comment more...

A little update on Aspect

by Jonathan on Jul.02, 2010, under Aspect

So it’s been a while since I’ve posted about Aspect. I was talking with a friend who had read up on Aspect here, and it seems I had forgotten to mention a rather key change I made: switching to Ruby. So I’ll take some time to explain exactly what’s up with Aspect right now.

Previously, I had explained that I was using Python, using the Tornado server. Unfortunately – and sorry, Python-lovers – I can’t stand Python. Development got to a point where I couldn’t make any progress because I was fighting the language. I had originally planned on using Ruby, in fact, but at the time I couldn’t find any Ruby libraries that did what I wanted.

Apparently I just didn’t look in the right place, because once I started looking again, I immediately found EventMachine, a Reactor-based library. EventMachine is pretty awesome, and suits my needs perfectly. But I still needed an HTTP server, and preferably one that could handle lots of concurrent conncetions. Thin fits the bill nicely. And lastly I needed a framework that could deal with holding onto connections until I have content ready. So far, Cramp handles that perfectly well.

Recently I’ve tossed Nginx into the mix, since I’m hoping I can have multiple Thin workers behind a front-end Nginx. Nginx also uses the event-based model, and it doesn’t give a thread to each connection. That would bring my server an early death, I think.

So that’s my network pipeline. From start to finish, it’s built to handle multiple concurrent, persistent connections, and it should be scalable too. Next time I’ll post about the messaging layer I’m building between the user and Aspect.

2 Comments : more...

First Thoughts on Nginx

by Jonathan on Jun.26, 2010, under Aspect

I’ve been experimenting a bit with Nginx lately. I haven’t had a chance to actually benchmark it – nor, in fact, do I have any idea how I would actually do that – but I’m pretty happy with the learning curve. My only problem was with getting Passenger working, since I use RVM to manage my rubies. It turned out to be as simple as “rvm ruby –passenger” and setting the passenger_ruby configuration option.

Nginx seems to be a better choice to run Aspect behind than Apache, because Nginx is built to manage simultaneous connections. Just like Thin, it’s event-driven instead of thread-based, and I’m going to have plenty of long-running (and recurring) connections.

Now that I have a local server running on my computer, too, I think I’ll have an easier time juggling my projects. I recently discovered /etc/hosts, too, and I just can’t describe how much nicer it is to use “aspect.local” instead of “127.0.0.1″ all the time.

In other news, the 9 key on my laptop broke. You may also know this key as the “left parenthesis” key. Now… If you’ve done any programming at all, you should know just how important this key is. Gah!!

Leave a Comment :, , more...

Aspect (plugin infrastructure)

by Jonathan on Mar.23, 2010, under Aspect

This is still entirely conceptual at the moment, so don’t get too excited. ;) Also, if you’re not too into technical stuff, you might want to read only the first paragraph.

I’m planning on making Aspect extensible to a certain extent. Users will be able to create plugins written in Lua, run on the server in a highly sandboxed environment, with a specialized API so the plugins can do useful stuff. One such API will probably be a widget interface, to create and manage HTML widgets on the browser side. This means you could create plugins that manage a visual set of stat gauges, or show a visual map, etc etc. in a similar vein to MUSHclient and Mudlet.

Now, plugins will be run on the server. This is because (a) I don’t think you can run Javascript sandboxed very easily, (b) it makes it much harder to communicate between the plugins and Aspect and the MUD, and (c) Lua is just so much nicer to script this kind of stuff with. For speed and manageability, I’ve also decided that all plugins will share the same Lua interpreter space. You have to be careful with security in all cases here, because third-party scripts will be running on the Aspect server. You don’t want arbitrary plugins to share player info, or clash with eachother in other ways.

Obviously I have to only give scripts access to utilities I want them to have access to. But what if the plugin goes into an infinite loop, or just takes too much time to execute? In the best case, it will lag everyone else connected to Aspect. That’s clearly not a good thing!

There is functionality in Lua to set a “hook” which can be called automatically every so often. Theoretically, I can use this to halt execution, go do other stuff, then come back to the plugins. But we also want some way to keep plugins completely separate, and we don’t want any one plugin to hog all the time so no other plugin gets a chance. It’s a bit of a sticky situation.

First of all, we can isolate each plugin by starting each one in a coroutine, and setting the coroutine functions environment to its own special table. Coroutines are like threads, but they’re cooperative rather than preemptive. That is to say, a coroutine has to say “Okay, I’m going to take a break now” or “Hey, you, do some work” explicitly in order for another coroutine to begin. And changing the environment keeps the plugins from sharing any state.

Now we’re down to one more problem (or at least the last major one for now). I mentioned before that I don’t want any one plugin to hog all the time. I also mentioned hook functions. It just so happens that you can set a hook function separately for each coroutine. If we could yield from that hook, it would be almost like preemptive threads. And while you can’t yield from a hook set with debug.sethook (in Lua), you can from a hook set with lua_sethook (in C). So I just have to figure out how to write C-based hook functions to yield every so often.

So that’s a rather verbose explanation of how I want to implement plugins. Well, it’s 2am and I needed to talk. :P Tl;dr summary: every plugin will be a completely isolated Lua coroutine that shares execution time with other plugins. Plugins will be able to do a lot of interesting things, like create and manage visual widgets on the user’s browser, like status bars and maps and stuff, and could even augment/replace the default output window and input bar. It’ll be extremely extensible.

Now if only I could figure out how lua_yield works.
~Jonathan

1 Comment :, more...

Looking for something?

Use the form below to search the site:

Still not finding what you're looking for? Drop a comment on a post or contact us so we can take care of it!

Visit our friends!

A few highly recommended friends...