Diesel: A Case Study In That Thing I Just Said

Thursday September 24, 2009
Thanks to jamwt for the shout-out on the announcement of Diesel.

Since the reaction to my reaction to tornado was so good (or at least so ... energetic), I figure I should comment on Diesel as well.  Spoiler alert: my reaction is ... largely similar, but since jamwt has been kind of nice to Twisted in the past, and didn't actually say anything mean this time, I'm somewhat reluctant to have that reaction.  Nevertheless, I swore a solemn oath to tell it like it is, keep it real, and soforth.  So I must.

Once again, I'm happy that event-driven programming is getting some love.  This time, I'm pleased that nobody is saying anything especially snarky or FUD-ish about Twisted.  I do feel like it's a little weird not to mention Twisted, or include some comparisons to Nevow or Orbited, both of which provide different, comprehensive approaches to COMET with Twisted.

(Worth noting: Orbited also originally started out using its own event-driven I/O layer, but switched to Twisted later, because Twisted is "crazy delicious".)

Diesel has many more interesting ideas at the level of async I/O than Tornado did.  I think the generator-based approach for implementing protocols is interesting and deserves some more exploration.  I'm not sold on it for every use-case, and I think the implementation might have some flaws, but it definitely has some advantages.

I'd give jamwt a hard time for not reporting issues and communicating with Twisted more before re-writing the core, but for three issues:
  1. jamwt's been around in the Twisted community for a while.  He's written a bunch of fairly deep Twisted code and he clearly knows what the framework is capable of.
  2. I've spoken with him on a number of occasions, and for all I know I might have discussed this with him.  I don't remember it, but it would be pretty embarrassing to write a big rant about how nobody talks to us only to have him paste some chat log where he explained why he was writing Diesel six months ago, and I said "oh, okay" ;-).
  3. Nobody is calling Twisted names or making vague, unsubstantiated accusations.  You're not obligated to examine Twisted, nor Nevow, nor Orbited, I just feel that you owe us some explanation if you publicly say that you tried it and found it wanting.  The tone on the Diesel announcement, in its one brief mention of Twisted, is "we tried it, but we kinda wanted to do our own thing".  So, good for them, they did their own thing, I hope they had fun.
Now, personally, I'd like to leave it at that, but there is a certain inevitable comparison that I think is going to take place.  Diesel has a nicer web page than Twisted.  They have entwittered ... twitified ... uh ... tweetened ... the project, and we haven't; we just have an old-fashioned "blog".  Diesel is smaller than Twisted, so it's easier to explain, and so the people approaching it will have a better idea of its scope.  This might give the immediate impression that it is a simpler, better, more "modern" replacement for Twisted's I/O layer, and this is not the case.  So I still feel it's important that I set the record straight.

Before I launch into my critique, I should say that I don't want to harsh on Diesel too bad. It's a neat little hack and you should go play with it.  And I feel bad pointing out problems with it, since as I mentioned above, nobody's dumping on Twisted.  So, Diesel fans, please take this in the spirit of a frank code-review, not a complaint about your behavior.

The interesting generator-munging bits could be easily adapted to run on top of Twisted's loop, which, arguably, they should have been in the first place; and the toy "hub" that they've written might be good enough for some simple applications where reliability under load is not a serious concern.  In fact, inlineCallbacks might provide a good deal of what is needed to support Diesel's programming style.  Alternately, Diesel might provide some hints as to how things like inlineCallbacks could be made more efficient.

That said, Diesel's I/O loop sucks.

It's disappointing to see the same mistakes getting made over and over again.  First and foremost: no tests.  Come on, Python community!  You can do better!  Write your damn tests first!

The #1 benefit that a brand-new I/O loop project could have over Twisted is that Twisted was written in the bad old days before everybody knew that TDD was the right way to write programs, so we don't have 100% test coverage.  But, we strive to get closer every day, while every new project decides that they don't need no stinking quality control.

Predictably, as it has no tests, Diesel's I/O layer is full of dead code, inaccurate  documentation, and unhandled errors.  Consider this gem, which I found about 30 seconds into reading the code: KqueueEventHub is documented to be "an epoll-based event hub", and its initializer defines an inner function which is never used.  I'm not going to belabor the point by enumerating all the typo bugs I found, but you may find the output of 'pyflakes diesel' interesting.

Instead of Tornado's inaccurate handling of EINTR, Diesel has no handling of EINTR, as far as I can tell.  It also doesn't handle EPERM, ENOBUFS, EMFILE, or even EAGAIN on accept().  To be fair, it has a catch-all exception handler all the way at the top of the stack, so none of these will cause instant crashes, but they will cause surprising behavior in odd situations (and possibly infinite traceback-spewing loops).

More surprisingly - I had to re-read the code about five times to make sure - it doesn't appear that sockets are ever set to be non-blocking, and EAGAIN is not handled from accept(), recv(), or send().  And yes, this can happen even if your multiplexor says your socket is ready for reading and/or writing.  The conditions are somewhat obscure, but nevertheless they do happen.  So, occasionally, Diesel will hiccup and block until some slow network client manages to send or receive some traffic.  In other words: Diesel is not really async.  It just fakes it convincingly, most of the time.

Once again, there's no way to asynchronously spawn a process, and no way to asynchronously connect a TCP client.  Sure, this looks like an asynchronous connect call, but it's misleading: it blocks on resolving the hostname, and it potentially blocks on the initial SYN/ACK/SYN+ACK exchange.  There's no asynchronous SSL support.  And no, that is not trivial.  Not to mention handling all the crazy errors that spew out of the Windows TCP stack.  And since the loop is implemented to be incompatible with Twisted, it's not obviously trivial to compatibly plug it in and get those features.

Again, I don't want to dump on Diesel here; for what it is, i.e. an experiment in how to idiomatically structure asynchronous applications, it's all right.  For that matter Twisted has its fair share of bugs too, which would be pretty easy to lay out in a similar post; you wouldn't even need to do the research yourself, just go look at our bug tracker.

But both Diesel and Tornado make the mistake of attempting to replace the years of trial-and-error, years of testing discipline, and years of portability and feature work that Twisted has accumulated with a few oversimplified, untested hacks.

What they could have done is contributed any extensions that they needed to Twisted's loop, or modifications to Twisted's packaging that would allow them to get a smaller sliver of Twisted's core to bootstrap, if that's what they needed.

My goal in pointing out all these flaws is not to illustrate any particular point about Diesel, but to reinforce the point I implicitly made in my Tornado post, which is that if you try to write a new mainloop (especially without tests) you will screw it up.  You will most likely screw it up in ways which will only surface later, under mysterious circumstances, when your servers are under load and you are under the gun for a deadline.

Or if I happen to get wind of it and write a blog post about it, of course.  Then you get to cheat a little.

It's not an indictment of Diesel that it screwed this up; everyone screws it up.  I would probably screw it up, if I didn't have Twisted sitting in front of me as a direct reference.  POSIX by itself is unreasonably subtle and difficult, but POSIX, plus the subtle variations in different platforms which implement it, plus the Windows APIs which are almost-but-not-quite-exactly-nothing-like the POSIX APIs, presents an inhuman challenge.

Hopefully Diesel will grow some tests.  Hopefully it will fix, or better yet shed, its somewhat unfortunate I/O hub.  I am hopeful that someone will follow Dustin's excellent lead (perhaps Dustin himself!) and port Diesel's API and generator system over to Twisted's I/O architecture and eliminate all these silly bugs.  Of course, it someone did that, you could use Dustin's tornado port with Diesel.

With the silly bugs from the I/O loop out of the way, the Diesel team can write tests for the more interesting pieces, and fix the bugs which aren't entirely silly :-).