[Twisted-Python] conceptually, why a deferred can't be used more than once?

Terry Jones terry at jon.es
Thu Jul 8 16:13:47 MDT 2010


Hi Jean

> Concerning the concept of the deferred, why is it more useful to go with
> a deferred which gets consumed and can only be fired once?

I can't speak for the Twisted devs, but my #1 reason would be that it's
just simpler that way (and Deferreds are already complex enough).

> In my small script I realize I need to take special care that the
> deferred has not been used, and that I must recreate explicitly a
> deferred for each network request. In a parallel world someone might have
> come up with a deferred concept which happily fires the callback as many
> times as there data coming back from the server. Is it a dumb idea?

No, I don't think so. I've done some thinking and playing along these
lines.  At one point I thought about a Deferred class that could be called
multiple times, but decided the code changes were too ugly.

Recently I did some experiments in which Deferreds could be re-used.  The
first observation was that the time taken to create Deferreds is really
negligible. That is, just re-using an already allocated object doesn't save
you much.

More interestingly, the adding of call/errback functions to a Deferred does
take significant time. I don't have the details of what I did handy, but
from memory I managed to get about a 2x speedup in a simple test when
Deferreds were having about 5 call/errbacks added to them.

I wrote a class that allowed you to reset a Deferred, leaving the
call/errback chain on it (snipping off other things that might have been
added to the chain), and there was also a DeferredPool class that could
hand you back an already prepared Deferred (or make you a new one if the
pool was empty). I.e., if you're in a situation where you're repeatedly
using Deferreds and putting exactly the same call/errbacks on them, you
might get a win by resetting and reusing already fired Deferred. On the
other hand, you'd also be a social pariah :-)

If anyone wants more details, I'm happy to dig up my experimental code. It
was fun and it seemed to work just fine.

> Could the deferred design be part of the solution to the network problem
> of two requests passing each other as each ends is not yet aware that the
> other has just sent a request? Buggy networks nodes would expect a
> response but get a request instead and go crazy...

OK, I'll leave that one for someone else...

T




More information about the Twisted-Python mailing list