[Web-SIG] PEP 444 / WSGI 2 Async

[Alice Bevan–McGregor]

On 2011-01-06 10:15:19 -0800, Antoine Pitrou said:

> Alice Bevan–McGregor <alice at ...> writes:
>>> Er, for the record, in Python 3 non-blocking file objects return None when
>>> read() would block.
>> 
>> -1
>> 
>> I'm aware, however that's not practically useful.  How would you detect
>> from within the WSGI 2 application that the file object has become
>> readable?  Implement your own async reactor / select / epoll loop?
>> That's crazy talk!  ;)
> 
> I was just pointing out that if you need to choose a convention for 
> signaling blocking reads on a non-blocking object, it's already there.

I don't.  I need a way to suspend execution of a WSGI application 
pending some operation, often waiting for socket or file read or write 
availability.  (Just as often something entirely unrelated to file 
descriptors, see my previous post from a few moments ago.)

> By the way, an event loop is the canonical implementation of 
> asynchronous programming, so I'm not sure what you're complaining 
> about. Or perhaps you're using "async" in a different meaning? (which 
> one?)

If you use non-blocking sockets, and the WSGI server provides a way to 
directly access the client socket (ack!), utilizing the none response 
on reads would require you to utilize a tight loop within your 
application to wait for actual data.  That's really, really bad, and in 
a single-threaded server, deadly.

> I don't understand why you want a "yield" at this level. IMHO, WSGI 
> needn't involve generators. A higher-level wrapper (framework, 
> middleware, whatever) can wrap fd-waiting in fancy generator stuff if 
> so desired. Or, in some other environments, delegate it to a reactor 
> with callbacks and deferreds. Or whatever else, such as futures.

WSGI already involves generators: the response body.  In fact, the 
templating engine I wrote (and extended to support flush semantics) 
utilizes a generator to return the response body.  Works like a hot 
damn, too.

Yield is the Python language's native way to suspend execution of a 
callable in a re-entrant way.  A trivial example of this is an async 
"ping-pong" reactor.  I wrote one ("you aren't a real Python programmer 
unless...") as an experiment and utilize it for server monitoring with 
tasks being generally scheduled against time, vs. edge-triggered or 
level-triggered fd operation availability.

Everyone has their own idea of what a "deferred" is, and there is only 
one definition of a "future", which (in a broad sense) is the same as 
the general idea of a "deferred".  Deferreds just happen to be 
implementation-specific and often require rewriting large portions of 
external libraries to make them compatible with that specific deferred 
implementation.  That's not a good thing.

Hell; an extension to the futures spec to handle file descriptor events 
might not be a half-bad idea.  :/

> By the way, the concurrent.futures module is new. Though it will be 
> there in 3.2, it's not guaranteed that its API and semantics will be 
> 100% stable while people start to really flesh it out.

Ratification of PEP 444 is a long way off itself.  Also, Alex Grönholm 
maintains a pypi backport of the futures module compatible with 2.x+ 
(not sure of the specific minimum version) and < 3.2.  I'm fairly 
certain deprecation warnings wouldn't kill the usefulness of that 
implementation.  Worrying about instability, at this point, may be 
premature.

>> +1 for pure futures which (in theory) eliminate the need for dedicated 
>> async versions of absolutely everything at the possible cost of 
>> slightly higher overhead.
> 
> I don't understand why futures would solve the need for a low-level 
> async facility.

You mis-interpreted; I didn't mean to infer that futures would replace 
an async core reactor, just that long-running external library calls 
could be trivially deferred using futures.

> You still need to define a way for the server and the app to wake each 
> other (and for the server to wake multiple apps).

Futures is a pretty convienent way to have a server wake an app; using 
a future completion callback wrapped (using partial) with the paused 
application generator would do it.  (The reactor Marrow uses, a 
modified Tornado IOLoop, would require calling 
reactor.add_callback(partial(worker, app_gen)) followed by 
reactor._wake() in the future callback.)

"Waking up the server" would be accomplished by yielding a futures 
instance (or fd magical value, etc).

> This isn't done "naturally" in Python (except perhaps with stackless or 
> greenlets). Using fds give you well-known flexible possibilities.

Yield is the natural way for one side of that, re-entering the 
generator on future completion covers the other side.  Stackless and 
greenlets are alternate ideas, but yield is built-in (and soon, so will 
futures).

> If you want to put the futures API in WSGI, think of the poor authors 
> of a WSGI server written in C who will have to write their own executor 
> and future implementation. I'm sure they have better things to do.

If they embed a Python interpreter via C, they can utilize native 
implementations of future executors, though these will obviously be 
slightly less performant than a native C implementation.  (That is, 
unless the stdlib version in 3.2 will have C backing.)

	- Alice.