I've been itching to play with some of the new features of python 3.4. The most anticipated feature is the new asynchronous module which should, the hope is, let us run concurrent code without going insane. This will be a code heavy post, with some very new features that the community at large hasn't really decided on any best practices yet. This will probably be a multi part series getting deeper into the intricacies of async programming in python. I haven't yet seen a good series of idioms yet, so this is mostly exploratory. This post will serve as an introduction to using async programming for the synchronously minded. Later I'll go into some of the extra APIs that let you really make the most of async programming. After that I'll try to make a third post about using these techniques to do something more practical than waiting for sleep().
import asyncio import contextlib import time @contextlib.contextmanager def timer(msg): print(msg) T = time.time() yield print(time.time() - T)
We'll do some simple context manager based benchmarks to figure out if we're doing things right.
#notacoroutine def a(A="a"): time.sleep(1) return A @asyncio.coroutine def b(B="b"): yield from asyncio.sleep(1) return B
Define some functions, a() is a normal synchronous python function that every one is used to, b() is a coroutine. By observation, they do the same thing except that b() has a "yield from" statement turning this into a generator. "yield from" is valid in python 3.3 onwards. The "asyncio" module is in the standard library in 3.4, but available in PyPI for 3.3. Here's how to use them.
def main(): with timer("Normal synchronous code"): for x in range(5): print(a()) if __name__ == '__main__': main() ## 5.005340099334717
These examples are going to call a function 5 times, and time how log it takes overall.
def main(): loop = asyncio.get_event_loop() with timer("Using asynchronous code synchronously"): for x in range(5): retval = loop.run_until_complete(b()) print(retval) if __name__ == '__main__': main() ## 5.007107496261597
With coroutines, you have to use them inside a function which is why my examples will be wrapped in a main() construct. This is because generators, and the yield function is that need something to yield to, the event loop. "loop" will be the default event loop, which has a pluggable interface so that the implementation can be changed without changing the API. We can then tell the loop to run the b() coroutine until completion and return as if this was synchronous code. If you ever need to convert a coroutine to a normal blocking function, then this is a useful construct as a last resort. It is a little bit messier than the synchronous calls, but anyone can follow this logic. There is no speed increase.
def main(): loop = asyncio.get_event_loop() with timer("Separate function call, and code running"): tasks =  for x in range(5): task = b() tasks.append(task) for task in tasks: retval = loop.run_until_complete(task) print(retval) ## 5.006737232208252
Using generators lets us split up defining functions, calling them and running them. It might be useful to do it this way, but the real benefit comes if we can run tasks in parallel.
def main(): loop = asyncio.get_event_loop() with timer("Briefer concurrency with map"): tasks = [asyncio.async(b()) for x in range(5)] print(*map(loop.run_until_complete, tasks)) ## 1.002532958984375
The key ingredient here is that we can call async() on the generator. This schedules the coroutine into the event loop and we can assume that it is running, but have no idea if it has finished. Calling run_until_complete() on each task in turn will drop out the results. While this is quick and quite readable, I think we can do a bit better.
def async_map(func, *iterables, event_loop=None): loop = event_loop or asyncio.get_event_loop() args_iter = zip(*iterables) tasks = [asyncio.async(func(*args)) for args in args_iter] for task in tasks: yield loop.run_until_complete(task) def main(): with timer("Async map, not using the event loop directly"): tasks = [b for x in range(5)] retvals = [r for r in async_map(lambda x: x(), tasks)] print(*retvals) ## 1.0016822814941406
It's useful to hide slightly tricky idioms behind convenience functions such as an asynchronous version of map(). Event loop handling boilerplate can be moved out and we're back to quick and readable code again.
def async_map(func, *iterables, event_loop=None): loop = event_loop or asyncio.get_event_loop() args_iter = zip(*iterables) if asyncio.iscoroutinefunction(func): coro = func else: @asyncio.coroutine def coro(*args): return func(*args) tasks = [asyncio.async(coro(*args)) for args in args_iter] for task in tasks: yield loop.run_until_complete(task) def main(): with timer("It still works with synchronous functions"): tasks = [a for x in range(5)] retvals = [r for r in async_map(lambda x: x(), tasks)] print(*retvals) ## 5.006326913833618
I've added a conditional so that this still works with normal functions.
def main(): apply = lambda x: x() with timer("Mixin' it up"): tasks = [b, a, b ,a, b] retvals = [r for r in async_map(apply, tasks)] print(*retvals) ## 3.004204511642456
And you can even mix synchronous and asynchronous code. Lessons learnt, we can use any paradigm we want. But there are potential speed-ups to be had in IO- bound code.