[Python-ideas] async/await and synchronous code (and PEP492 ?)
koos.zevenhoven at aalto.fi
Tue May 5 15:55:56 CEST 2015
I am excited about seeing what's going on with asyncio and PEP492 etc. I
really like that Python is becoming more suitable for the increasing
amount of async code and that the distinction between async functions
and generators is increasing.
In addition, however, I would also like to see the async functions and
methods come even closer to regular functions and methods. This is
something that is keeping me from using asyncio at the moment even if I
would like to. Below I'll try to explain what and why, and a little bit
of how. If it is not clear, please ask :)
One of the best things about asyncio and coroutines/async functions is
that you can write asynchronous code as if it were synchronous, the
difference in many places being just the use of "await" ("yield from")
when calling something that may end up doing IO (somewhere down the
function call chain) and that the code is run from an event loop.
When writing a package that does IO, you have the option to make it
either synchronous or asynchronous. Regardless of the choice, the code
will look roughly the same. But what if you want to be able to do both?
Should you maintain two versions, one with "async" and "await"
everywhere and one without?
Besides the keywords "async" and "await", async code of course differs
from synchronous code by the functions/coroutines that are used for IO
at the end of the function call chain. Here, I mean the end (close to)
where the "yield" expressions are hidden in the async versions. At the
other end of the calling chain, async code needs the event loop and
associated framework (almost always asyncio?) which hides all the async
scheduling fanciness etc. I'm not sure about the terminology, but I will
use "L end" and "Y end" to refer to the two ends here. (L for event
Loop; Y for Yield)
The Y and L ends need to be compatible with each other for the code to
work. While asyncio and the standard library might provide both ends in
many cases, there can also be situations where a package would want to
work with different combinations of L and Y end, or completely without
an event loop, i.e. synchronously.
In a very simple example, one might want to wrap different
implementations of sleep() in a function that would pick the right one
depending on the context. Perhaps something like this:
async def any_sleep(seconds):
if __async__.framework is None:
elif __async__.framework is asyncio:
raise RuntimeError("Was called with an unsupported async
[You could of course replace sleep() with socket IO or whatever, but
sleep is nice and simple. Also, a larger library would probably have a
whole chain of async functions and methods before calling something like
But if await is only allowed inside "async def", then how can
any_sleep() be conveniently run in non-async code? Also, there is
nothing like __async__.framework. Below, I describe what I think a
potential solution might look like.
This is simplified version; for instance, as "awaitables", I consider
only async function objects here. I describe the idea in three parts:
Add a keyword argument "async_framework" (or whatever) to next(...) with
a default value of None. When an async framework, typically asyncio,
starts an async function object (coroutine) with a call to next(...), it
would do something like next(coro, async_framework = asyncio). Here,
asyncio could of course be replaced with any object that identifies the
framework. This information would then be somehow attached to the async
(2) __async__.framework or something similar:
Add something like __async__ that has an attribute such as .framework
that allows the code inside the async function to access the information
passed to next(...) by the framework (L end) using the keyword argument
of next [see (1)].
(3) Generalized "await":
[When the world is ready:] Allow using "await" anywhere, not just within
async functions. Inside async functions, the behavior of "await" would
be the same as in PEP492, with the addition that it would somehow
propagate the __async__.framework value to the awaited coroutine.
Outside async functions, "await" would do roughly the same as this function:
next(async_func_obj) # same as next(async_func_obj,
async_framework = None)
except StopIteration as si:
raise RuntimeError("The function does not support synchronous
(This function would, of course, work in Python 3.4, but it would be
mostly useless because the async functions would not know that they are
being called in a 'synchronous program'. IIUC, this *function* would be
valid even with PEP492, but having this as a function would be ugly in
the long run.)
Some random thoughts:
With this addition to Python, one could write libraries that work both
async and non-async. When await is not inside async def, one would
expect it to potentially do blocking IO, just like an await inside async
def would suggest that there is a yield/suspend somewhere in there.
For testing, I tried to see if there is a reasonable way to make a hack
with __async__.framework that could be set by next(), but did not find
an obvious way. For instance, coro.gi_frame.f_locals is read-only, I
An alternative to this approach could be that await would implicitly
start a temporary event loop for running the coroutine, but how would it
know which event loop? This might also have a huge performance overhead.
Relation to PEP492:
This of course still needs more thinking, but I wanted to post it here
now in case there is desire to prepare for something like this already
in PEP492. It is not completely clear if/how this would need to affect
PEP492, but some things come to mind. For example, this could
potentially remove the need for __aenter__, __aiter__, etc. or even
"async for" and "async with". If __aenter__ is defined as "async def",
then a with statement would do an "await" on it, and the context manager
would have __async__.framework (or whatever it would be called)
available, for determining what behavior is appropriate.
Was this clear enough to understand which problem(s) this would be
solving and how? I'd be happy to hear about any thoughts on this :).
More information about the Python-ideas