[Python-Dev] PEP 492 vs. PEP 3152, new round
Guido van Rossum
guido at python.org
Fri Apr 24 19:03:48 CEST 2015
I've tried to catch up with the previous threads. A summary of issues
1. precise syntax of `async def` (or do we need it at all)
2. do we need `async for` and `async with` (and how to spell them)
3. syntactic priority of `await`
4. `cocall` vs. `await`
5. do we really need `__aiter__` and friends
7. compatibility with asyncio and existing users of it
(I've added a few myself.)
I'll try to take them one by one.
*1. precise syntax of `async def`*
Of all the places to put `async` I still like *before* the `def` the best.
I often do "imprecise search" for e.g. /def foo/ and would be unhappy if
this didn't find async defs. Putting it towards the end (`def foo async()`
or `def foo() async`) makes it easier to miss. A decorator makes it hard to
make the syntactic distinctions required to reject `await` outside an async
function. So I still prefer *`async def`*.
*2. do we need `async for` and `async with`*
Yes we do. Most of you are too young to remember, but once upon a time you
couldn't loop over the lines of a file with a `for` loop like you do now.
The amount of code that was devoted to efficiently iterate over files was
tremendous. We're back in that stone age with the asyncio `StreamReader`
class -- it supports `read()`, `readline()` and so on, but you can't use it
with `for`, so you have to write a `while True` loop. `asyncio for` makes
it possible to add a simple `__anext__` to the `StreamReader` class, as
async def __anext__(self):
line = await self.readline()
if not line:
A similar argument can be made for `async with`; the transaction commit is
pretty convincing, but it also helps to be able to wait e.g. for a
transport to drain upon closing a write stream. As for how to spell these,
I think having `async` at the front makes it most clear that this is a
(Though maybe we should consider `await for` and `await with`? That would
have the advantage of making it easy to scan for all suspension points by
searching for /await/. But being a verb it doesn't read very well.)
*3. syntactic priority of `await`*
Yury, could you tweak the syntax for `await` so that we can write the most
common usages without parentheses? In particular I'd like to be able to
return await foo()
with await foo() as bar: ...
foo(await bar(), await bletch())
(I don't care about `await foo() + await bar()` but it would be okay.)
I think this is reasonable with some tweaks of the grammar (similar to what
Greg did for cocall, but without requiring call syntax at the end).
*4. `cocall` vs. `await`*
Python evolves. We couldn't have PEP 380 (`yield from`) without prior
experience with using generators as coroutines (PEP 342), which in turn
required basic generators (PEP 255), and those were a natural evolution of
Python's earlier `for` loop.
We couldn't PEP 3156 (asyncio) without PEP 380 and all that came before.
The asyncio library is getting plenty of adoption and it has the concept of
separating the *getting* of a future from *waiting* for it. IIUC this
is also how `await` works in C# (it just requires something with an async
type). This has enabled a variety of operations that take futures and
produce more futures.
 I write `future` with a lowercase 'f' to include concepts like
coroutine generator objects.
*I just can't get used to this aspect of PEP 3152, so I'm rejecting it.*
Sorry Greg, but that's the end. We must see `await` as a refinement of
`yield from`, not as an alternative. (Yury: PEP 492 is not accepted yet,
but you're getting closer.)
One more thing: this separation is "Pythonic" in the sense that it's
similar to the way *getting* a callable object is a separate act from
*calling* it. While this is a cause for newbie bugs (forgetting to call an
argument-less function) it has also enabled the concept of "callable" as
more general and more powerful in Python: any time you need to pass a
callable, you can pass e.g. a bound method or a class or something you got
from `functools.partial`, and that's a useful thing (other languages
require you to introduce something like a lambda in such cases, which can
be painful if the thing you wrap has a complex signature -- or they don't
support function parameters at all, like Java).
I know that Greg defends it by explaining that `cocal f(args)` is not a
`cocall` operator applied to `f(args)`, it is the *single* operator `cocall
...(args)` applied to `f`. But this is too subtle, and it just doesn't jive
with the long tradition of using `yield from f` where f is some previously
*5. do we really need `__aiter__` and friends*
There's a lot of added complexity, but I think it's worth it. I don't think
we need to make the names longer, the 'a' prefix is fine for these methods.
I think it's all in the protocols: regular `with` uses `__enter__` and
`__exit__`; `async with` uses `__aenter__` and `__aexit__` (which must
Ditto for `__aiter__` and `__anext__`. I guess this means that the async
equivalent to obtaining an iterator through `it = iter(xs)` followed by
`for x over it` will have to look like `ait = await aiter(xs)` followed by
`for x over ait`, where an iterator is required to have an `__aiter__`
method that's an async function and returns self immediately. But what if
you left out the `await` from the first call? I.e. can this work?
ait = aiter(xs)
async for x in ait:
The question here is whether the object returned by aiter(xs) has an
`__aiter__` method. Since it was intended to be the target of `await`, it
has an `__await__` method. But that itself is mostly an alias for
`__iter__`, not `__aiter__`. I guess it can be made to work, the object
just has to implement a bunch of different protocols.
I'm not sure about this. The motivation given in the PEP seems to focus on
the need for `__anext__` to be async. But is this really the right pattern?
What if we required `ait.__anext__()` to return a future, which can either
raise good old `StopIteration` or return the next value from the iteration
when awaited? I'm wondering if there are a few alternatives to be explored
around the async iterator protocol still.
*7. compatibility with asyncio and existing users of it*
This is just something I want to stress. On the one hand it should be
really simple to take code written for pre-3.5 asyncio and rewrite it to
PEP 492 -- simple change `@asyncio.coroutine` to `async def` and change
`yield from` to `await`. (Everything else is optional; existing patterns
for loops and context managers should continue to work unchanged, even if
in some cases you may be able to simplify the code by using `async for` and
But it's also important that *even if you don't switch* (i.e. if you want
to keep your code compatible with pre-3.5 asyncio) you can still use the
PEP 492 version of asyncio -- i.e. the asyncio library that comes with 3.5
must seamlessly support mixing code that uses `await` and code that uses
`yield from`. And this should go both ways -- if you have some code that
uses PEP 492 and some code that uses pre-3.5 asyncio, they should be able
to pass their coroutines to each other and wait for each other's coroutines.
That's all I have for now. Enjoy!
--Guido van Rossum (python.org/~guido)
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Python-Dev