[New-bugs-announce] [issue40916] Proposed tweak to allow for per-task async generator semantics

Joshua Oreman report at bugs.python.org
Mon Jun 8 19:23:21 EDT 2020

New submission from Joshua Oreman <oremanj at gmail.com>:

The current async generator finalization hooks are per-thread, but sometimes you want different async generator semantics in different async tasks in the same thread. This is currently challenging to implement using the thread-level hooks. I'm proposing a small backwards-compatible change to the existing async generator hook semantics in order to better support this use case. I'm seeking feedback on the proposal and also on how "major" it would be considered. Does it need a PEP? If not, does it need to wait for 3.10 or could it maybe still make 3.9 at this point?

TL;DR: if the firstiter hook returns a callable, use that as the finalizer hook for this async generator instead of using the thread-level finalizer hook.

== Why would you want this? ==

The use case that brought me here is trio-asyncio, a library that allows asyncio and Trio tasks to coexist in the same thread. Trio is working on adding async generator finalization support at the moment, which presents problems for trio-asyncio: it wouldn't work to finalize asyncio-flavored async generators as if they were Trio-flavored, or vice versa. It's easy to tell an async generator's flavor from the firstiter hook (just look at which flavor of task is running right now), but hard to ensure that the corresponding correct finalizer hook is called (more on this below).

There are other possible uses as well. For example, one could imagine writing an async context manager that ensures all async generators firstiter'd within the context are aclose'd before exiting the context. This would be less verbose than guarding each individual use of an async generator, but still provide more deterministic cleanup behavior than leaving it up to GC.

== Why is this challenging to implement currently? ==

Both of the above use cases want to provide a certain async generator firstiter/finalizer behavior, but only within a particular task or tasks. A task-local firstiter hook is easy: just install a thread-local hook that checks if you're in a task of interest, calls your custom logic if so or calls the previous hook if not. But a task-local finalizer hook is challenging, because finalization doesn't necessarily occur in the same context where the generator was being used. The firstiter hook would need to remember which finalizer hook to use for this async generator, but where could it store that information? Using the async generator iterator object as a key in a regular dictionary will prevent it from being finalized, and as a key in a WeakKeyDictionary will remove the information before the finalizer hook can look it up (because weakrefs are broken before finalizers are called). About the only solution I've found is to store it in the generator's f_locals dict, but that's not very appealing.

== What's the proposed change? ==

My proposal is to allow the firstiter hook to return the finalizer hook that this async generator should use. If it does so, then when this async generator is finalized, it will call the returned finalizer hook instead of the thread-level one. If the firstiter hook returns None, then this async generator will use whatever the thread-level finalizer was just before firstiter was called, same as the current behavior.

== How disruptive would this be? ==

Async generator objects already have an ag_finalizer field, so this would not change the object size. It's just providing a more flexible way to determine the value of ag_finalizer, which is currently not accessible from Python.

There is a theoretical backwards compatibility concern if any existing firstiter hook returns a non-None value. There wouldn't be any reason to do so, though, and the number of different users of set_asyncgen_hooks() currently is likely extremely small. I searched all of Github and found only asyncio, curio, uvloop, async_generator, and my work-in-progress PR for Trio. All of these install either no firstiter hook or a firstiter hook that returns None.

The implementation would likely only be a handful of lines change to genobject.c.

components: asyncio
messages: 371053
nosy: Joshua Oreman, asvetlov, njs, yselivanov
priority: normal
severity: normal
status: open
title: Proposed tweak to allow for per-task async generator semantics
type: enhancement
versions: Python 3.9

Python tracker <report at bugs.python.org>

More information about the New-bugs-announce mailing list