asyncio with two interpreter instances

[mbadaire]

BPO	47219
Nosy	@asvetlov, @ericsnowcurrently, @1st1

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = None
created_at = <Date 2022-04-04.19:57:08.003>
labels = ['type-bug', 'expert-asyncio']
title = 'asyncio with two interpreter instances'
updated_at = <Date 2022-04-04.20:38:00.422>
user = 'https://bugs.python.org/mbadaire'

bugs.python.org fields:

activity = <Date 2022-04-04.20:38:00.422>
actor = 'JelleZijlstra'
assignee = 'none'
closed = False
closed_date = None
closer = None
components = ['asyncio']
creation = <Date 2022-04-04.19:57:08.003>
creator = 'mbadaire'
dependencies = []
files = []
hgrepos = []
issue_num = 47219
keywords = []
message_count = 1.0
messages = ['416694']
nosy_count = 4.0
nosy_names = ['asvetlov', 'eric.snow', 'yselivanov', 'mbadaire']
pr_nums = []
priority = 'normal'
resolution = None
stage = None
status = 'open'
superseder = None
type = 'behavior'
url = 'https://bugs.python.org/issue47219'
versions = []

• PR: GH-91375: Port _asyncio static types to heap types and module state #99122

• PR: GH-91375: Purge asyncio variables from globals-to-fix list #101288

[mbadaire]

Hi,
I have an issue when using asyncio and two interpreter instances each launched and used in a seperated thread.
I am getting a asyncio loop for each thread .However asyncio is getting me the same loop because of this code in get_running_loop. Indeed when I have two interpreter, the ts_id would be the same for both my threads and therefore I will get the cached value of the first thread. cached_running_holder being static, it is the same value for all instances of interpreter.
Maybe we should check if we are in the same interpreter or same thread ,.. I am not sure how it could be fixed.

_asynciomodule.c:
get_running_loop(PyObject **loop)
{
    PyObject *rl;

    PyThreadState *ts = _PyThreadState_GET();
    uint64_t ts_id = PyThreadState_GetID(ts);
    if (ts_id == cached_running_holder_tsid && cached_running_holder != NULL) {

If it does not make sense, I have some sample code but it is not just 10 lines.

[pipiche38]

hello, is there anything we can do to get some attention ?
This is a blocking issue in our project which spawn several instances from a Python embedded framework

Thanks in advance

[ericsnowcurrently]

It looks like asyncio has not been updated to support use in multiple interpreters -- the module uses a whole bunch of global variables to store state. Until that is dealt with, asyncio cannot be used reliably with multiple interpreters. See PEP 687. @erlend-aasland

Also note that this is unlikely to be considered a bug, so no backports. Likewise it is unlikely to happen for 3.11 since the feature freeze is literally today. Thus the soonest you may see a fix is 3.12 (Fall 2023), unfortunately.

[erlend-aasland]

Yeah, as soon as PEP 687 land, we can start the process of implementing this. I believe I've got a WIP branch lying around already.

[pipiche38]

Thanks for the update, we will be patient. just a shame not to have back port, but understood the amount of work.
Now, we would be really please to test it when you want .

Potential testing platforms OS: raspian or fedora

[erlend-aasland]

I can push it to my fork next week, so you can play with it.

[kumaraditya303]

One workaround currently is to not use _asyncio speed up module but to use pure python version as that does not has this limitation.

[pipiche38]

One workaround currently is to not use _asyncio speed up module but to use pure python version as that does not has this limitation.

interesting. How to you prevent using _asyncio ?

[kumaraditya303]

Before importing asyncio add this line:

import sys
sys.modules["_asyncio"] = None

[pipiche38]

@kumaraditya303 if ayncio is imported in many module. does it enough to do it only once while importing asyncio for the first time ?

[edit] at least I did it prior the first occurence of import and this looks pretty good .

[kumaraditya303]

Yes it is only required once when asyncio is imported for the first time.

[pipiche38]

@kumaraditya303 big big big thanks. This is an awesome workaround .
You have unlocked our main issue preventing us to work with multi-instance.

[gvanrossum]

Yeah, as soon as PEP 687 land, we can start the process of implementing this. I believe I've got a WIP branch lying around already.

Since PEP 687 is accepted, how about it? Seems a good idea to get this out of the way.

[gvanrossum]

Also, looking at the code a bit more, it looks like PyThreadState_GetID() returns tstate->id. I wonder if, regardless of all the other things we should do, maybe we could make that truly unique, even across interpreters? It looks like it just increments interp->threads.next_unique_id and uses that. It looks like a simple counter. Maybe that counter could be moved to the runtime state, so it's truly unique? @ericsnowcurrently

[erlend-aasland]

Since PEP 687 is accepted, how about it? Seems a good idea to get this out of the way.

Thanks for the ping. I'll try to find time for this in the coming week.

[ericsnowcurrently]

Maybe that counter could be moved to the runtime state, so it's truly unique?

That would work. We would use the existing _PyRuntimeState.interpreters.mutex to protect "next_unique_id".

An alternative would be to rely on tstate->thread_id instead of tstate->id, since the former is globally unique.

That said, once _asynciomodule.c has been updated, the cached_running_holder_tsid value would be located on the module state, which is always interpreter-specific. So making tstate->id globally unique (or using tstate->thread_id) might not provide much value.

[gvanrossum]

I feel that the promise in the comment ("unique ID") is violated by the current code. Either the comment should be updated to something like "unique per interpreter" or (IMO better) we should follow the spirit of the comment. Your proposed implementation sounds fine.

[kumaraditya303]

Thanks for the ping. I'll try to find time for this in the coming week.

If you don't mind, please add me as a reviewer.

[gvanrossum]

@erlend-aasland Did you ever get to this? It reared its head again.

[erlend-aasland]

@erlend-aasland Did you ever get to this? It reared its head again.

Sorry; for various reasons I haven't been able to devote much time to CPython the last weeks. I picked up my work on this last week, though. Currently, the diff stat is ±500 lines, but I expect it to double or tiple by the time I'm done (there'll be some clinic changes that will explode the diff). I also got a ref. leak after converting TaskType to heap type. It's been time-consuming to debug it on my machine; a ./python.exe -m test -R : test_asyncio takes hours (last run took close to 4 hours!).

I'll create a draft PR so anyone interested can join in and help with debugging and progress.

[kumaraditya303]

I'll create a draft PR so anyone interested can join in and help with debugging and progress.

Looking forward to the draft PR. Thanks

It's been time-consuming to debug it on my machine; a ./python.exe -m test -R : test_asyncio takes hours (last run took close to 4 hours!).

Debugging on Linux is much faster so I will be able to debug it faster.

[erlend-aasland]

FTR, here's my WIP branch: https://github.com/erlend-aasland/cpython/tree/isolate-asyncio
Feel free to take over, Kumar!

[kumaraditya303]

Draft PR here #99122

[ericsnowcurrently]

Thanks for working on this, @kumaraditya303!

[erlend-aasland]

Reopening: the future iter object freelist is still a static global, and we forgot to clean up Tools/c-analyzer/cpython/globals-to-fix.tsv.

[kumaraditya303]

Reopening: the future iter object freelist is still a static global, and we forgot to clean up Tools/c-analyzer/cpython/globals-to-fix.tsv.

I left that freelist intentionally as it is not an issue until we have per interpreter GIL. @markshannon has ideas for a global better freelist so hopefully we should be able to remove this freelist entirely.

[erlend-aasland]

I left that freelist intentionally as it is not an issue until we have per interpreter GIL. @markshannon has ideas for a global better freelist so hopefully we should be able to remove this freelist entirely.

I see, thanks for the heads-up.

[gvanrossum]

I left that freelist intentionally as it is not an issue until we have per interpreter GIL. @markshannon has ideas for a global better freelist so hopefully we should be able to remove this freelist entirely.

Hm. I wouldn't necessarily count on Mark's single freelist idea to be implemented before we get a per-subinterpreter GIL (esp. since the work on mimalloc seems stalled, alas -- Christian Heimes seems to be occupied by other things).

So I think it would behoove us to move fi_freelist and fi_freelist_len to the "per runtime" structure? Or what am I missing?