Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Support the deprecation of asyncore in python 3.12 #260

Merged
merged 5 commits into from
Nov 27, 2023

Conversation

fruch
Copy link

@fruch fruch commented Oct 11, 2023

since we need to deprecate asyncore which was the default event loop manager, we need to extend the testing of some of the other so we can select a new default

@fruch fruch requested a review from Lorak-mmk October 11, 2023 12:30
@fruch
Copy link
Author

fruch commented Oct 11, 2023

@Lorak-mmk
started tweaking the CI a bit so we could cover other event loops
and test the python 3.12 support (as we'll start addressing it)

@fruch fruch force-pushed the make_asyncore_unused branch 5 times, most recently from 32919c0 to a9bad78 Compare October 11, 2023 15:47
@fruch
Copy link
Author

fruch commented Oct 12, 2023

@Lorak-mmk

FYI, the situation with asyncio on python-3.8 is o.k.

One extra failing test

We'll need to test more python versions

@fruch
Copy link
Author

fruch commented Oct 12, 2023

I'm setting a goal to this PR (so won't spread too much)

CI working for 3.8 - asyncore/libev/asyncio

3.12 - only libev

Other python versions aren't mandatory
for now

@Lorak-mmk
Copy link

@Lorak-mmk

FYI, the situation with asyncio on python-3.8 is o.k.

One extra failing test

We'll need to test more python versions

Wow, I'm very surprised that it works

@fruch fruch force-pushed the make_asyncore_unused branch 12 times, most recently from fc7e475 to ebd44d3 Compare October 17, 2023 21:16
@fruch
Copy link
Author

fruch commented Oct 17, 2023

I'm setting a goal to this PR (so won't spread too much)

CI working for 3.8 - asyncore/libev/asyncio

3.12 - only libev

Other python versions aren't mandatory for now

so at the end I've did much more in this PR
and it seems to have the full support for asyncio working and passing integration tests
and python3.12 to be workable, i.e. we can use asyncio and libev, which sound good enough for me.

followup PRs would be

  • wheel building for python 3.12
  • testing the extra loop managers, they seem to be having issue, and we'll need to pick on which version to test them regularly

@fruch fruch force-pushed the make_asyncore_unused branch from ebd44d3 to 85a7211 Compare October 17, 2023 21:21
@fruch fruch requested a review from Lorak-mmk October 17, 2023 21:25
@fruch
Copy link
Author

fruch commented Oct 17, 2023

@Lorak-mmk

from my POV we can merge this one, once I get a green CI

but I still want you to take a look at the actual commit (not the CI ones, they are quite straightforward)

@fruch fruch force-pushed the make_asyncore_unused branch from b93369c to 48399d6 Compare October 18, 2023 21:20
@fruch
Copy link
Author

fruch commented Oct 18, 2023

#264 seems to be flaky, and not sure exactly why...
I'lll leave it for now,

@Lorak-mmk this one is ready from my POV

Copy link

@Lorak-mmk Lorak-mmk left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Requested some changes. We should probably offer this PR upstream after we merge it,

cassandra/cluster.py Outdated Show resolved Hide resolved

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This file has a ConnectionTests class which is then subclassed (AsyncoreConnectionTests and LibevConnectionTests). It would be good if a part of this PR was to add subclasses for other reactors. If that's too much work, a TODO will suffice.
I also wonder if a subclass for each reactor is the best approach - it causes us to run all of them always, no matter what is the value of EVENT_LOOP_MANAGER, maybe it should be conditional?

One more thing, in case of import error, relevant var is set to None - I don't get why it doesn't cause tests to fail - I don't see any None checks, and ConnectionTests.setUp calls self.klass.initialize_reactor() which should fail - so there is something I don't understand in this code...

# need to run them with specific configuration like `gevent.monkey.patch_all()` or under async functions
unsupported_connection_classes = [GeventConnection, AsyncioConnection, EventletConnection]
# unsupported_connection_classes = [GeventConnection, AsyncioConnection, EventletConnection]

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't immediately see why wouldn't AsyncioConnection work here. Did you try to run this test with it?
Also, is commenting out this line and 3 imports necessary?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Asyncio wasn't working at all when those tests were written, so I didn't introduce it now

As for the comment, I rather not import the unused classes, it was for documenting only, and import of some of them now complicates things

.github/workflows/integration-tests.yml Outdated Show resolved Hide resolved
Comment on lines 92 to 97
loop_args = dict()
if sys.version_info[0] == 3 and sys.version_info[1] < 10:
loop_args['loop'] = self._loop
self._write_queue = asyncio.Queue(**loop_args)
self._write_queue_lock = asyncio.Lock(**loop_args)

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I really don't like such conditionals...
This could be solved differently - before 3.10, when loop argument is not provided, current loop is used (fetched during __init__). So we could put a task that creates those objects on a loop and get the result, something like this (it can most likely be a bit prettier, it's just a PoC):

async def asyncify(f, *args, **kwargs):
    return f(*args, **kwargs)
self._write_queue = asyncio.run_coroutine_threadsafe(asyncify(asyncio.Queue)(), self._loop).result()
self.._write_queue_lock = asyncio.run_coroutine_threadsafe(asyncify(asyncio.Lock)(), self._loop).result()

However I'm not really sure it's better than (if python < 3.10) version...
If we stay with if approach, let's add a TODO to remove it after dropping support for 3.9

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

seems like in upstream they removed completely
datastax@e9136f4
which is quite weird, since it's break python 3.8

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sorry, missed this one earlier. The params in question were removed in PYTHON-1313. There's more detail in the ticket but in short asyncio will default to the event loop for the current thread since 3.7. And since the asyncio reactor explicitly sets the created loop for the current thread via set_event_loop() immediately after it's creation there didn't appear to be any obvious need to pass it around.

The relevant tests pass on Python 3.8 on our Jenkins instance, but if I've missed something there I'm certainly open to changes.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

again, from my own local tests, and tests running on github actions,
python3.8 break totally when those parameters are not passed to the locks

cassandra/io/asyncioreactor.py Outdated Show resolved Hide resolved
cassandra/io/asyncioreactor.py Outdated Show resolved Hide resolved

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

relevant_warnings can probably be made empty if we apply my suggestion instead of version-conditional code.

from cassandra.io.asyncorereactor import AsyncoreConnection as DefaultConnection # NOQA
from cassandra.io.asyncioreactor import AsyncioConnection as DefaultConnection # NOQA

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We could also think about removing asyncore in a follow up PR.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

There's no rush on that,

And we can sync with upstream about when todo it, i.e. not to break current users that might have hardcore it

@absurdfarce
Copy link

Hey all, just a quick note on upstream plans related to what you guys are working on for this ticket.

We're gearing up for a 3.29.0 release which aims to follow our policy of supporting all non-EOL releases at time of release. That means we're aiming to support 3.8 through 3.12 with 3.29.0. The removal of asyncore in 3.12 is being pursued in PYTHON-1366. I had originally intended to make the asyncio reactor the default (replacing asyncore) for 3.29.0 but I'm not convinced the asyncio reactor is stable enough to support that yet. So for 3.29.0 I'm settling on messaging telling users to use one of the third-party event loops for now; asyncio will become the default once we can make it more stable.

To that end: if you guys are inclined to contribute any fixes to the asyncio reactor upstream we'd be more than happy to take a look at them. It's probably worthwhile to separate that work from any work around 3.12 compat, though, in order to avoid any confusion of goals.

Comments welcome. :)

@fruch
Copy link
Author

fruch commented Nov 7, 2023

Hey all, just a quick note on upstream plans related to what you guys are working on for this ticket.

We're gearing up for a 3.29.0 release which aims to follow our policy of supporting all non-EOL releases at time of release. That means we're aiming to support 3.8 through 3.12 with 3.29.0. The removal of asyncore in 3.12 is being pursued in PYTHON-1366. I had originally intended to make the asyncio reactor the default (replacing asyncore) for 3.29.0 but I'm not convinced the asyncio reactor is stable enough to support that yet. So for 3.29.0 I'm settling on messaging telling users to use one of the third-party event loops for now; asyncio will become the default once we can make it more stable.

I've managed at least to get the asyncio passing all the integration suite

Would love to hear what are the other thing you think would be need to title it ready for production (maybe we help in some parts)

To that end: if you guys are inclined to contribute any fixes to the asyncio reactor upstream we'd be more than happy to take a look at them. It's probably worthwhile to separate that work from any work around 3.12 compat, though, in order to avoid any confusion of goals.

Yeah I know I was trying to do multiple things here, but I'll can separate the asyncio only change into an upstream PR.

Comments welcome. :)

@fruch
Copy link
Author

fruch commented Nov 7, 2023

Requested some changes. We should probably offer this PR upstream after we merge it,

Yes, we probably would need to break it a bit

@fruch fruch changed the title CI: run integration tests on multiple EVENT_LOOP_MANAGER and python versions Support the depreciation of asyncore in python 3.12 Nov 9, 2023
@fruch fruch force-pushed the make_asyncore_unused branch from 48399d6 to 3a5609d Compare November 9, 2023 10:36
@mykaul mykaul changed the title Support the depreciation of asyncore in python 3.12 Support the deprecation of asyncore in python 3.12 Nov 9, 2023
@fruch fruch force-pushed the make_asyncore_unused branch 3 times, most recently from 926ae2d to 8bf2590 Compare November 9, 2023 22:31
@absurdfarce
Copy link

absurdfarce commented Nov 13, 2023

Hey @fruch , appreciate the follow-up!

An update to my previous comment: I've now merged the changes in PYTHON-1366 into the code base. As mentioned above these changes throw exceptions when trying to use asyncore with Python 3.12. These warnings coupled with documentation and messaging around the release should make clear why asyncore isn't supported on Python 3.12 and what the user can do about it (i.e. use another event loop).

I decided it was easier to take an incremental approach like this (not failing completely when trying to use asyncore with Python 3.12 while also providing clear instructions) in the hope that it would be less intrusive for our users. Which brings me to your other question: what do we need to see to make asyncio the default?

I'm impressed with the work you guys have done to get asyncio passing all the integration tests and would love to incorporate those changes into upstream as well. Other than that I'd like to see how this reactor behaves over time. We typically run "duration tests" with each release that run for several days in order to identify any regressions that only appear over time. I'd like to see an asyncio reactor pass such a test. I'd also like to see some more load testing, just to make sure it won't have any issues if we stress it a bit.

I'm not at all opposed to making asyncio the default reactor for future driver releases; with the removal of asyncore in 3.12 that seems like the correct way to go. I just want to do it in a step-wise fashion so that we don't drop a ton of change on our users within a single release.

Thanks again!

@fruch
Copy link
Author

fruch commented Nov 13, 2023

Hey @fruch , appreciate the follow-up!

An update to my previous comment: I've now merged the changes in PYTHON-1366 into the code base. As mentioned above these changes throw exceptions when trying to use asyncore with Python 3.12. These warnings coupled with documentation and messaging around the release should make clear why asyncore isn't supported on Python 3.12 and what the user can do about it (i.e. use another event loop).

I did thought we might be able to default to asyncio only once asyncore isn't available,
cause we are using that in cqlsh that we bundle with scylla (and I think cassandra does exactly the same)
it would break, since it's using the pure python implementation without any compilation
i.e. that default ux of be able to run with any changes, or selecting between the different options

I decided it was easier to take an incremental approach like this (not failing completely when trying to use asyncore with Python 3.12 while also providing clear instructions) in the hope that it would be less intrusive for our users. Which brings me to your other question: what do we need to see to make asyncio the default?

I'm impressed with the work you guys have done to get asyncio passing all the integration tests and would love to incorporate those changes into upstream as well. Other than that I'd like to see how this reactor behaves over time. We typically run "duration tests" with each release that run for several days in order to identify any regressions that only appear over time. I'd like to see an asyncio reactor pass such a test. I'd also like to see some more load testing, just to make sure it won't have any issues if we stress it a bit.

Yeah I thought that would helpful to grant trust, we didn't touch those test in this repo yet
one other idea I has was to use our dtest repo, and run it with asyncio, and flush out possible issues.

I'm not at all opposed to making asyncio the default reactor for future driver releases; with the removal of asyncore in 3.12 that seems like the correct way to go. I just want to do it in a step-wise fashion so that we don't drop a ton of change on our users within a single release.

that's great, that things are moving fast, having python 3.12 support is very important to us
our users (i.e. developers) are very keen on running latest versions of their linux distros
and python 3.12 is coming in the next version of a lot of the distros (and I think already know that)

Thanks again!

to conclude, I think we won't be merging those change, and we'll wait for the upstream release (which I hope is just around the corner)

as for the asyncio related changes, please see: #271

@fruch fruch requested a review from Lorak-mmk November 19, 2023 16:33
running on python 3.12, we get this error, we should ignore
it until eventlet is fixed

```
ImportError while loading conftest '/home/runner/work/python-driver/python-driver/tests/integration/conftest.py'.
tests/integration/__init__.py:16: in <module>
    from cassandra.cluster import Cluster
cassandra/cluster.py:103: in init cassandra.cluster
    from cassandra.io.eventletreactor import EventletConnection
cassandra/io/eventletreactor.py:18: in <module>
    import eventlet
.test-venv/lib/python3.12/site-packages/eventlet/__init__.py:17: in <module>
    from eventlet import convenience
.test-venv/lib/python3.12/site-packages/eventlet/convenience.py:7: in <module>
    from eventlet.green import socket
.test-venv/lib/python3.12/site-packages/eventlet/green/socket.py:21: in <module>
    from eventlet.support import greendns
.test-venv/lib/python3.12/site-packages/eventlet/support/greendns.py:45: in <module>
    from eventlet.green import ssl
.test-venv/lib/python3.12/site-packages/eventlet/green/ssl.py:25: in <module>
    _original_wrap_socket = __ssl.wrap_socket
E   AttributeError: module 'ssl' has no attribute 'wrap_socket'
```

Ref: eventlet/eventlet#812
this isn't being used anyhow, and breaking support
for python 3.12
since asyncore isn't available in python 3.12,
we should be gracfully handle it, and enable any other event
loop implementions to work
since python 3.12 is deprecating asyncore, we should make
asyncio the default fallback event loop when asyncore isn't available

asyncio now that it's fixed and we verified it's working (passing the integration suite)
in multiple python versions we support (from 3.8 - 3.12)
@fruch
Copy link
Author

fruch commented Nov 20, 2023

@Lorak-mmk
Can you take a look at this one, this would enable us to release python3.12 wheel

And then I think we should release one last minor release, before syncing with upstream 3.28.0, or 3.29.0 once it would be available

@Lorak-mmk
Copy link

I left 2 questions in the comments, apart from that looks good to me. However I'd prefer to not release until #271 (comment) is addressed

@fruch fruch merged commit 1cc6ccc into scylladb:master Nov 27, 2023
20 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants