Test Gunicorn Worker

[Kludex]

The idea here is to create a setup to test the gunicorn worker classes that we have.

I don't know how is the best way to achieve this, as gunicorn test suite didn't give me much insight on how we can test this integration without unittest.mock. Suggestions are welcome.

If you are interested on working on this, go ahead. Some clarifications:

1. Avoid integration kind of setup. I want to keep it simple.
2. Avoid unittest.mock - if inevitable, explain the reason.
3. Add test coverage up to 100% on the workers.py.

Notes

When doing this, please change those lines:

uvicorn/setup.cfg

Lines 39 to 43 in eec7d22

	[coverage:run]
	omit = venv/*
	include = uvicorn/*, tests/*
	plugins =
	coverage_conditional_plugin

To something more like: https://github.com/encode/starlette/blob/048643adc21e75b668567fc6bcdd3650b89044ea/setup.cfg#L41-L42

Important

• We're using Polar.sh so you can upvote and help fund this issue.
• We receive the funding once the issue is completed & confirmed by you.
• Thank you in advance for helping prioritize & fund our backlog.

[br3ndonland]

I can take this on. I agree we should keep it simple and focus on unit testing.

One way I've approached unit testing without mocks is by starting a Gunicorn server with the --print-config option. This does actually start the Gunicorn server, but Gunicorn will just print out the config and then shut down. You can then parse the output and assert that Gunicorn is using the appropriate settings.

I'll think about some other complementary ways to approach testing of the worker.

[br3ndonland]

Just dropping a note here to say that I'm working on this and making some slow progress.

I started off with the --print-config strategy I mentioned in #1834 (comment), but it wasn't very useful. We're not doing much modification of the Gunicorn config here, other than just setting the worker class, and I've also found that this strategy doesn't really test functionality of the server at all. For example, I can run a command like gunicorn --print-config --worker-class "this.worker.DoesNotExist" "example_package.example_module:app", and Gunicorn will start up, print out the config stating that the worker class is this.worker.DoesNotExist, and exit successfully, even though that config would clearly not work.

Instead, I'm working on some testing capabilities that will:

• Provide a running (not mocked), context-managed Gunicorn server with a client for testing requests and responses. This will probably just be an httpx.Client instance, almost like a minimal version of the Starlette TestClient.
• Provide a means of programmatically hooking into the Gunicorn process manager (the "arbiter"). This would be helpful for testing signal handling, and we have had a number of issues and PRs that relate to signal handling (such as [BUG] Received duplicated signal from Gunicorn #1116, uvicorn eats SIGINTs, does not propagate exceptions #1579 and Add SIGQUIT handler to UvicornWorker #1710). It seems like we need some means of testing how signals are handled by both the Uvicorn worker and the Gunicorn arbiter. In the future, developers working on PRs like Add SIGQUIT handler to UvicornWorker #1710 should have a simple way of testing changes to the workers.

Hope to have a PR for review soon.

[br3ndonland]

I'm working through a few tedious limitations right now. One of them relates to conditional test coverage.

As correctly suggested in the description, we need to change our coverage.py config from include = uvicorn/*, tests/* to source_pkgs = uvicorn, tests. This is because the current include syntax actually excludes uvicorn.workers from test coverage. Switching to source_pkgs will include all the modules.

The limitation is that we can't cover uvicorn.workers or the associated tests on Windows. Gunicorn doesn't test on Windows at all because it depends on various Unix utilities like fcntl.

How do we tell coverage.py, "please exclude uvicorn.workers and tests.test_workers from coverage on Windows"? Seems like there should be some obvious way to do this, but I'm not sure.

Coverage.py has an --omit flag, but I don't think this flag can handle platform-specific conditionals like we need here.

I tried setting pragmas at the top-level of the modules, but coverage-conditional-plugin doesn't support this (wemake-services/coverage-conditional-plugin#2).

Am I missing something? Or will I need to do a PR that resolves wemake-services/coverage-conditional-plugin#2 in order to unblock this?

I guess the simplest option would be to add a # pragma: py-win32 comment to every statement that's not covered. It would add a lot of comments.

[br3ndonland]

Am I missing something? Or will I need to do a PR that resolves wemake-services/coverage-conditional-plugin#2 in order to unblock this?

PR submitted (wemake-services/coverage-conditional-plugin#221). This feature will enable entire modules to be conditionally omitted from test coverage measurement.

[br3ndonland]

Coding for this is done, and we now have a nice way to omit the worker code from test coverage measurement on Windows, thanks to coverage-conditional-plugin 0.9.0.

PR coming up shortly.

[Kludex]

Thanks 🙏

[Kludex]

• The workers are moved to the Uvicorn Worker package.

I'll deprecate the workers within Uvicorn.