Implemented contextvars isolation in test runner using distinct tasks for each scope

[jhominal]

This is an implementation of contextvars isolation between tests, using distinct tasks and queues for each pytest scope.

I am sure that there is a lot to review (e.g. the _TaskManager name is not very inspired ;)), but it seems that the basic functionality works. (The documentation should also be updated to document this new behaviour of the test runner)

I am going to implement the solution to issue #614 on top of the implementation in this branch, when I have a bit more time, but in the meantime I welcome any and all feedback.

[agronholm]

Nice! I'll try to find time for a review in the coming week.

[agronholm]

Looking at this, I have just one concern: do contextvars still propagate from higher scoped fixtures to lower scoped fixtures (e.g. module -> function)?

[jhominal]

Yes, higher scoped fixtures still propagate to lower scoped fixtures - that is thanks to pytest ordering guarantees, that higher scoped fixtures run before lower scoped ones. However, this propagation is implemented by making lower scoped fixtures run on a copy of the higher level context, meaning that lower scoped fixtures cannot modify the context of the higher level context. I guess I will add another test to confirm that.

However, trying to fix issue #614 is something of a pain with this system of queuing coroutines to background tasks, because suddenly I have to try and run even synchronous fixtures and tests in these background tasks and I have a hard time identifying when that is possible/desirable. I also suspect that reentrancy related to calling getfixturevalue would be difficult. Ideally I would like to run asynchronous fixtures in the "current context" without spinning a context copy (meaning, without spawning a new task), but I do not know if that is even possible.

[agronholm]

Hm, I suspected this might not be so easy. I've only taken a fairly cursory look, but I'll delve deeper when I get the chance. Maybe I can think of something then.

[jhominal]

So, after thinking about the issue during the night:

• I am beginning to think that context isolation is too much of a deviation from "regular pytest" to be something that should be introduced as a matter of course in the anyio test runner;
  • Isolation between tests can actually be handled via a decorator on the test function, same as what people would be able to do with a sync or async test;
• I think I have identified two possible, not too invasive, ways of running all fixtures and tests, including the synchronous ones, in the same context;
  • One which already promises to be extremely hacky if I can make it work at all but would allow me to avoid touching to the synchronous fixtures and tests at all;
  • One which would be slightly more invasive but could probably work by using an explicit, pytest session-scoped contextvars.Context object;

As a result of that, I am going to mark this PR as a draft, and I will submit PRs related to these two options in the coming days.

[jhominal]

Given how #618 has turned out, I am going to close this PR - I do not think anymore that this PR should be integrated in anyio.