(fixtures): Replace fixture represenation with a class

[Glyphack]

Closes #11525

During the sprint we discussed fixing the above issue and thought maybe it's a better idea to add a better represenation to fixtures. To do this without patching the object more, this PR refactors fixtures to have a class with attributes.
The main rational behind that is:

• Now we have a type for fixtures makes it easier to check for fixture types and no need to perform cast on objects. Removed monkey patching.
• They are no longer functions that carry a class within them.
• When you decorate a function with a fixture it's not possible to call it anymore. So it makes more sense for it to not be a function or method.

Example

Previously we had:

===================================== short test summary info ======================================
FAILED tmp31nzhe4b.py::test_something - assert fixt == 42
======================================== 1 failed in 0.03s =========================================

where fixt is a pytest fixture function that is not replaced by it's value(directly used)

Now we print:

===================================== short test summary info ======================================
FAILED tmp31nzhe4b.py::test_something - assert pytest_fixture(<function fixt at *>) == 42 
======================================== 1 failed in 0.03s =========================================

[Glyphack]

Hey @The-Compiler @RonnyPfannschmidt I rebased the PR could you please take a look?

[webknjaz]

Could you assert against a more specific exception? As I understand, this is not testing that something went wrong but it's not clear what exactly. Also, it's usually a good idea to use the match= regex with this helper.

[Glyphack]

Yes will apply. This tries to document the behavior more than testing anything. Before introducing the FixtureFunctionDefinition it was possible to run inspect.getsource on fixtures but now it's not. This is just describing the behavior. I'm open to removing it as well.

[Glyphack]

I changed it to a more specific error. Do you think we can remove the test? This just tests that you cannot use inspect on fixture functions.

[Glyphack]

I think we can remove this test. I removed it.

[webknjaz]

Hey @Glyphack, could you resolve the merge conflicts / rebase this?

[Glyphack]

Hey @Glyphack, could you resolve the merge conflicts / rebase this?

Yes will do today.

[RonnyPfannschmidt]

Iove where this is going

It opens up nice follow ups

[RonnyPfannschmidt]

I believe we can drop fixture definition from that , we should pass the definitions attribute

[Glyphack]

Are you suggesting to pass the FixtureFunctionDefinition object to this function? If I understand correctly this function is expected to take any object and return the real object according to this test testing the behavior:
https://github.com/Glyphack/pytest/blob/3d7416c2dd5dcf2d60a8b30636dcc90978c64215/testing/test_compat.py#L63

[RonnyPfannschmidt]

As fixture definitions are no longer wrapping functions into other functions

The suggestion is to no longer consider them wrappers and passing the function directly

[Glyphack]

@RonnyPfannschmidt Sorry I didn't fully understand your comment.

The current unwrapping behavior here is only for two cases:

1. The pytest fixture is applied on a function so I can display this case with Fixture(f)
2. The pytest fixture is decorated so we have deco(Fixture(f))

The first case is handled by obj = obj._get_wrapped_function() line and second case is handled by obj = inspect.unwrap(obj, stop=lambda x: type(x) is FixtureFunctionDefinition).

If we drop any of them we cannot handle one of those two cases.

But I feel I am missing your point but unwrapping to me are these two lines.

[RonnyPfannschmidt]

Decorators cannot apply to a fixture definition

So that should trigger a error

It's wrong even now

[Glyphack]

I was wrong about that but the behavior is still needed. For example this test fails if we don't do unwrapping in this function:
https://github.com/Glyphack/pytest/blob/a6318c944f907bc350922c3961365e9ecd8d3e81/testing/python/integration.py#L11

The reason is that getfslineno is using this function to unwrap a decorator:
https://github.com/Glyphack/pytest/blob/a6318c944f907bc350922c3961365e9ecd8d3e81/src/_pytest/_code/code.py#L1341

I think the solution is to make a separate function for the cases when we want to perform the unwrapping? What is your suggestion?

[RonnyPfannschmidt]

What I meant is that we no longer ought to pass fixture definition to anything that unwrapped

Definition is no longer a function

Perhaps we should drop the call altogether

[Glyphack]

Thanks I realized the reason my code required this check to be here was that I was getting objects using the fixture name instead of their actual name and mistakenly I calling this on non fixture objects. After fixing that I could remove the get_real_method and simplify this function.

[Glyphack]

Thank you both for the review. I left some TODO comments I will address them in a follow up PR.

[bluetech]

@bluetech not really, I meant to apply those later (I only created a Draft PR to not lose the local commits I made).

Ah sorry for the assumption. The reason I thought it would be better to combine is that merging as-is would introduce a typing regression (fixture functions would lose their signature). But it's only temporary so not a problem.

[bluetech]

Looks good to me!

[nicoddemus]

Thanks @Glyphack!

[youtux]

FYI, this breaks pytest-factoryboy, but that's something I have to deal with.
Just wanted to let you know, in case maybe other users could have the same issue.

That being said, __pytest_wrapped__ was an internal API, so 100% agree that there is no guaranteed behavior on these.