Module overrides member

[patrick91]

Reproduction here: https://github.com/patrick-bug-repro/griffe-member-module

In my case I have a module with the same name as one of the function in a __init__.py file, it's a bit odd, but it works in python, unfortunately it looks like griffe is overriding it.

 .
├──  README.md
└──  strawberry
    ├──  __init__.py  # here there's a type function
    └──  type
        └──  __init__.py

output:

{
  "strawberry": {
    "kind": "module",
    "name": "strawberry",
    "labels": [],
    "members": [
      {
        "kind": "module",
        "name": "type",
        "labels": [],
        "members": [
          {
            "kind": "function",
            "name": "hello",
            "lineno": 1,
            "endlineno": 2,
            "labels": [],
            "members": [],
            "decorators": [],
            "parameters": [],
            "returns": null
          }
        ],
        "filepath": "strawberry/type/__init__.py"
      }
    ],
    "filepath": "strawberry/__init__.py"
  }
}

I'd expect to see the type function inside the members there too :)

[patrick91]

This seems to be handled here:

griffe/src/griffe/mixins.py

Lines 186 to 208 in 039a702

	if not member.is_alias:
	# When reassigning a module to an existing one,
	# try to merge them as one regular and one stubs module
	# (implicit support for .pyi modules).
	if member.is_module and not (member.is_namespace_package or member.is_namespace_subpackage):
	with suppress(AliasResolutionError, CyclicAliasError):
	if value.is_module and value.filepath != member.filepath:
	with suppress(ValueError):
	value = merge_stubs(member, value) # type: ignore[arg-type]
	for alias in member.aliases.values():
	with suppress(CyclicAliasError):
	alias.target = value
	self.members[name] = value # type: ignore[attr-defined]
	if self.is_collection: # type: ignore[attr-defined]
	value._modules_collection = self # type: ignore[union-attr]
	else:
	value.parent = self # type: ignore[assignment]
	else:
	self.members[parts[0]].set_member(parts[1:], value) # type: ignore[attr-defined]
	class ObjectAliasMixin:
	"""A mixin for methods that appear both in objects and aliases, unchanged."""

I'm not 100% familiar with the codebase, but I wonder if members should be come a dict[tuple[str, kind], MemberType] so you can have members of different kinds in the dictionary?

[pawamoy]

Thanks for the repro!

Yes, that's a known issue. It's actually a design flaw. When I designed Griffe, I centered it around the syntax to access objects: loader.modules_collection["package.module.submodule.Class.NestedClass.attribute"]. I failed to account for the fact that a module can have both a submodule and a member (imported or not) of the same name.

Instead of using (name, kind) tuples, the solution I envision is to split submodules out of the members dictionary, and update the syntax to support colons: package.module.submodule to access the submodule or the member of the same name when there's no such submodule (no ambiguity), and package.module:submodule to force access to the member and not the submodule of the same name when it exists.

As you can guess, that's quite a refactor, as there are a lot of places in the code that will need an update, so I kinda focused my energy elsewhere in the meantime 😅 But this issue is definitely in the 1.0 milestone (which exists only in my head): I will not release v1 before it is fixed.

Now if someone wants to take a stab at it, I'd gladly review a PR. The spec is simple (new submodules attribute in the Object class, support : in keys in the member mixins). But I think the implementation will have some traps and surprises 🙂

[pawamoy]

I'll close as duplicate of #124.