inspect.getframeinfo forces all lazy modules to be loaded

[SpecLad]

Bug report

Bug description:

Create two files as follows:

main.py:

import importlib.util
import inspect
import sys

# https://docs.python.org/3/library/importlib.html#implementing-lazy-imports
def lazy_import(name):
    spec = importlib.util.find_spec(name)
    loader = importlib.util.LazyLoader(spec.loader)
    spec.loader = loader
    module = importlib.util.module_from_spec(spec)
    sys.modules[name] = module
    loader.exec_module(module)
    return module
    
lazy_import('m')

f = inspect.currentframe()
inspect.getframeinfo(f)

m.py:

print("I'm loaded!")

Now run:

$ python main.py 
I'm loaded!

This behavior doesn't make much sense, since f isn't even inside m.

CPython versions tested on:

3.13

Operating systems tested on:

Linux

[ramicats]

The inspect module's getframeinfo(frame, context) calls getmodule(frame, file) and that creates a module / filename cache by iterating through sys.modules & accessing their attributes. Since lazy_import(name) adds the lazy module to sys.modules, sys.modules[name] = module - the lazy module gets initialized when accessing the attribute (see "__getattribute__" in LazyModule for more info on accessing attrs).

Need someone with more expertise in stdlibs to comment on if there's a possible workaround but the mutation checks for the lazy module attributes is the reason the module is being loaded.

[ramicats]

On hindsight, it seems like the inspect module only accesses the 'file' attribute of sys.modules. If it's ok to have an early return on certain immutable attributes that are derived from ModuleSpec metadata in LazyModule (like file in this instance), then we could evade the issue. Looking forward to thoughts from others, I have a working fix if this gets the ok.

[xuantengh]

Just my two cents: I think it's the expected behavior? The lazy module should be triggered to load at the time it is firstly accessed, wherever this happens, in user code or in first-party libraries.

[ramicats]

Yep, I agree accessing a lazy module's attributes triggering a load is expected. I still think it's not ideal that an introspection tool has a side effect of executing lazy modules. The mutated values tracking only checks for non-metadata attributes i.e. '__dict__', therefore I think a change might be warranted (?) to early return metadata attributes. @picnixz

[picnixz]

I agree with xuantengh. While it's not ideal as a side-effect, lazy modules are really meant to be lazy until they are needed. Here, we "somehow" need it as a side effect but making special cases isn't enough IMO. inspect is known to create side-effects, either by evaluating signatures, or other stuff like that. With PEP-810 being accepted, I'm not sure that it will be the case anymore, so I would first close the issue now and possibly re-open it for 3.15 and later if this behavior is still inacceptable.

OTOH, if we were to pre-load some attributes on the lazy module, maybe.. it could work?

cc @serhiy-storchaka