filterwarnings causes warnings.warn to raise TypeError on non-string messages

[zanieb]

Bug report

When using filterwarnings, a TypeError is raised when a non-string value is passed to warnings.warn which differs from the default behavior of warnings.warn.

import warnings

# Does not raise a type error
warnings.warn(1)

# Raises a type error
with warnings.catch_warnings():
    warnings.filterwarnings("ignore", "test")
    warnings.warn(1)

The traceback does not demonstrate that this is a detail of warning filtering:

/Users/mz/eng/src/madkinsz/cpython-103577/example.py:4: UserWarning: 1
  warnings.warn(1)
Traceback (most recent call last):
  File "/Users/mz/eng/src/madkinsz/cpython-103577/example.py", line 9, in <module>
    warnings.warn(1)
TypeError: expected string or bytes-like object

Filtering warnings should not change the accepted types of warnings.warn.

Reproduction example at https://github.com/madkinsz/cpython-103577
Originally raised at pytest-dev/pytest#10865

Your environment

Reproduced Ubuntu with Python 3.7-3.12: https://github.com/madkinsz/cpython-103577/actions/runs/4714383423/jobs/8360711980

Also reproduced manually on macOS with Python 3.10/3.11

[zanieb]

I'm happy to help contribute a fix if ya'll agree this behavior isn't ideal :)

[sobolevn]

Are there any real-life use-cases where using warnings.warn with non-str messages make sense?
I've only seen str and Warning subclasses.

Docs also say the same: https://docs.python.org/3/library/warnings.html#warnings.warn
Typeshed also annotates message as either str or Warning: https://github.com/python/typeshed/blob/7dcec5b3db887f518114486b3e4783dde244186d/stdlib/_warnings.pyi#L7-L10

In you case you can use:

>>> with warnings.catch_warnings():
...     warnings.filterwarnings("ignore", "test")
...     warnings.warn(Warning(1))
... 
<stdin>:3: Warning: 1

I don't think that allowing more types in is a good thing.

[zanieb]

Hey @sobolevn — I agree that it doesn't make sense to allow more types but it is very surprising for the behavior to change when you enable filtering. This means that you can have a library that works totally fine until warning filtering is used then suddenly a type error will be thrown.

In practice, I encountered this working on prefect where we display a warning based on an exception:

https://github.com/PrefectHQ/prefect/blob/b6d0433a15b8172c0f897a6eac2b4d747350fd1f/src/prefect/logging/handlers.py#L123

Before we added the str cast, the code worked fine in some environments but failed in others.

[sobolevn]

This means that you can have a library that works totally fine until warning filtering is used then suddenly a type error will be thrown.

Yes, this is an eventual side-effect of dynamic typing. It all works untill it does not :)

Typechecker should catch errors like this. I don't think that anything should be done in this case from CPython's side.

Do you agree? Maybe you have ideas of better docs wordings?

[zanieb]

I hesitantly agree :) it's a sneaky problem since the error message / traceback does not include any context about filtering. Perhaps:

1. Add a warning to the documentation of filterwarnings about this case
2. Improve the traceback i.e. by moving the filtering to a separate utility function so there's a frame indicating that it's related to filtering
3. Throw a warning when warnings.warn is called with an invalid type (backwards compatible and helpful but not very Pythonic)

[Zac-HD]

IMO we should change from "errors if you called filterwarnings()" to "always errors".

[zanieb]

@Zac-HD I'm happy with that too — seems like a breaking change but a reasonable one