From a quick glance at the patch, the use of PyErr_Clear here seems like it must be wrong. If something is setting an error then we almost always should exit and propagate the exception, not ignore the error and continue. And in the rare cases where PyErr_Clearis the right thing to use, it should be called immediately after the potentially-failing operation, with a comment explaining why we want to ignore errors there.

The usage of PyErr_WarnEx seems wrong here. Also _array_from_buffer_3118 seems to not follow some Python conventions. It returns -1, but doesn't set an error. With the usage of this function, errors are ignored anyway. And I'm not sure if the warnings should ever make the function fail.

The issue itself arises after _descriptor_from_pep3118_format fails and sets an error which was never handled. Would you accept it if PyErr_Clear is called just before PyBytes_FromFormat, or would you prefer rewriting this part completely?

Why does it give a warning and not just error out at all? Not following python conventions is maybe not nice, but not so bad for an internal function, but I think the problem here may be that the function should be able to signal three and not two possible results:

1. Everything is good
2. This is not a buffer at all (some error may have happened, but we should have cleared it)
3. It seems to be an invalid buffer

And it seems for 3. it actually sets a warning in a pretty nasty way, since that warning can also easily be an exception. The second clear seemes to be more likely correct to do something of the logic if PyErr_Occurred(): goto fail to be honest. I am not sure the warnings should be warnings at all and not errors though. Can it happen that we want to warn, then continue and create an object array or so?! We may need a deprecation to turn it to an unconditional error?

EDIT: Or maybe split it into two functions, one that can't error and decides whether to try or not, the other that can error and the error will always be shown (possibly after deprecation).

Looks like _descriptor_from_pep3118_format needs a fair bit of work here too...