On 8/6/2026 6:26 AM, Richard Purdie via lists.openembedded.org wrote:
> On Tue, 2026-08-04 at 22:04 +0000, Alejandro Hernandez Samaniego via lists.openembedded.org wrote:
>> python distutils/sysconfig.get_platform() falls back to
>> distutils.util.get_platform() when the _PYTHON_HOST_PLATFORM
>> environment variable is unset, and returns the *build host* arch.
>> The result is baked into the produced wheel's filename
>> (pkg-<ver>-cp3XX-cp3XX-<PLAT>.whl) and its <pkg>.dist-info/WHEEL
>> metadata.
>>
>> Without this export, extension-module wheels tag themselves with
>> whichever arch the autobuilder worker happens to have, which makes
>> the resulting rpm non-reproducible across mixed-arch autobuilder
>> pools.
>>
>> Since python 3.14 + wheel >=0.44 also rejects wheels with an empty
>> platform tag ("Bad wheel filename"), the value has to be non-empty
>> in every class scope. HOST_ARCH is class-scoped by bitbake
>> (target/native/cross/nativesdk), so `linux-${HOST_ARCH}` covers
>> every case with a single unconditional assignment. Setting per-class
>> overrides to "" (as an earlier draft did) trips the wheel filename
>> check and must be avoided.
>>
>> Pure-python wheels ignore _PYTHON_HOST_PLATFORM and remain tagged
>> "any", so this change only affects wheels that ship compiled
>> extensions (cffi, cryptography, numpy, bcrypt, markupsafe, psutil,
>> rpds-py, websockets, ...).
>>
>> Assisted-by: AI - OpenAI
>> Signed-off-by: Alejandro Hernandez<alhe@linux.microsoft.com>
>> ---
>>   meta/classes-recipe/python_pep517.bbclass | 16 ++++++++++++++++
>>   1 file changed, 16 insertions(+)
>>
>> diff --git a/meta/classes-recipe/python_pep517.bbclass b/meta/classes-recipe/python_pep517.bbclass
>> index d6246af5c2..06b23d3cba 100644
>> --- a/meta/classes-recipe/python_pep517.bbclass
>> +++ b/meta/classes-recipe/python_pep517.bbclass
>> @@ -30,6 +30,22 @@ PEP517_INSTALL_PYTHON:class-native = "nativepython3"
>>   # pypa/installer option to control the bytecode compilation
>>   INSTALL_WHEEL_COMPILE_BYTECODE ?= "--compile-bytecode=0"
>>   
>> +# Force the wheel's platform tag to reflect the target machine rather than the
>> +# build host. Without this, extension-module wheels bake the build host's arch
>> +# (e.g. linux_x86_64) into <pkg>.dist-info/WHEEL via distutils/sysconfig's
>> +# get_platform(), which breaks reproducibility across autobuilder workers of
>> +# different architectures. Pure-python wheels ignore this variable and remain
>> +# tagged "any".
>> +#
>> +# HOST_ARCH is class-scoped by bitbake:
>> +#   - target:     HOST_ARCH == TARGET_ARCH (reproducible across workers)
>> +#   - native:     HOST_ARCH == BUILD_ARCH  (matches worker arch, unpackaged)
>> +#   - cross:      HOST_ARCH == BUILD_ARCH
>> +#   - nativesdk:  HOST_ARCH == SDK_ARCH
>> +# so a single expression covers every class without producing an empty tag
>> +# (which python 3.14 + wheel >=0.44 rejects with "Bad wheel filename").
>> +export _PYTHON_HOST_PLATFORM = "linux-${HOST_ARCH}"
> The fix looks right, thanks!
>
> I'm not sure we need the 10+ lines of explanation from AI ;-) The
> commit message could be more concise too. We're seening way too much
> text being added by AI in general and it will make the codebase harder
> to understand in the long run :(.
>
> Cheers,
>
> Richard

I agree, I can send a v2 or just keep it in mind for next time, let me know.

Cheers,
Alejandro

>
>
>
> -=-=-=-=-=-=-=-=-=-=-=-
> Links: You receive all messages sent to this group.
> View/Reply Online (#242937):https://lists.openembedded.org/g/openembedded-core/message/242937
> Mute This Topic:https://lists.openembedded.org/mt/120602086/4354175
> Group Owner:openembedded-core+owner@lists.openembedded.org
> Unsubscribe:https://lists.openembedded.org/g/openembedded-core/unsub [alhe@linux.microsoft.com]
> -=-=-=-=-=-=-=-=-=-=-=-
>