[Biopython-dev] Using NumPy/SciPy/scikit-learn approach to building wheels

[Peter Cock]

A brief update - the multibuild system works in that we can
build wheels for Linux, Mac, and Windows (32-bit and 64-bit).

https://github.com/biopython/biopython-wheels

Logs:

https://travis-ci.org/biopython/biopython-wheels/builds
https://ci.appveyor.com/project/biopython/biopython-wheels/history

I've not yet asked for the RackSpace credentials to actually
save the wheels prior to uploading them to PyPI.

Unfortunately, getting all the tests to pass is proving frustrating -
the multibuild platform has highlighted some locale issues:

https://github.com/biopython/biopython/issues/855
https://github.com/biopython/biopython/issues/1319
https://github.com/biopython/biopython/issues/1320
https://github.com/biopython/biopython/issues/1321

And a strange floating point nan issue:

https://github.com/biopython/biopython/issues/1316

Plus a minor int/long doctest which I fixed (5 vs 5L), fixed:

https://github.com/biopython/biopython/commit/a1d14584f3fa8002b42ebcf9b79ff26fe82235d7

On the bright side, so far there is nothing which makes me
want to rush out another release.

I'm happy now with the Windows tests (skipping two problem
tests), so perhaps we should use these wheels as is?

https://ci.appveyor.com/project/biopython/biopython-wheels/build/1.0.23

Peter

On Fri, Jul 7, 2017 at 4:31 PM, Peter Cock <p.j.a.cock at googlemail.com> wrote:
> Matthew Brett is happy with this plan :)
>
> I'll be creating a new repository for this (probably next week):
>
> https://github.com/biopython/biopython-wheels
>
> Once that is up and running, the goal would be to make a
> trivial commit there to update the version tag, and then TravisCI
> and AppVeyor would be triggered to fetch that tagged commit
> from the main repository, and build all the compiled wheels for us.
>
> e.g. An example from AstroPy:
>
> https://github.com/MacPython/astropy-wheels/commit/cf5f4e80d3a6644f7aadc6375662209be85ca2c1
>
> For now it would be a manual step (after final QC) to upload
> the wheels to PyPI for public use.
>
> Regards,
>
> Peter
>
> On Wed, Jul 5, 2017 at 11:48 AM, Peter Cock <p.j.a.cock at googlemail.com> wrote:
>> Dear Biopythoneers,
>>
>> RE: https://github.com/biopython/biopython/issues/1296
>>
>> There has been an impressive amount of work in the wider
>> NumPy ecosystem to automate building compiled wheels for
>> Linux, Mac and Windows:
>>
>> Here is how NumPy etc does their wheels:
>>
>> https://github.com/MacPython/numpy-wheels
>>
>> A more relevant example of another project using NumPy:
>>
>> https://github.com/MacPython/astropy-wheels
>>
>> Background to the setup:
>>
>> https://github.com/MacPython/wiki/wiki
>>
>> https://github.com/matthew-brett/multibuild
>>
>> Unless anyone has a better idea, I plan to ask Matthew Brett
>> if Biopython can also use the Rackspace donated resources
>> originally for the scikit-learn team for hosting the wheels prior
>> to uploading to PyPI:
>>
>> https://github.com/MacPython/wiki/wiki/Managing-Rackspace
>>
>> Regards,
>>
>> Peter