Multibuild served us well for serveral years, however have moved to using cibuildwheel running on GHA on matplotlib/matplotlib.
This repo is archived for future refernce.
We automate wheel building using this custom github repository that builds on the travis-ci OSX machines and the travis-ci Linux machines.
The travis-ci interface for the builds is https://travis-ci.org/MacPython/matplotlib-wheels
The driving github repository is https://github.com/MacPython/matplotlib-wheels
The wheel-building repository:
- does a fresh build of any required C / C++ libraries;
- builds a matplotlib wheel, linking against these fresh builds;
- processes the wheel using delocate (OSX) or auditwheel
repair
(Manylinux1).delocate
andauditwheel
copy the required dynamic libraries into the wheel and relinks the extension modules against the copied libraries; - uploads the built wheels to http://wheels.scipy.org (a Rackspace container kindly donated by Rackspace to scikit-learn).
The resulting wheels are therefore self-contained and do not need any external dynamic libraries apart from those provided as standard by OSX / Linux as defined by the manylinux1 standard.
The .travis.yml
file in this repository has a line containing the API key
for the Rackspace container encrypted with an RSA key that is unique to the
repository - see http://docs.travis-ci.com/user/encryption-keys. This
encrypted key gives the travis build permission to upload to the Rackspace
directory pointed to by http://wheels.scipy.org.
You will likely want to edit the .travis.yml
file to specify the
BUILD_COMMIT
before triggering a build - see below.
You will need write permission to the github repository to trigger new builds on the travis-ci interface. Contact us on the mailing list if you need this.
You can trigger a build by:
- making a commit to the matplotlib-wheels repository (e.g. with git commit --allow-empty); or
- clicking on the circular arrow icon towards the top right of the travis-ci page, to rerun the previous build.
In general, it is better to trigger a build with a commit, because this makes a new set of build products and logs, keeping the old ones for reference. Keeping the old build logs helps us keep track of previous problems and successful builds.
The matplotlib-wheels repository will build the commit specified in the
BUILD_COMMIT
at the top of the .travis.yml
and appveyor.yml
files.
This can be any naming of a commit, including branch name, tag name or commit
hash.
Be careful, http://wheels.scipy.org points to a container on a distributed content delivery network. It can take up to 15 minutes for the new wheel file to get updated into the container at http://wheels.scipy.org.
The same contents appear at https://3f23b170c54c2533c070-1c8a9b3114517dc5fe17b7c3f8c63a43.ssl.cf2.rackcdn.com; you might prefer this address because it is https.
When the wheels are updated, you can download them to your machine manually, and then upload them manually to pypi, or by using twine. You can also use a script for doing this, housed at : https://github.com/MacPython/terryfy/blob/master/wheel-uploader
For the wheel-uploader
script, you'll need twine and beautiful soup 4.
You will typically have a directory on your machine where you store wheels, called a wheelhouse. The typical call for wheel-uploader would then be something like:
VERSION=2.0.0 CDN_URL=https://3f23b170c54c2533c070-1c8a9b3114517dc5fe17b7c3f8c63a43.ssl.cf2.rackcdn.com wheel-uploader -r warehouse -u $CDN_URL -s -v -w ~/wheelhouse -t macosx matplotlib $VERSION wheel-uploader -r warehouse -u $CDN_URL -s -v -w ~/wheelhouse -t manylinux1 matplotlib $VERSION
where:
-r warehouse
uses the upcoming Warehouse PyPI server (it is more reliable than the current PyPI service for uploads);-u
gives the URL from which to fetch the wheels, here the https address, for some extra security;-s
causes twine to sign the wheels with your GPG key;-v
means give verbose messages;-w ~/wheelhouse
means download the wheels from to the local directory~/wheelhouse
.
matplotlib
is the root name of the wheel(s) to download / upload, and
2.0.0
is the version to download / upload.
In order to use the Warehouse PyPI server, you will need something like this
in your ~/.pypirc
file:
[distutils] index-servers = pypi warehouse [pypi] username:your_user_name password:your_password [warehouse] repository: https://upload.pypi.io/legacy/ username: your_user_name password: your_password
So, in this case, wheel-uploader
will download all wheels starting with
matplotlib-2.0.0-
from http://wheels.scipy.org to ~/wheelhouse
, then
upload them to PyPI.
Of course, you will need permissions to upload to PyPI, for this to work.