No description, website, or topics provided.
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.
multibuild @ e7a0246
pandas @ 07739aa


Building and uploading pandas wheels

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

The driving github repository is

How it works

The wheel-building repository:

  • does a fresh build of any required C / C++ libraries;
  • builds a pandas wheel, linking against these fresh builds;
  • processes the wheel using delocate (OSX) or auditwheel repair (Manylinux1). delocate and auditwheel copy the required dynamic libraries into the wheel and relinks the extension modules against the copied libraries;
  • uploads the built wheels to (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 This encrypted key gives the travis build permission to upload to the Rackspace directory pointed to by

Triggering a build

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 pandas-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.

Which pandas commit does the repository build?

The pandas-wheels repository will build the commit specified in the BUILD_COMMIT at the top of the .travis.yml file. This can be any naming of a commit, including branch name, tag name or commit hash.

Uploading the built wheels to pypi

Be careful, 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

The same contents appear at; 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 :

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:

wheel-uploader -r warehouse -u $CDN_URL -s -v -w ~/wheelhouse -t macosx pandas $VERSION
wheel-uploader -r warehouse -u $CDN_URL -s -v -w ~/wheelhouse -t manylinux1 pandas $VERSION


  • -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.

pandas is the root name of the wheel(s) to download / upload, and 0.18.1 is the version to download / upload.

In order to use the Warehouse PyPI server, you will need something like this in your ~/.pypirc file:

index-servers =


username: your_user_name
password: your_password

So, in this case, wheel-uploader will download all wheels starting with pandas-0.18.1- from to ~/wheelhouse, then upload them to PyPI.

Of course, you will need permissions to upload to PyPI, for this to work.