[Wheel-builders] Specifying configuration options and platform of built wheels

[Fridolín Pokorný]

Hi all,

we are trying to provide built wheels of TensorFlow. TensorFlow can be
built with various configuration options that include/exclude certain
functionality for the resulting wheel artifact - we would like to provide
multiple wheels based on desired configuration. An example can be support
for kafka, s3, xla, hdfs. The main issue we are struggling with is a lack
of specifying such features in python packages for users. Considering the
current wheel binary package format (PEP 491) there is no way to specify
such options in package names in the wheel naming convention.

Even if there would be such ability, there are issues on how to specify
different configurations from the available ones. Considering different
wheels (e.g. tensorflow-1.8.0-py3-cp36m-linux_x86_64+hdfs+s3+kafka+xla.whl)
there is no straightforward way to specify desired features that an
installed wheel should provide. As a user, I don’t want to consume packages
that are named “tensorflow+hdfs+s3+kafka+xla” so I think this deserves
broader discussion in community and design decisions for providing
configuration and platform specific features for wheels.

Are there any plans from community to incorporate such configuration
options for wheels? Also, how these compile time options should be included
in the resulting wheel as metadata?

Thanks,
Fridolin
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/wheel-builders/attachments/20180702/9588cbe1/attachment.html>