Since the inclusion of LLVM, arrow wheels stopped working with TensorFlow again (on some configurations at least).
While we are continuing to discuss a more permanent solution in https://groups.google.com/a/tensorflow.org/d/topic/developers/TMqRaT-H2bI/discussion, I made some progress in creating tensorflow compatible wheels for an unmodified pyarrow.
They won't adhere to the manylinux1 standard, but they should be as compatible as the TensorFlow wheels because they use the same build environment (ubuntu 14.04).
I'll create a PR with the necessary changes. I don't propose to ship these wheels but it might be a good idea to include the docker image and instructions how to build them in the tree for organizations that want to use tensorflow with pyarrow on top of pip. The official recommendation should probably be to use conda if the average user wants to do this for now.