Bazel doesn't recognize that /usr/bin/python is Python 3, not 2.

[philwo]

Description of the problem / feature request:

On systems where "python" on the PATH is Python 3 (e.g. Arch Linux has /usr/bin/python = Python3, also some macOS Homebrew installations seem to have /usr/local/bin/python as Python3), Bazel doesn't realize that and believes it's using Python 2.

Thus py_binary's that use PY2 (which is still the default for Bazel, so that would be almost all binaries) and are not compatible with Python 3 will mysteriously fail to run.

Some examples of where this caused confusion:

• Make Python scripts in build_defs compatible with Python 3. intellij#452
• python SyntaxError: invalid syntax when building plugin intellij#453

Bugs: what's the simplest, easiest way to reproduce this bug? Please provide a minimal example if possible.

Bazel should detect the correct Python 2.x and 3.x versions from the PATH, by inspecting the version output of "python", "python2" and "python3" and then choosing appropriate binaries.

What operating system are you running Bazel on?

Arch Linux

[philwo]

@brandjon Don't know if you're the right person to assign this to - if not, please reroute. :) Thank you!

[jerryzj]

That's a very stupid bug. When I tried to build TensorFlow r1.12 with CUDA 10, I have to install the outdated python instead of python3...

[brandjon]

I am the right person. I'll add this to the overall python mode tracking bug #6444. Note that due to #4815, even when the Python mode is set, Bazel doesn't select the "python3" alias in preference to "python".

[brandjon]

FYI the solution to this will be that the default Python toolchain will verify the version of the Python runtime that it autodetects. Follow #7375 for progress on Python toolchain work.

[brandjon]

I'm closing this as redundant with #4815 and #7375. When toolchains are enabled (--incompatible_use_python_toolchains, see #7899), the autodetecting toolchain (non-Windows platforms only) specifically checks that the python command it finds has the expected version. If there's a problem, you can either change your system installation of Python or define an explicit Python toolchain for your platform.