pip installing fails if numpy is not already installed

[jdavidheiser]

Since setup.py tries to import numpy before even defining any dependencies, it's possible for the numpy import to be attempted before the dependencies trigger installing numpy. This made it difficult to install fastparquet as part of a larger requirements.txt.

[martindurant]

Yes, I can see how that might be a problem. Can you think of a way around it?

[jbrockmendel]

Just ran into this problem. An issue over in pip has a discussion about this that mentioned bottleneck. bottleneck has a solution that includes a comment crediting stackoverflow.

# workaround for installing bottleneck when numpy is not present
class build_ext(_build_ext):
    # taken from: stackoverflow.com/questions/19919905/
    # how-to-bootstrap-numpy-installation-in-setup-py#21621689
    def finalize_options(self):
        _build_ext.finalize_options(self)
        # prevent numpy from thinking it is still in its setup process
        __builtins__.__NUMPY_SETUP__ = False
        import numpy
        # place numpy includes first, see gh #156
        self.include_dirs.insert(0, numpy.get_include())

[martindurant]

@jjhelmus , your solution appears to work; is there any benefit to the one above mentioned by @jbrockmendel ?

[jjhelmus]

I believe that the solution suggested by @jdavidheiser and in the stack overflow post works around the numpy dependency but Cython would need to be install before calling setup.py or pip. The stack overflow comments mention this limitation and do not offer a work around.

The solution in #171 works by examining the command line arguments which allows pip to be run without NumPy or Cython installed.