Use setup_requires to ensure cython and numpy are present

[j-santander]

Addressing issue #256
Summary of changes:

I'm using setup_requires to indicate the dependencies on "setup" time, with install_requires to indicate the dependencies on "run" time.

I'm also ensuring the setup_requires is evaluated before making any reference to them.

As numpy version I've used the declared dependency for Pandas 0.17.0

[codecov-io]

Current coverage is 87.25% (diff: 100%)

Merging #260 into master will increase coverage by 0.06%

@@             master       #260   diff @@
==========================================
  Files             5          5          
  Lines           414        416     +2   
  Methods           0          0          
  Messages          0          0          
  Branches          0          0          
==========================================
+ Hits            361        363     +2   
  Misses           53         53          
  Partials          0          0          

Powered by Codecov. Last update 6d93b2f...9c661cc

[wesm]

Need to make sure this doesn't conflict with conda environments -- setuptools has a way of wanting to build NumPy or Cython from source when it's already installed in your site-packages

[tdsmith]

A problem with using setup_requires this way is that, if they're missing, setuptools will install a temporary copy of cython and numpy for each setup.py invocation. Unfortunately, tools like pip will call setup.py several times at least twice to do things like extract metadata.

One solution is to cargo-cult this function around: https://github.com/pyca/cryptography/blob/8e66ca6813016d9fc6f57d5f1e50530fc39f78ae/setup.py#L114 -- the cryptography project runs into a similar chicken-egg problem with cffi.

Mercifully, this will be resolved in pip 10 by PEP 518 (pyproject.toml).

pip PEP 518 tracking bug: pypa/pip#3691

[tdsmith]

-- although the cryptography setup.py doesn't model a complete solution for feather since cryptography manages to avoid actually importing cffi directly, relying on cffi's setuptools extension hooks instead; packaging is hard.

[j-santander]

Yes, I know :) Thanks for looking into this.

[danpalmer]

Ping on this? I just tried to install feather-format and it failed. Not a great first-run experience for someone new to the format.

It also means we can't use a requirements.txt, as pip doesn't know the dependency and therefore can't install numpy and Cython first.

[danpalmer]

The correct name of the package is Cython – using the lowercase version will result in issues with some tooling as it is not the canonical name.

[wesm]

Closing as Cython is no longer required to create the feather Python package (which is just a convenience interface to pyarrow.feather)