Make aplpy conda install work with older numpy versions

[cdeil]

I'm trying to improve the Gammapy build matrix and am getting this issue in the builds testing against older numpy versions and installing the optional dependencies like aplpy: https://travis-ci.org/gammapy/gammapy/jobs/77556117#L479

Hint: the following packages conflict with each other:
  - numpy 1.8*
  - aplpy

@astrofrog – Shouldn't aplpy work with Numpy 1.8?

Maybe the numpy version requirement isn't declared properly for the aplpy conda package?

[cdeil]

Same issue with Numpy 1.6 and 1.7. I updated the title to reflect the more general question / feature request.

[astrofrog]

I thought it was possible to specify the dependency as numpy > 1 and it would all with all Numpy versions, but I may be wrong.

[mwcraig]

The problem is that conda used to pin the numpy version when a package was built to the version of numpy installed in the build environment. The pinning was baked into the last part of the file name, e.g. _np19py27. That's great if you have a package that really needs a specific version of the numpy ABI, but not so helpful if the package handles a range of numpy versions.

This behavior was changed in a recent version of conda-build, so that it should be possible to build the affiliated packages in a numpy-version-agnostic way now. The alternative would be to maintain a complicated build matrix of numpy versions that would probably vary from package to package (i.e. not really an option).

Will hopefully get a chance to fix it up this weekend -- if someone wants to take a crack at it before then, I think all that would be necessary is remove CONDA_NPY from the build matrix (see https://github.com/conda/conda-build/pull/493)

[cdeil]

@mwcraig – Thanks!

If it doesn't work out, it's not a huge deal, it just complicates the setup of dependencies from my CI builds a bit, for end users just having one numpy version available should be sufficient (unless some of their packages pin it to different versions).

[mwcraig]

@cdeil -- good news, all new builds of affiliated packages will not have the numpy version pinned unless the recipe or setup.py pins the numpy version. This is due to a change in conda build, and should simplify CI both for this and for astropy on older versions of numpy.

[mwcraig]

@cdeil -- one other thought. If you want me to nuke the current aplpy builds and replace them with numpy agnostic versions, let me know. Wouldn't take much effort.

[mwcraig]

@cdeil -- this is now fixed, finally...