Conda-build: "numpy=1.11" is enforced if "numpy" w/o version is a build requirement

[sscherfke]

Actual Behavior

1. I have a recipe with the following build requirements:

   requirements:
     build:
       - python
       - setuptools
       - numpy
   

2. Our internal conda index contains numpy 1.12 and newer versions

3. conda build <recipe> fails with:

   ...
   conda.exceptions.ResolvePackageNotFound:
     - numpy=1.11
   
   ...
       raise DependencyNeedsBuildingError(exc, subdir=subdir)
   conda_build.exceptions.DependencyNeedsBuildingError: Unsatisfiable dependencies for platform linux-64: {'numpy=1.11'}
   

4. The user is completely confused where the numpy=1.11 requirement comes from why the available numpy versions are ignored.

5. Changing the requirments (in every package that requires numpy) works around the issue:

   requirements:
     build:
       - python
       - setuptools
       - numpy >= 1.12
   

   This would install, e.g., numpy=1.14.5 during the build.

Expected Behavior

Either one (in descending order):

1. If no numpy-version is specified for the build, install the latest available version.
2. If no numpy-version is specified for the build, try to use 1.11. Fall back to a newer version if 1.11 is not available. Maybe emit a warning.
3. If no numpy-version is specified and it is for whatever reasons important that 1.11 is used in this case, emit a proper error message that tells users what the actual problem is and how to fix it.

Steps to Reproduce

$ mkdir /tmp/example
$ vim /tmp/example/meta.yaml
package:
  name: example
  version: 1.0

requirements:
  build:
    - python
    - numpy
    # - numpy >=1.12

build:
  script: 'echo onoes'
$ conda build /tmp/example

Output of conda info

     active environment : None
            shell level : 0
       user config file : /home/stefan/.condarc
 populated config files : /home/stefan/emsconda/.condarc
          conda version : 4.5.9
    conda-build version : 3.12.0
         python version : 3.6.5.final.0
       base environment : /home/stefan/emsconda  (writable)
           channel URLs : https://.../conda/stable/linux-64
                          https://.../conda/stable/noarch
          package cache : /home/stefan/emsconda/pkgs
                          /home/stefan/.conda/pkgs
       envs directories : /home/stefan/emsconda/envs
                          /home/stefan/.conda/envs
               platform : linux-64
             user-agent : conda/4.5.9 requests/2.19.1 CPython/3.6.5 Linux/4.17.11-100.fc27.x86_64 fedora/27 glibc/2.26
                UID:GID : 1119:1119
             netrc file : None
           offline mode : False

[sscherfke]

The version 1.11 is hard-coded in https://github.com/conda/conda-build/blob/4e6b1f68f966291770e055141252ef13aeb293c8/conda_build/variants.py#L23

[mingwandroid]

Don't use this file then? We deliberately target old numpy as it's forward
compatible.

On Wed, Aug 15, 2018, 7:41 AM Stefan Scherfke [e-mail]
wrote:

The version 1.11 is hard-coded in
https://github.com/conda/conda-build/blob/4e6b1f68f966291770e055141252ef13aeb293c8/conda_build/variants.py#L23

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
https://github.com/conda/conda-build/issues/3081#issuecomment-413108272,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AA_pdAAu8Wx7Ye8BRX2krXLYtn3Rk4YDks5uQ8J8gaJpZM4V9ooL
.

[msarahan]

We could probably get rid of this. We depend on conda_build_config.yaml to set this for ourselves: https://github.com/AnacondaRecipes/aggregate/blob/master/conda_build_config.yaml#L117

You should probably do the same yourself. If we remove this, your unpinned recipes will float up to the latest numpy, and they won't be backwards compatible.

[sscherfke]

Don't use this file then? We deliberately target old numpy as it's forward compatible.

I cannot simply not use this file. This is a built-in file from conda-build. If I don't have any explicit conda_build_config.yaml files, this built-in is used.

I’m okay with keeping 1.11 as default but if this version cannot be installed, conda-build should either use any version that it can find or print a better error message. The current error message looks as if it is the user's fault b/c they have somewhere pinned numpy to 1.11.