Installing extras using conda

Question

When writing a setup.py I can specify extras_require and give a list of dependencies that are needed for additional functionality of my tool like this:

setup(
    name = "mypackage",
    install_requires = ["numpy"],
    extras_require = {
        "plotting": ["matplotlib"],
    }
)

I uploaded my package to PyPI and a conda channel and tried to install it, including all extras. From PyPI I can install the extras using:

$ pip install mypackage[plotting]

However, when installing my package from conda, I fail to install the extras. Is there a similar option for conda?

Answer 1

You can do this by creating a metapackage for the optional dependency. A good example for this is the matploptlib-feedstock. Here you have the main package matplotlib-base that contains all code and the metapackage matplotlib that depends on matplotlib-base and its optional Qt dependencies.

With the matplotlib example in mind, you could have the following outputs in your recipe:

package:
  name: some_pkg

…usual recipe contents…

outputs:
  - name: some_pkg

  - name: some_pkg_with_optional_dep
    requirements:
      host:
        - python
      run:
        - python
        - optional_dependency
        - {{ pin_subpackage('some_pkg', exact=True) }}
    test:
      imports:
        - some_pkg

Note that when you require a specific version of an optional dependency, you can specify it in the optional package but then the version constraint will only be applied when you have some_pkg_with_optional_dep installed.

To have the version constraint on the optional dependency respected independently of the presence of some_pkg_with_optional_dep, you should specify it in the run_constrained section:

requirements:
  …
  run_constraint:
    - optional_dependency >=1.2

Answer 2

As of 18-APR-2017, this is not currently possible. See https://github.com/conda/conda/issues/3299 and https://github.com/conda/conda/issues/2984

The solution that I generally use is just to make everything that might be optional as a mandatory dependency. This tends to work out, since conda packages are available for most platforms, particularly for common dependencies, and if a dependency isn't available, writing a conda recipe and uploading it to Anaconda.org is relatively easy.