Add section discussing conda and Jupyter notebooks

[davidrpugh]

@jakevdp has an excellent blog post that summarizes the key issues.

Can this post be summarized in a callout box? Or is more discussion required?

[jakevdp]

That post predates newer functionality; IPython 7.3 introduced pip and conda magic functions so that %conda install package essentially does what you want it to, without having to put too much thought into the issues discussed in that post.

[davidrpugh]

@jakevdp Thanks for the tip. I had always avoided installing packages from inside notebooks in favor of maintaining a single environment.yml or requirements.txt file. Installing packages from inside the notebook feels like an anti-pattern that could lead to problems. However once I started using Google Colab and Kaggle kernels I started installing packages from within Notebooks out of necessity.

I just tried to see if the %conda magic command was available via Google Colab and it appears that it is not available (the %pip magic command was available) also 'conda was not available either via !conda (guess that conda is not installed in the Docker image used by Colab?); on Kaggle conda was available using !conda from a code cell but the %conda magic command was not available.

I am considered adding a callout box encouraging the use of environment.yml files wherever possible, but use of %conda magic as the preferred method of installing packages from within a Jupyter notebook using conda (where necessary).

Would be interested in your thoughts on this.

[jakevdp]

I think there are two different modes of work: interactive exploration & data analysis (which suits itself well to Jupyter notebooks and packages installed within the active kernel on the fly) and development of reproducible analyses & tools (which suits itself well to version-controlled .py files and well-defined & reproducible environments). I don't think either is an anti-pattern; they are modes of work that both have their place in any data science workflow.