If both Python 2.x and 3.x are available, reticulate prefers Python 2.x despite being deprecated

[s-u]

On most systems both Python 2.x and 3.x are available as python and python3 respectively, but despite Python 2 being deprecated, reticulate will give it preference to Python 3.x. For example:

> reticulate::py_discover_config()
python:         /usr/bin/python
[...]
NOTE: Python version was forced by PATH

python versions found: 
 /usr/bin/python3
 /usr/bin/python

This is suboptimal since reticulate itself warns that Python 2 is deprecated so it should prefer Python 3.

It requires explicit setting via use_python or RETICULATE_PYTHON in order to force the use of Python 3 (which is the only version that actually works with reticulate since Python 2 will fail with undefined symbols, but that's another story).

Given that the choice is obvious and that reticulate already knows that both Python versions are available (see above) it should pick Python 3 over Python 2.

[eddelbuettel]

If this is Ubuntu or Debian based consider sudo apt install python-is-python3. I have no idea why that is not a default.

reticulate follow the value in $PATH, and given all the complexity in different setups, also allows you to override in various ways that I mostly lost track of. But I do have this in ~/.Renviron, probably did for umpteen years but also think I set it at the to avoid being dragged into various conda setups I generally try to avoid.

## reticulate, please use my system python
RETICULATE_PYTHON=/usr/bin/python3

[t-kalinowski]

I opened a PR which should fix the issue. Thanks for reporting.

I think this was a rare edgecase. It's literally last in the order of discovery. The default experience for users that don't express a preference for a particular python installation is to create and use a r-reticulate virtual env. I think most users eventually end up expressing a preference in one form or another (e.g., by setting RETICULATE_PYTHON)

When falling back to the PATH, we also explicitly already look for python3 before python:

reticulate/R/config.R

Lines 339 to 340 in b1c33d5

	Sys.which("python3"),
	Sys.which("python")

The only circumstance where we would have prioritized python over python3 would have been if python3 was the wrong arch or was missing numpy, while python was both the correct arch and had numpy installed.

[s-u]

@t-kalinowski thanks, your comments describe at least part of the logic which is very helpful as it also explains why one version is preferred over the other. Can you, please, document the full logic in the help file - e.g. in ?py_discover_config?

That said, the PR does not change the behavior in the case where numpy is available in Python2 and not Python3. In the case that triggered this investigation it was caused by an R package that doesn't use numpy so even though all module dependencies of the package were satisfied in Python3 reticulate still picked Python2 because of the numpy requirement which was in that case unnecessary. I'm ok if this is simply documented (if that is the desired behavior) since then at least the user can fix the problem.

[t-kalinowski]

Thanks @s-u, I updated the docs.