Conda.add may downgrade Python

[tkf]

If a package installed via Conda.add is not available for installed Python version, conda may downgrade Python:
https://discourse.julialang.org/t/help-understanding-pycall-related-travis-failure/15957

The error is cryptic when it's installed via pyimport_conda since the Python installed in disk is different from the one loaded in PyCall.

Maybe we should pin Python version and provide a nicer error message pointing to a command that re-builds Conda and PyCall? Or detect change in Python version, automatically run Pkg.build, and then throw an error with message saying that you need to restart Julia?

[stevengj]

Pinning Python or having a better error message in pyimport_conda seems like a good idea, but an issue for PyCall.jl, not Conda.jl?

I'm beginning to think that #108 was a mistake … Python 3 is just causing way too many problems. Maybe we should default to installing a slightly older version of Python 3?

[tkf]

I think switching to Python 3 is a good choice. matplotlib 3.0 just dropped Python 2 and pandas is going to switch to Python 3-only soon https://python3statement.org/#sections40-timeline

a slightly older version of Python 3

Yeah, this issue is more like 3.6-vs-3.7 than 2-vs-3. I guess it's safer to run conda install python=3.(x-1) in build.jl after Python 3.x is released (say for 6 months or maybe even one year).

an issue for PyCall.jl, not Conda.jl

Yeah, I wasn't sure which issue tracker I should post this.