@cjdoris could you approve the CI run?

Codecov Report

Merging #328 (8ad6e50) into main (e374e25) will increase coverage by 0.01%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main     #328      +/-   ##
==========================================
+ Coverage   41.50%   41.52%   +0.01%     
==========================================
  Files          76       76              
  Lines        4652     4655       +3     
==========================================
+ Hits         1931     1933       +2     
- Misses       2721     2722       +1     

... and 1 file with indirect coverage changes

What is missing to merge this commit? The bug that this commit fixes is causing issues with my co-workers when they are trying to install PythonPlot...

Pinging @cjdoris for any comments on what's blocking merge. He may just be on vacation, as I don't see any GitHub activity for about a month.

Hi yeah I've been away.

My worry with this is what happens if a package Foo that depends on PythonCall has an __init__ function which does pyimport("something")? Unless Foo.__init__ is also guarded in the same way, it assumes that PythonCall.__init__ has run already, and therefore will error because it hasn't.

Or have I misunderstood and it all works fine? I'm going to experiment a bit.

Also the linked issues all look like the same issue of conda install mysteriously failing during precompilation, which I think has been reported before but not resolved. Really efforts should be concentrated on fixing that.

Edit: it's #266

I would love to figure out the underlying issue, but this band-aid PR makes this (very annoying) problem go away. Can we merge it as a minimally-intrusive stopgap?

It is not minimally intrusive, it actually very easily breaks any downstream packages in the way I suspected above.

If I make a package Foo containing just

using PythonCall
function __init__()
    pyimport("sys")
end

then precompiling Foo fails because it runs Foo.__init__() which calls pyimport("sys") which only works if PythonCall.__init__() has already run.

I don't think this general idea can work, so I'm going to close this PR. Another option may be to initialise Python not in __init__ but whenever pyimport and friends are actually called - but this would be a larger change to the codebase.

But Foo shouldn't call __init__() during precompilation in the first place - the fact that it does is a bug that will probably need to be resolved within Julia itself. PythonPlot.jl already uses the suggested escape hatch to avoid the same bug: https://github.com/JuliaPy/PythonPlot.jl/blob/63ed915e9a3abcffa24552129d00b7f1b9099a56/src/init.jl#L147