Load CondaPkg lazily, and drop Pkg as a dependency

[JamesWrigley]

Combined, these reduce the load time with JULIA_PYTHONCALL_EXE set from ~1.3s to ~0.3s on my machine.

Note that removing Pkg as a dependency could be breaking if packages are depending on accessing Pkg in the juliacall module. But, only Main is documented as being available so it is technically ok to remove it in a nonbreaking release: https://juliapy.github.io/PythonCall.jl/stable/juliacall-reference/#Constants

[JamesWrigley]

(bump)

[cjdoris]

Thanks, it's a nice idea, however doing runtime imports like this is not allowed in Julia as it breaks precompilation. I don't totally understand how but there are issues/PRs about it somewhere...

What we CAN do though is, once the PR for using Preferences.jl is merged, avoid loading CondaPkg if the exe pref is set to something other than @CondaPkg. This is because preferences can be used at precompile-time. We do something similar in CondaPkg itself to avoid loading MicroMamba_jll if we don't need it.

Closing as the approach is unsound but happy to take a PR for the preferences way.

[JamesWrigley]

@eval using CondaPkg is indeed not allowed and you get an error about it, but AFAIK Base.require() is safe.

What we CAN do though is, once the PR for using Preferences.jl is merged, avoid loading CondaPkg if the exe pref is set to something other than @CondaPkg. This is because preferences can be used at precompile-time. We do something similar in CondaPkg itself to avoid loading MicroMamba_jll if we don't need it.

Ok, that works for me too 👍