Add precompile workload

[rikhuijzer]

In response to the comments at JuliaAI/MLJ.jl#1016, this PR suggests adding a precompile workload here to MLJBase.jl.

I'm running a bit into an issue though. I can't run evaluate! because that one calls MLJModelInterface.nrows at src/resampling.jl#1289. As far as I understand, the problem is that the interface mode in MLJModelInterface (INTERFACE_MODE) is not set during precompilation yet. However, I also cannot set the global constant because that will likely "bake in" the constant during precompilation via constant propagation, which then could give problems later on.

Is there a way we could get rid of the global constant? Maybe that could make precompilation easier. This might also have some other benefits down the line. I do understand that it's a big ask, though, so we can also continue without calling evaluate!.

[codecov-commenter]

Codecov Report

Merging #900 (3e0ed3b) into dev (a7f2814) will decrease coverage by 0.78%.
The diff coverage is 0.00%.

📣 This organization is not using Codecov’s GitHub App Integration. We recommend you install it so Codecov can continue to function properly for your repositories. Learn more

@@            Coverage Diff             @@
##              dev     #900      +/-   ##
==========================================
- Coverage   85.81%   85.03%   -0.78%     
==========================================
  Files          40       41       +1     
  Lines        3567     3562       -5     
==========================================
- Hits         3061     3029      -32     
- Misses        506      533      +27     

Impacted Files	Coverage Δ
src/MLJBase.jl	60.00% <ø> (ø)
src/precompile.jl	0.00% <0.00%> (ø)

... and 7 files with indirect coverage changes

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

[ablaom]

Thanks for this. Let me give a little context to see if that helps you see a solution. The idea that INTERFACE_MODE is a switch to allow new model implementations to avoid certain package dependencies (such as Tables.jl) when they depend on MLJModelInterface.jl. (We had tried a Requires.jl solution at the time but got issues.) As we document here, these implementations don't actually function until MLJBase is imported, which changes the value of the constant.

Is there a way we could get rid of the global constant?

Unfortunately, I don't see a way to do this that isn't disruptive. I guess we could use the new optional dependencies feature of 1.9, but that would still be quite a bit of refactoring and would be breaking. I'm not sure breaking changes to MLJModelInterface.jl are even conceivable at this point. That's for MLJ2 ha ha.

A key point is that this: once MLJBase is imported, there is no reason for INTERFACE_MODE to have any value other than FullInterface.

Currently the switch happens in MLJBase.jl/init.jl. Is it possible we move it out of init.jl? I have gotten into trouble trying to modify global constants in other packages outside of init.jl, so I could be missing something here.

[rikhuijzer]

I unfortunately cannot figure out how to get a reasonable speedup without hitting the global constant, so will close this PR

[ablaom]

Thanks @rikhuijzer for spending time looking into this. Appreciate it.