Should @main be run after julia -L file.jl -e "do_something()"?

[lukas-weber]

I noticed recently that julia will run the new @main entry point even after running an expression with the -e flag. This is a bit surprising. Consider the following workflow.

# runscript.jl
using MyPackage

config = generate_config()

function (@main)(args)
    run_heavy_calculation(args)
end

This script can then be submitted e.g. via slurm for some production run. It would be really nice if I could do the following

julia -L runscript.jl -e "run_diagnostics(config)"

to do some sanity checks or print some configuration parameters without running the heavy calculation. However, in julia 1.11.3, @main will be run after run_diagnostics.

A workaround is to do julia -L runscript.jl -e "(@main)(_) = (run_diagnostics(config);0)". However, I would expect -e to behave like running julia -L runscript.jl and then entering the commands myself, which is clearly not the case. The same occurs with the -E flag, which is even more surprising.

Maybe changing this behavior would be a breaking change and we are stuck with it. Maybe it is surprising enough to be considered a bug that can be fixed?

[Keno]

I think there's a question of what happens if you do -e include("runscript.jl"). Does main run? I think people probably expect it would because they'd expect it to be equivalent to running runscript.jl directly which does run main. I think this points to the problem being -L. I could see an argument for adding a special case that suppresses main when passed to -L (to facilitate precisely this use case), but it's a bit weird.

[lukas-weber]

I think if -e 'include("runscript.jl"); do_something()' behaves differently from -L runscript.jl -e 'do_something()', it would be more confusing than the current situation.

An alternative solution would be to add a --nomain flag after all.

[Keno]

The other thing I can think of is that -e could be semantically considered as defining a main internally, so it would only run main if that main was defined inside -e. I think that could be acceptable.

[lukas-weber]

That’s probably the perfect solution for my use case. 😄