World age problems when using @everywhere in a function

[josia-pool]

I have a julia (1.12) file that is structured something like this:

using Distributed

function main()
  @everywhere variable = [0, 1, 2, 3, 4, 5]

  for x in variable
    println(x^2)
  end
end

main()

From my understanding @everywhere should define the variable on every worker, including the main process.
But when I run the code, it give me this error:
UndefVarError: `variable` not defined in `Main` The binding may be too new: running in world age 38716, while current world is 38717.

In a jupyter notebook, I can run main() again a second time, and then it works, since I assume then the world age matches...

Can someone explain what is going on and if there is some way to circumvent this problem? I can fix it by defining it again in the main process variable = [0, 1, 2, 3, 4, 5], or by taking the code out of the main() function but to me that doesn't feel intended.

[simeonschaub]

The semantics around globals have been reworked significantly to be more consistent in Julia 1.12. It errors for the same reasons something like the following now errors, because the global has only been declared in a later world age:

julia> let
           eval(:(foo = 1))
           foo
       end
ERROR: UndefVarError: `foo` not defined in `Main`
The binding may be too new: running in world age 38696, while current world is 38697.
Stacktrace:
 [1] top-level scope
   @ REPL[1]:3

It might make sense for @everywhere to insert (latestworld-if-toplevel) though, similar to what @eval does, which would fix the specific case here. (It wouldn't fix this particular case, but similar cases, if the binding occurs outside a function.)

You need to either add invokelatest to the binding access, something like:

function main()
  @everywhere variable = [0, 1, 2, 3, 4, 5]

  for x in @invokelatest getglobal(Main, :variable)
    println(x^2)
  end
end

or declare the global explicitly like so:

function main()
  global variable
  @everywhere variable = [0, 1, 2, 3, 4, 5]

  for x in variable
    println(x^2)
  end
end

[josia-pool]

Thanks for your explanation, I understand now. Would there also be a way to have the variable globally on the workers, but in local scope on the master process?