Make InterruptException capturable in a program

[tkf]

In julia 1.0 (probably since 0.3?), Julia program launched outside REPL (and -e and -E) cannot catch InterruptException:

$ cat catch-sleep.jl
try
    println("sleeping...")
    sleep(10)
catch err
    @show err
finally
    println("woke up")
end

$ julia --startup-file=no catch-sleep.jl
sleeping...
^C
signal (2): Interrupt

This PR makes it work:

$ julia --startup-file=no catch-sleep.jl
sleeping...
^Cerr = InterruptException()
woke up

It seems ccall(:jl_exit_on_sigint, Void, (Cint,), 1) is added in #6754 but I don't see the discussion why it was required. I also checked the PRs touching this code but couldn't find any relevant discussion (#25162, #23775, #12679, #11347, #9482).

I think it makes sense for various ways of running Julia program

$ julia script.jl
$ julia -e 'include("script.jl")'
$ julia
julia> include("script.jl")

behave consistently. Are there any reasons why exit_on_sigint is set to 1 only in the case of julia script.jl?

Even if this change is reasonable, since this breaks the API, I suppose this has to be blocked until 2.0? Still, adding CLI option --handle-interrupt or a documented Julia function wrapping jl_exit_on_sigint would be nice.

closes #25308

[JeffBezanson]

I think the idea of this was that interrupting code is mostly useful in interactive mode, while in a "program" you expect it to just stop (like it does for C programs).

[tkf]

The usecase I have in mind is some kind of long-running computation with some buffer which has to be flushed before exit:

handle = prepare_computation()
try
    while !finished(handle)
        dump_computation("output/file", handle)
        do_some_computation_for_a_while!(handle)
    end
finally
    dump_computation("output/file", handle)
end

(which can be nicely converted into a do block with some wrapper).

In HPC clusters, you often have an option to send a signal before the computation limit time. This PR is useful for such case since you can (relatively) safely shutdown the computation without loosing intermediate results. Of course, you can do it with atexit but then it makes it hard to write correct code. For example, imagine that you need to do two of the above computations in the same Julia script. You may not want to trigger the tear-down code for the first computation while sigint is received during the second computation. It's doable, but isn't try-catch/finally much nicer to write?

Also, since Python can catch KeyboardInterrupt in script, being able to catch InterruptException was my first expectation.

[JeffBezanson]

I think your best bet for now is to call ccall(:jl_exit_on_sigint, Cvoid, (Cint,), 0). I agree we should add a documented wrapper for this; that would be a new feature and so can be added in 1.x.

[tkf]

Yes, it makes sense that you can't change it in 1.x. But can it get a 2.0 milestone? It looks like I'm not the only one that finds this behavior strange:

#25308
#19222
https://stackoverflow.com/questions/40937921/julia-0-4-7-how-to-catch-sigterm
https://stackoverflow.com/questions/20831290/julia-handle-keyboard-interrupt/29270413