check if DataFrames.jl is affected by Threads.@spawn bug

[bkamins]

x-ref JuliaLang/julia#40626, JuliaData/CSV.jl#1046, JuliaData/CSV.jl#1045

[bkamins]

I have checked and DataFrames.jl is affected by this issue.

[bkamins]

The question is if running this function after every @sync of spawned tasks:

function _clear_thread_state()
    Threads.@threads :static for _ in 1:Threads.nthreads()
        nothing
    end
end

is good enough to ensure clearing thread state and will not cause problems in case some other tasks are spawned independently what DataFrames.jl does.

CC @tkf @nalimilan @quinnj

[quinnj]

PR to CSV.jl with an alternative approach that I think may be more all-around correct: JuliaData/CSV.jl#1058 (it links to 2 other issues in JuliaLang/julia & CSV.jl where I discuss recent findings/thoughts in more detail).

[bkamins]

After JuliaServices/ConcurrentUtilities.jl#12 discussion is finalized this issue should be resolved.

We need to ensure that both passed parameters and returned values in Task can be eventually GC'ed.
If it is not possible to GC returned values we need to redesign DataFrames.jl to avoid returning values from Task.

@quinnj (maybe @tkf or @vchuravy):

would just setting the result field in Task manually after it is processed be enough for the second problem?
What I mean is if after the line

DataFrames.jl/src/groupeddataframe/splitapplycombine.jl

Line 760 in 1b9d037

postprocessf = fetch(t)

we would add

t.result = nothing

would it solve the issue of return value?

[quinnj]

Yes, that would avoid holding references to the return value, as far as I understand.

[vchuravy]

Yes that would suffice, if you can guarantee that only one user will call fetch alternativly you could use a 1-element channel. Which would have the additional benefit of being potentially type-stable.