Attempting to use Distributed and consistently getting a MethodError on sorted_deppairs

[jmuchovej]

When attempting to solve many POMDPs in parallel (each process is a POMDP-solver pair, no communication between pairs is needed), I consistently run in to an issue where sorted_deppairs cannot be dispatched, yet when I run the same functions on the main threads, everything proceeds accordingly.

The error, in brief:

On worker 6:
MethodError: no method matching sorted_deppairs(::Type{BoxWorld}, ::NTuple{4, Symbol})
Closest candidates are:
  sorted_deppairs(::Type{<:POMDP}, ::Any) at [~/.local/conda/lib/julia/packages/POMDPs/XBTe5/src/gen_impl.jl:66](https://file+.vscode-resource.vscode-cdn.net/Users/john/10-projects/10.01-agent-programs/agent-programs/_research/~/.local/conda/lib/julia/packages/POMDPs/XBTe5/src/gen_impl.jl:66)
  sorted_deppairs(::Type{<:MDP}, ::Any) at [~/.local/conda/lib/julia/packages/POMDPs/XBTe5/src/gen_impl.jl:56](https://file+.vscode-resource.vscode-cdn.net/Users/john/10-projects/10.01-agent-programs/agent-programs/_research/~/.local/conda/lib/julia/packages/POMDPs/XBTe5/src/gen_impl.jl:56)
  sorted_deppairs(::Dict{Symbol, Vector{Symbol}}, ::Any) at [~/.local/conda/lib/julia/packages/POMDPs/XBTe5/src/gen_impl.jl:77](https://file+.vscode-resource.vscode-cdn.net/Users/john/10-projects/10.01-agent-programs/agent-programs/_research/~/.local/conda/lib/julia/packages/POMDPs/XBTe5/src/gen_impl.jl:77)
Stacktrace:
...

BoxWorld is a POMDP, as evidence:

       From worker 7:	typeof(pomdp) = BoxWorld
       From worker 7:	typeof(pomdp) <: POMDP = true
       From worker 8:	typeof(pomdp) = BoxWorld
       From worker 8:	typeof(pomdp) <: POMDP = true
       From worker 4:	typeof(pomdp) = BoxWorld
       From worker 4:	typeof(pomdp) <: POMDP = true
       From worker 2:	typeof(pomdp) = BoxWorld
       From worker 2:	typeof(pomdp) <: POMDP = true
       From worker 3:	typeof(pomdp) = BoxWorld
       From worker 3:	typeof(pomdp) <: POMDP = true
       From worker 5:	typeof(pomdp) = BoxWorld
       From worker 5:	typeof(pomdp) <: POMDP = true
+      From worker 6:	typeof(pomdp) = BoxWorld
+      From worker 6:	typeof(pomdp) <: POMDP = true
       From worker 9:	typeof(pomdp) = BoxWorld
       From worker 9:	typeof(pomdp) <: POMDP = true

What's most perplexing is that if I don't use Distributed, everything runs just fine.

Potential causes for error: I'm using a Jupyter Notebook through VSCode, but I've also replicated the error on Jupyter Lab and in a shell script.

do_null_program_sim does some extra processing before running do_pomdp_sim, which is:

function do_pomdp_sim(
    seed::Int, items::Vector, boxes::Vector{Box}, target
)
    # @show seed
    Random.seed!(seed)

    program = @agent items begin end

    solver = QMDPSolver(;
        max_iterations=100,
        belres=1e-3,
        verbose=false
    )

    pomdp = FruitWorld(
        program;
        boxes=boxes,
        items=items,
        target=target
    )

    @show typeof(pomdp)
    @show typeof(pomdp) <: POMDP

    policy = solve(solver, pomdp)

    hr = HistoryRecorder(; max_steps=30)
    hist = simulate(hr, pomdp, policy)

    return (seed=seed, pomdp=pomdp, policy=policy, hist=hist)
end

FruitWorld and BoxWorld are synonymous (they need to be differently named for other reasons).

[zsunberg]

Well, this is perplexing and frustrating!

I don't know why the dispatch isn't working. It probably has something to do with genout (

POMDPs.jl/src/gen_impl.jl

Line 19 in 851535f

@generated function genout(v::DDNOut{symbols}, m::Union{MDP,POMDP}, s, a, rng) where symbols

) being a @generated function

[zsunberg]

If you are really wanting to get something working as fast as possible, you could try make a specific implementation of POMDPs.genout for your problem.

[zsunberg]

@jmuchovej I tried to reproduce the error with this code:

procs = addprocs(2)
@everywhere using POMDPs, Random
@everywhere begin
    struct DistributedDeterministic{T}
        x::T
    end
    Base.rand(rng::AbstractRNG, d::DistributedDeterministic) = d.x

    struct DistributedTestMDP <: MDP{Int, Int} end
    POMDPs.transition(::DistributedTestMDP, s, a) = DistributedDeterministic(s+a)
    m = DistributedTestMDP()
    @gen(:sp)(m, 1, 1)
end
rmprocs(procs)

Unfortunately, this code executed without error. Can you either share all your code or help me find the differences between this minimal example and your code that results in an error.

[zsunberg]

Also it's probably worth checking which version of julia you are using and trying it with the latest just in case this is a bug in julia that has been fixed.

[jmuchovej]

Thanks for testing this out! 🙂

While looking into this issue – it seems like another library I've been using might be causing a problem (how still remains unclear).

I'm using DrWatson to do stuff like ensure we're using the right Julia project environment and the like, but for whatever reason when it's paired with Distributed things start "acting out".

• Odd behavior with Distributed JuliaDynamics/DrWatson.jl#269
• Usage together with Distributed can cause precompile issues JuliaDynamics/DrWatson.jl#270

Definitely not a satisfying answer, but once I swapped out DrWatson for traditional Julia Pkg.activate(...) commands, everything seems to be working correctly, so I'll go ahead and mark this as answered and hopefully this is helpful to anyone else trying to use the combo. 😅