@rikhuijzer i realised i don't really know what i'm doing 😅 can you take a look at adding more precompiles for our dynamic things?

@rikhuijzer i realised i don't really know what i'm doing sweat_smile can you take a look at adding more precompiles for our dynamic things?

Yeah sure. Gladly. Can you tell give in code or pseudocode are list of methods are called? Or, I can just see what the firebasey sends to the backend and trigger those?

Firebasey.diff gets called on two outputs of notebook_to_js, and it recursively calls itself on their children to compare. Run the Firebasey notebook to see what I mean

I've been going deeper on Pluto.run since we should be able to shave some time off there. So, similarly to #1934, I've used SnoopCompile to figure out where the time is spent. That gives this flamegraph:

I've obtained this by running the following notebook:

begin
    using Pkg
    
    Pkg.activate(; temp=true)
    Pkg.add([
        "SnoopCompile",
        "ProfileSVG"
    ])
    Pkg.develop(; path="/home/rik/git/Pluto.jl")
end

begin
    using SnoopCompile
    using ProfileSVG
    using Pluto
end

tinf = @snoopi_deep begin
    port = 13435
    options = Pluto.Configuration.from_flat_kwargs(; port, launch_browser=false, workspace_use_distributed=true, require_secret_for_access=false, require_secret_for_open_links=false)
    🍭 = Pluto.ServerSession(; options)
    server_task = @async Pluto.run(🍭)
    
    retry(; delays=ExponentialBackOff(n=5, first_delay=0.5)) do
        HTTP.get("http://localhost:$port/edit").status == 200
    end
    server_task
end

let
     g = flamegraph(tinf)
     ProfileSVG.view(g)
end

This flamegraph is good news. All inference is rooted on the same block, so if we can get the right precompile statement somewhere on from_flat_kwargs, then we can possibly get some proper improvements. I'll fiddle with it and report back here once I got it.

EDIT: precompile(from_flat_kwargs, ()) is great

Without precompile

$ julia --project -ie 'using Pluto'
...

julia> @time @eval Pluto.Configuration.from_flat_kwargs();
  2.174783 seconds (3.60 M allocations: 194.278 MiB, 8.07% gc time, 99.96% compilation time)

With precompile

julia> @time @eval Pluto.Configuration.from_flat_kwargs()
  1.395046 seconds (2.64 M allocations: 143.641 MiB, 12.13% gc time, 99.92% compilation time)

This takes roughly 0.8 seconds off the time for Pluto.run on my system. I'll add it to #1983. I have hard time measuring this via @time @eval Pluto.run() because via that call I have to pass keyword arguments.

As another possibility, maybe we should just build PackageCompiler.jl into Pluto? Like,

julia> using Pluto

julia> Pluto.compile()

julia> Pluto.run()

to have a super speedy Pluto. Only need to call Pluto.compile once after updating Pluto. And to have the normal Pluto, don't call Pluto.compile() after updating Pluto.

julia> using Pluto

julia> Pluto.run()

EDIT: Nope. Let's hope for JuliaLang/julia PR 44527 and the successors to be merged in newer Julia versions.

With JET.jl, I've been digging down in why SessionActions.open keeps taking > 10 seconds even though precompile is called on it. I've found one example of a problem. The compiler has a really bad time with maybe_macroexpand:

julia> using Pluto

julia> precompile(Pluto.ExpressionExplorer.maybe_macroexpand, (Expr,))
true

julia> @time @eval Pluto.ExpressionExplorer.maybe_macroexpand(:(x = 1))
  0.213755 seconds (397.54 k allocations: 20.237 MiB, 31.72% gc time, 100.06% compilation time)

julia> @time @eval Pluto.ExpressionExplorer.maybe_macroexpand(:(x = 1))
  0.000607 seconds (75 allocations: 3.484 KiB)

So that's more than 100% compilation time 🤣 😬 JET shows 120 lines of runtime dispatches and optimization failures.

EDIT: Fixed most of it. Got this now in a fresh session by fixing some type inference issues:

julia> using Pluto

julia> @time @eval Pluto.ExpressionExplorer.maybe_macroexpand(:(@time 1))
  0.084008 seconds (86.97 k allocations: 4.649 MiB, 99.64% compilation time)

PR at #1991.

AMAZING 🌟🌟