Make UnicodePlots an optional dependency

[andreasnoack]

This package is an indirect dependency of ours and UnicodePlots adds more than eight seconds to the load time of our package so it would be helpful if UnicodePlots could be made an weak dependency as described in https://pkgdocs.julialang.org/dev/creating-packages/#Transition-from-normal-dependency-to-extension

[chriselrod]

Using @time_imports, I get

  20394.7 ms  UnicodePlots
      2.1 ms  UnicodePlots → UnitfulExt
     19.5 ms  CategoricalDistributions

That's 20 seconds loading UnicodePlots. That dependency is not worth it.

[ablaom]

I'd support this change which makes sense.

We keep the first show and the second goes into the UnicodePlots glue code:

CategoricalDistributions.jl/src/methods.jl

Line 82 in 5c62c4e

Base.show(io::IO, mime::MIME"text/plain",

This the only place UnicodePlots is used.

I can't make this a priority just now, but can promise a timely review if someone wants to make a PR.

[t-bltg]

That's 20 seconds loading UnicodePlots.

This is weird, it should be around 4 s. On what OS is that ?

JuliaGraphics/ColorVectorSpace.jl#186 will also improve load times by reducing dependencies (merged, but not yet released).

[chriselrod]

Fedora, Linux.
In a fresh Julia session with --startup=no, I get

julia> @time using UnicodePlots
  2.875718 seconds (5.95 M allocations: 346.315 MiB, 11.14% gc time, 3.13% compilation time: 24% of which was recompilation)

When using my startup, it is similar (recompilation goes up as a percent of compilation time, but compilation time goes down from its already low value)

julia> @time using UnicodePlots
  2.868708 seconds (5.85 M allocations: 337.308 MiB, 10.91% gc time, 1.19% compilation time: 65% of which was recompilation)

The 20s comes from @time_imports when loading other packages that depend on CategoricalArrays.

[t-bltg]

Thanks, I'd be interested in reading suggestions on how to make the UnicodePlots experience smoother.

There is some latency involved by loading compiled code generated by precompile statements (tracked by JuliaPlots/UnicodePlots.jl#324, with no immediate solution (except disabling the precompilation workload with the effect of worsening TTFP)):

$ julia
julia> VERSION
v"1.9.0-beta4"
julia> using SnoopPrecompile, Preferences
julia> Preferences.set_preferences!(SnoopPrecompile, "skip_precompile" => ["UnicodePlots"]; force = true)
julia> exit(0)
$ julia -e 'using UnicodePlots'  # precompile
$ julia
julia> @time using UnicodePlots
  2.536887 seconds (3.35 M allocations: 230.944 MiB, 6.89% gc time, 8.91% compilation time: 17% of which was recompilation)

# previously, with precompile workload: 
  4.223118 seconds (5.61 M allocations: 368.600 MiB, 6.43% gc time, 4.18% compilation time: 22% of which was recompilation)

Some recent improvements (2s, 32%) have been made in JuliaPlots/UnicodePlots.jl#332 following JuliaPlots/UnicodePlots.jl#331 using package extensions - weak deps.

IMO the load time in the range [2s, 4s] is reasonable for a plotting package, but this is subjective (tradeoff for TTFP) and feedback would be appreciated.

[ablaom]

@t-bltg Thanks for the update. It's reassuring to know there is clearly a concerted effort to optimise. I don't have any suggestions on how to improve things there.

I think the case for making UnicodePlots a weak dependency of CategoricalDistributions remains strong. It's only there for pretty printing and CategoricalDistributions will sit pretty low in the package food chain.

[ablaom]

Note to self:

Current load time for UnicodePlots 3.4.0 is about 13 times that of Distributions 25.82 in a @time_imports using CategoricalDistributions benchmark. Julia 1.8.5.