[BUG] Cycling arguments

[briochemc]

Plots.jl cycling arguments (but only sometimes) makes for a confusing experience. With Plots.jl v1.6.4, depending the backend,

scatter(1:10, 1:3)

either:

• plots by repeating 1:3 with, e.g., GR and PyPlot:
  

• plots only the first 3 points with, e.g., UnicodePlots:
  

• or errors with, e.g., Plotly:

  julia> plotly()
  Plots.PlotlyBackend()
  
  julia> scatter(1:10, 1:3)
  Error showing value of type Plots.Plot{Plots.PlotlyBackend}:
  ERROR: BoundsError: attempt to access 3-element UnitRange{Int64} at index [1:10]

I think it should either always error, or at least throw a warning.

Related issue: Note I read that

Plots is intended to cycle input arguments, this allows you to do things like scatter(1:4, [1]). But it's a good question why the backends are inconsistent! Plotlyjs has a third behaviour - it plots (1:100, 1:100). I hope you're OK with my renaming of the issue.

Originally posted by @mkborregaard in #1151 (comment)

But I think think this cycling of arguments is not a good idea. (My opinion being biased by having been sent into a 3-day rabbit-hole expedition because of that "feature".)

While I understand the appeal of being able to do things like scatter(1:4, [1]), I think it exposes the risk that someone (like me) mistakenly plots scatter(x,y) with x and y of different lengths, and cannot figure out what is happening. This is true especially for large-ish datasets that produce a somehow dense cloud of data. I would personally prefer the behavior to be only valid when calling a scalar, i.e. for scatter(1:4, 1) to actually do what scatter(1:4, [1]) currently does, but I think that may conflict with other things (e.g., what is discussed in #2129).

Backends

This bug occurs on ( insert x below )

Backend	yes	no	untested
gr (default)	x
pyplot	x
plotly		x
plotlyjs			x
pgfplotsx			x
inspectdr			x
unicodeplots	x

Versions

• Plots.jl version: v1.6.4
• Backend version (]st -m):
  • GR v0.52.0
  • PyPlot v2.9.0
  • UnicodePlots v1.3.0
• Output of versioninfo():

  julia> versioninfo()
  Julia Version 1.5.1
  Commit 697e782ab8 (2020-08-25 20:08 UTC)
  Platform Info:
    OS: macOS (x86_64-apple-darwin19.5.0)
    CPU: Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz
    WORD_SIZE: 64
    LIBM: libopenlibm
    LLVM: libLLVM-9.0.1 (ORCJIT, skylake)

---

Edit: Fixed the MWE (1:10 instead of 1:20)

[mkborregaard]

FWIW I've changed my mind on this, and agree it's unfortunate.

[BeastyBlacksmith]

The question is, what behaviour do we want? Personally, I like the zip-like behaviour of the unicode backend.

[briochemc]

I don't know what you will decide, but IMHO if it does not error, I really think it should throw a warning!

[mkborregaard]

I don't know what the unicode backend does - but wouldn't it be most consistent to error but have a recipe for scatter(1:10, Iterators.cycle(1:3))?

[briochemc]

I don't know what the unicode backend does - but wouldn't it be most consistent to error but have a recipe for scatter(1:10, Iterators.cycle(1:3))?

I showed what the output of UnicodePlots up there! With it, scatter(1:10, 1:3) plots the first 3 points only (the "missing" points are not plotted, but the xlim covers (1,10)). Here it is again:

[mkborregaard]

Oh sorry! Yeah I can't think of a situation when that would be the desired behaviour.

[yha]

I don't know what the unicode backend does - but wouldn't it be most consistent to error but have a recipe for scatter(1:10, Iterators.cycle(1:3))?

I agree that erroring would be best, and if we do this breaking change, I would extend it to other cases where we currently cycle, which seems to be some pseudo-random backend-dependent subset of plots attributes taking vector values. Would be nice if we can abolish Plots._cycle.
Maybe we can directly support arbitrary iterables, so things like scatter(1:10, Iterators.cycle(1:3)) can work?
I think these look quite nice:

using Iterators: repeated, cycle
plot(cycle([0,1]), y)
scatter(x, repeated(0))
plot(y, color=cycle(1,2), fillrange=repeated(0))
# each column against the same vector:
plot(repeated(x), eachcol(Y))
# or maybe
plot(repeated(x), Y)

I think this can work by zipping and collecting the positional inputs, and only then proceeding to dispatch by types according to the recipe system. Just need to error early if none of the inputs has known finite length (based on Base.IteratorSize).

[t-bltg]

This weird "un-official, un-documented, scatter only" cyclic behavior should be fixed removed in 2.0.

As of the scatter(1:4, [1]) thing, supporting Iterators from Base in a recipe as suggested in #2980 (comment) would be acceptable to me.