Scaling of legends when composing plots into layout, v1.25

[mcabbott]

Something has changed with the scaling of subplots within a layout, which causes the legend to become very wide.

Previously, they retained the size they had within individual plots. Now, the legend within one subplot seems to depend on the size of the overall layout. Especially for scatter plots, this causes the point to be marooned out in space, far from the label. Nothing else (that I know of) scales like this -- other features of a plot remain the same when it is placed into a layout.

Example:

julia> using Plots  # with GR default

julia> p1 = plot(cumsum(rand(30, 5), dims=1); size=(300,400)) # normal short legend lines

julia> plot(p1, p1, p1, p1; size=(1200, 400), layout=(1,4))  # much longer lines in legend

julia> plot!(title="Plots v1.25.7")

Individual plots do not seem to be as strongly affected:

[Edit - subplot sizes corrected so that they are exactly a multiple of the layout size.]

Details

Backends

This bug occurs on ( insert x below )

Backend	yes	no	untested
gr (default)	X
pyplot			X
plotlyjs		X
pgfplotsx			X
unicodeplots			X
inspectdr			X
gaston			X

Versions

(jl_A9O98Y) pkg> st
Status `/private/var/folders/yq/4p2zwd614y59gszh7y9ypyhh0000gn/T/jl_A9O98Y/Project.toml`
  [28b8d3ca] GR v0.63.1
  [58dd65bb] Plotly v0.4.1
  [a03496cd] PlotlyBase v0.8.18
  [91a5bcdd] Plots v1.25.7

julia> versioninfo()
Julia Version 1.8.0-DEV.1440
Commit e3b681c687 (2022-02-02 01:39 UTC)
Platform Info:
  OS: macOS (arm64-apple-darwin21.1.0)
  CPU: Apple M1
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-13.0.0 (ORCJIT, cyclone)
Environment:
  JULIA_NUM_THREADS = 4

julia> versioninfo()  # same results with rosetta & Julia 1.7
Julia Version 1.7.0
Commit 3bf9d17731 (2021-11-30 12:12 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin19.5.0)
  CPU: Apple M1
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, westmere)

[t-bltg]

This is an effect of #3855.

In the mentioned PR, I expressed my opinion (specifically here) against hard-coding a scale factor when a dynamic function of the plot area was implemented before.

Those eventually got merged in #4055.

I would vote in favor of reverting these changes.

Example output with the PR reverted:

Ping @BeastyBlacksmith and @isentropic for decision.

[mcabbott]

Thanks for digging. One further change which this doesn't revert is that the y-axis ticks are calculated differently. Not sure which is better (my real case fixed them by hand).

[t-bltg]

One further change which this doesn't revert is that the y-axis ticks are calculated differently

Oh I see that only now. That's quite difficult to track down, but I'll try to find out what changed this.
Some significant changes were made in JuliaPlots/PlotUtils.jl#124, which got in after 1.24.3 got out, maybe you can try reverting this PR and see if it fixes the ticks.

[BeastyBlacksmith]

To be honest, I don't like both versions for this example. But I agree, that we should have a more dynamic default and maybe also a way to customize this.

[mcabbott]

Maybe the scatter example is closer to what initially surprised me, as the points seem marooned out in space. (I had turned the legend box off too.)

[isentropic]

I also think there might be a better solution than both, but the one now looks better with default figure size (unlike this example), do you agree?

[mcabbott]

I see no difference at the default figure size:

[isentropic]

Yes that's intentional, how about 4 subplots in one figure, with default figure sizing (600x400?)

[isentropic]

The new version works better

[mcabbott]

4 plots within the default seems a very crowded case no matter what, I guess I would vote for the legend being pretty tight here rather than taking up too much of the graph?

Super-hard to tune all cases. But inserting so much more space between the symbol and the label, compared to that between different symbols, seems to break up wha the legend is trying to convey.

But #4092 (comment) is close to what I have in reality -- small plots which work fine alone, at the intended size, but change quite dramatically when placed into a larger plot, still at the same size.

That's the coupling which seems odd to me. Why should the legend of one plot care whether it's a small stand-alone plot or a panel (the same size) in a larger one?

I know nothing about the internals of this package but assumed this was tied to the font size or the size of the plot points used in the legend, and perhaps to the plot the legend belongs to, but not to the outermost layout it is placed into.

[isentropic]

In the 4 plot example above, the font size doesn't scale, the tick marks (little marks above ticks) don't scale, title font doesn't scale, legend font doesn't scale, size of the markers in the legend doesn't scale. Why should the legend box scale? It just does not make sense.
You plotted a scatter, which looks okay in the older version, but line plot doesn't look as well. Next, I was thinking about backend consistency, new GR is also on par with pyplot and pgfpltosx.

[isentropic]

At this point it's a debate of tastes I feel. My arguments are:

1. Stay consistent with other backends
2. Stay consistent with other plot elements

In my opinion, the legendbox should always keep its aspect ratio (defined by the legendbox fontsize), regardless of the figure size. Perhaps this hasn't been achieved yet.

[mcabbott]

Why should the legend box scale?

To be clear, the original surprise above is that the legend box does now scale, when subplots are combined into a larger plot. While the fonts, the markers, etc, do not.

I don't know why that's happening. Perhaps "keep its aspect ratio" is being defined relative to the whole layout, rather than to the axes on which the legend is located.