add smart fillto to enable logscale y-axis histogram barplot

[Moelf]

Fix #3000

This implements what's suggested on Slack, namely, detect if transformation is one of the dangerous logs (example of non-dangerous log is Makie.pseudolog10), if so, use the minimal non zero value divided by 2 as the "bottom" to avoid illegal transformation 0 -> -Inf when applying the log.

Some examples that used to error but now works:

using CairoMakie

a = "/tmp/jl_k9M7ELsL2h.png"

save(a, hist(randn(10^6); axis=(; yscale=log10)))

save(a, hist(randn(10^6); axis=(; yscale=log2)))

save(a, hist(filter!(x-> x<0 || x > 1.5, randn(10^6)); axis=(; yscale=log10)))

[Moelf]

(should maybe also add a test?)

[asinghvi17]

Ideally, test + example in the barplot page. The test can be added here:

https://github.com/Moelf/Makie.jl/blob/log_yscale_barplot_smart_fillto/ReferenceTests/src/tests/examples2d.jl

[Moelf]

when you say example do you mean in the docs? I don't think we need that right?

[jkrumbiegel]

example do you mean in the docs? I don't think we need that right?

yes a docs example would be good because we want to explain every single attribute visually in the long run. And the visual effect of fillto = automatic would have to be shown

[jkrumbiegel]

we could also think about a drop_zeros attribute which would remove the bars at zero. This makes no difference for linear scale because you can't see them anyway, but for log scale it would make an otherwise invalid plot valid.

[Moelf]

I would say drop zero is not needed because in both cases it should be invisible, that seems to be the only reasonable behavior

[jkrumbiegel]

how would a value of 0 be invisible in the log scale case? Or do you mean you would always drop zeros so no setting needed

[Moelf]

See the third example in op. I'm saying we always drop zero and it's already implemented, but maybe there's a better way to implement it

[Moelf]

the last commit makes x-scale work too:

barplot(
           tbl.x, tbl.height,
           dodge = tbl.grp,
           color = tbl.grp,
           bar_labels = :y,
           axis = (xticks = (1:3, ["left", "middle", "right"]),
                   title = "Dodged bars horizontal with labels", xscale=log10), # <---------------------------
           colormap = [:red, :green, :blue],
           color_over_background=:red,
           color_over_bar=:white,
           flip_labels_at=0.85,
           direction=:x,
       );

this used to error, but now:

[Moelf]

is this ready? I don't know what example I should add, fillto itself is documented in the barplot page pretty well

[Moelf]

this seems ready 👀

[SimonDanisch]

@jkrumbiegel wanted to do a final review before merging

[Moelf]

franklin error seems unrelated

[SimonDanisch]

Hm, a new Franklin version was released just now: https://github.com/tlienart/Franklin.jl/releases/tag/v0.10.86
I guess that's most likely the cause?

[Moelf]

yeah I think so, this PR didn't touch the the slide grid markdown

[Moelf]

Looks like it Passed?

[jkrumbiegel]

Made some small changes to the test to exercise the x direction code, too. I think this is good to go now. I haven't tested out all combinations of switching back and forth in GLMakie, but it looked like we easily get limit errors that way which are not really related to the recipe, rather to the limit update code and when what value is updated. So not relevant here.

There should be a docs example added at some point, but before I spend time doing that in the current format, I'd rather wait until we switch the plotting functions to direct attribute docs like we have for Blocks. Shouldn't block this right now.