Discrete x/y scales reserve space for unused limits

[wch]

For example, from the docs

d <- ggplot(subset(diamonds, carat > 1), aes(cut, clarity)) +
      geom_jitter()
d + scale_x_discrete(limits=c("Fair","Ideal"))

In the 0.9.3.1 docs, it looks like this:

[wch]

I've bisected it to 89893c7.

[hadley]

Probably related to #1589

[thomasp85]

This is two unrelated problems. This one is easily fixed whereas #1589 is a bit less straightforward.

@hadley can you by chance remember what you tried to achieve in 89893c7? The message is that it improves the defaults, but since it is quite broken it is not obvious what was wrong with the old.

[hadley]

The goal was to fix #1542 - ie when you accidentally use a discrete scale with only continuous data. (You have to be able to use a mix of cont and disc so that jittering and similar work)

[thomasp85]

I somehow missed that reference - sorry 'bout that

[thomasp85]

It seems your change in get_limits didn't anticipate that the function was used in other places than dimensions.

Was the change supposed to do avoid drawing all breaks when numeric ranges were converted to discrete?

[hadley]

Your guess is as good as mine. (But I don't think it was supposed to)

[thomasp85]

Also is it intentional that continuous data on a discrete scale does not receives any expansion?

In general can you explain the calls to expand_range? It seems that the continuous range only receives additive expansion and discrete ranges only receives multiplicative expansion - shouldn't both receive both?

[hadley]

I'm not sure I can - I've done some exploration recently and I think multiplicative looks better for both. I don't remember my original reasoning

[thomasp85]

Makes sense - additive is very dependent on the number of levels. But no matter what looks best, shouldn't it honor what is passed in through the expand argument instead of just silently ignoring some of the settings?

[hadley]

The goal was always to get an aesthetic pleasing amount of space between the data and the axes. In an ideal world, that would be the same for both x and y axes, but there's no way to do that.

You're right that we should respect both options - I'm not sure why I made that change.

For this problem, the fix is probably straightforward - what we need to make sure to capture is a few different motivating examples of mixing continuous and discrete values on a discrete axis.

[thomasp85]

I'll cook up a PR with a potential fix and let you stress test it as you seems to have a better idea about what is the intended functionality... (coding in the blind FTW :-)