Unexpected behavior in element_text() horizontal justification

[zlskidmore]

In the upcoming release of ggplot2 (2.3.0) the behavior of element_text() has changed. Specifically it looks like the justification anchors have changed in relation to the text. For Example:

In the current release:

In the development version:

mtcars$name <- rownames(mtcars)

ggplot(mtcars, aes(x=name, y=carb)) + geom_bar(stat="identity") + theme(axis.text.x=element_text(angle=60, hjust=1))

[karawoo]

Thanks for reporting this. I believe this bug is due to the change in how xp and yp are calculated in #2577.

[karawoo]

@hadley @clauswilke I think we should revert this change, it's going to break a lot of users' plots.

[clauswilke]

@karawoo I'll look into it. However, I'm not yet convinced about the cause of this problem, because I checked for this particular issue when I made the patch, if I remember correctly. I think this has actually to do with changed grob heights due to fixed descenders.

A workaround exists, but I agree it's not intuitive (why vjust = 0.65?)

library(ggplot2)
mtcars$name <- rownames(mtcars)

ggplot(mtcars, aes(x=name, y=carb)) + 
  geom_bar(stat="identity") + theme(axis.text.x=element_text(angle=60, hjust=1, vjust = 0.65))

Created on 2018-05-23 by the reprex package (v0.2.0).

[clauswilke]

When looking at this in debug mode, we see that for axis tick labels along the x axis the reference point travels vertically for different vjust values rather than orthogonally to the direction of the text. As far as I recall, this behavior is identical to the currently released ggplot2 version, but it's different from the behavior for hjust (which is along the text direction) and it's also different from hjust and vjust for axis titles, legend text and titles, and so on.

library(ggplot2)
mtcars$name <- rownames(mtcars)

ggplot(mtcars, aes(x=name, y=carb)) + 
  geom_bar(stat="identity") + 
  theme(axis.text.x=element_text(angle=60, hjust=1, vjust = 0, debug = TRUE))

ggplot(mtcars, aes(x=name, y=carb)) + 
  geom_bar(stat="identity") + 
  theme(axis.text.x=element_text(angle=60, hjust=1, vjust = 0.5, debug = TRUE))

ggplot(mtcars, aes(x=name, y=carb)) + 
  geom_bar(stat="identity") + 
  theme(axis.text.x=element_text(angle=60, hjust=1, vjust = 1, debug = TRUE))

There are two possible alternative ways to fix this:

• Make the reference point travel inside the range indicated by the colored square, so that vjust = 0 indicates the bottom of the square and vjust = 1 indicates the top.

• Make the reference point travel orthogonally to the text direction at all times.

[clauswilke]

I see what the problem is now. It arises in title_spec() when one position coordinate is set and one is not. I can solve things for this particular case by replacing these lines:

ggplot2/R/margins.R

Lines 47 to 48 in eecc450

	x <- x %||% unit(rep(just$hjust, n), "npc")
	y <- y %||% unit(rep(just$vjust, n), "npc")

with

  if (is.null(x)) {
    if (is.null(y)) {
      x <- unit(rep(just$hjust, n), "npc")
      y <- unit(rep(just$vjust, n), "npc")
    } else {
      x <- unit(rep(hjust, n), "npc")
    }
  } else if (is.null(y)) {
    y <- unit(rep(vjust, n), "npc")
  }

but that's not a general solution (it fails at other angles). Let me ponder this some more. I'd rather fix this in a forward-looking manner than going back to behavior that is broken in other ways.

[clauswilke]

I've been pondering this issue the whole afternoon and I'm not sure how to fix this. (We can revert, of course.) The bigger problem I'm seeing is that we are using two different justification systems throughout ggplot, and things break one way or another when we mix them (as is the case for axis tick labels): One system uses a fixed reference point and moves the text relative to that point. E.g., hjust = 0 means the entire string is to the left of that point, and hjust = 1 means the entire string is to the right. The other system uses a fixed extent and moves the string inside that extent. E.g., hjust = 0 means the string is at the left boundary of that extent, and hjust = 1 means it's on the right boundary. For x axis tick labels, we want the first justification system along the x axis and the second orthogonal to the x axis. But we also want to use hjust and vjust in the internal coordinate system of the text label, so that hjust = 1 right-justifies all the text labels regardless of angle. I haven't yet found a way bring all these conflicting constraints into one consistent framework that makes sense at all angles.

Since I don't even know what the correct behavior is in the general case, I just created a PR that reverts to the previous model specifically when the justification model is mixed: #2654

With this PR:

library(ggplot2)
mtcars$name <- rownames(mtcars)

ggplot(mtcars, aes(x=name, y=carb)) + 
  geom_bar(stat="identity") + theme(axis.text.x=element_text(angle=60, hjust=1))

Created on 2018-05-23 by the reprex package (v0.2.0).

[lock]

This old issue has been automatically locked. If you believe you have found a related problem, please file a new issue (with reprex) and link to this issue. https://reprex.tidyverse.org/