Annotations with zero-length aesthetics are broken

[paleolimbot]

In the revdep checks (#3303), an issue was identified where zero-length aesthetics in annotate() caused newly broken plots (interflex package).

library(ggplot2)
ggplot() +
  annotate("point", x = numeric(0), y = numeric(0), colour = "red")
#> Error: Elements must equal the number of rows or 1

In 3.1.1, this did not error:

library(ggplot2)
ggplot() +
  annotate("point", x = numeric(0), y = numeric(0), colour = "red")

[clauswilke]

This and possibly some of the other problems could be solved if new_data_frame() returned an empty data frame in case some columns have length zero. I think that's reasonable behavior. See examples.

library(ggplot2)

# empty data frame creates empty plot
df <- data.frame(x = numeric(0), y = numeric(0))
ggplot(df, aes(x, y)) + geom_point()

# data.frame() throws an error in this case; this is what we're currently mirroring
data.frame(x = numeric(0), y = numeric(0), z = "red")
#> Error in data.frame(x = numeric(0), y = numeric(0), z = "red"): arguments imply differing number of rows: 0, 1

# however, tibble() does not throw an error; I think this behavior is reasonable
tibble::tibble(x = numeric(0), y = numeric(0), z = "red")
#> # A tibble: 0 x 3
#> # … with 3 variables: x <dbl>, y <dbl>, z <chr>

^{Created on 2019-05-08 by the reprex package (v0.2.1)}

@thomasp85 Any concerns about the proposed solution?

[thomasp85]

No concerns in general, other than the possibility of hard-to-debug bugs

[clauswilke]

I'm a big proponent of functions that behave gracefully when they're given empty input, and I find that many frustrating bugs are actually caused by functions that insist they receive non-empty input. If I remember correctly, @lionel- once quoted some fancy LISP principle to this effect, but I don't remember the details.

[thomasp85]

I agree, but as I understand it this is not about zero input. This is about multiple inputs among which some are zero, right?

[clauswilke]

It's about multiple inputs when some are zero and others are constants (length 1). If some are zero and others are length > 1 you still get an error. This mirrors exactly what tibble is doing. (See also here: https://tibble.tidyverse.org/articles/tibble.html#recycling)

ggplot2:::data_frame(x = numeric(0), color = character(0))
#> [1] x     color
#> <0 rows> (or 0-length row.names)
ggplot2:::data_frame(x = numeric(0), y = 1, color = "red")
#> [1] x     y     color
#> <0 rows> (or 0-length row.names)
ggplot2:::data_frame(x = numeric(0), y = 1, color = c("red", "blue"))
#> Error: Elements must equal the number of rows or 1

tibble::tibble(x = numeric(0), color = character(0))
#> # A tibble: 0 x 2
#> # … with 2 variables: x <dbl>, color <chr>
tibble::tibble(x = numeric(0), y = 1, color = "red")
#> # A tibble: 0 x 3
#> # … with 3 variables: x <dbl>, y <dbl>, color <chr>
tibble::tibble(x = numeric(0), y = 1, color = c("red", "blue"))
#> Tibble columns must have consistent lengths, only values of length one are recycled:
#> * Length 0: Column `x`
#> * Length 2: Column `color`

^{Created on 2019-05-11 by the reprex package (v0.2.1)}

[lionel-]

once quoted some fancy LISP principle to this effect, but I don't remember the details.

Was it nil punning? That's a bit different from recycling to zero (which might be viewed as empty punning). Aren't lisp principles quaint rather than fancy though ;)

I agree the issue seems to be about recycling to zero.

[thomasp85]

As I said, I don’t have a big stake in this and we can change how it works if that makes everything else work. I was just foreseeing that this would lead to some hard-to-debug errors at some point🙂

[clauswilke]

Lionel: Yes, that's what I meant.

Thomas: Could you take a look at the actual PR to make sure it's Ok and doesn't degrade performance? #3320

[thomasp85]

There shouldn’t be any meaningful degradation of performance, AFAIK