Adding scale_*_discrete() to date variable can crash R

[cpsievert]

With scales 0.3, this code will error:

x <- c(Sys.Date(), Sys.Date() - 1e4)
y <- c(3, 4)
d <- data.frame(x, y)
library(ggplot2)
ggplot(d, aes(x, y)) + geom_line() + scale_x_discrete()
#> Error: Discrete value supplied to continuous scale

Running this code using the latest dev version of scales doesn't result in error, and ggplot tries to draw a plot with 1e4 ticks. I know this is a cornercase, but I would prefer an error since it can easily wreak havoc on an R session, especially with date-times.

[cpsievert]

Here is the relevant change -- r-lib/scales@cd0395d

[hadley]

Are you sure that's the change? It seems weird that a check in continuous scales would break discrete scales

[cpsievert]

Yes, it appears both scale_x_discrete()/scale_y_discrete() use a continuous range. I'm not sure whether or not that is intentional, but it does seem odd:

> scale_x_discrete
function (..., expand = waiver()) 
{
    sc <- discrete_scale(c("x", "xmin", "xmax", "xend"), "position_d", 
        identity, ..., expand = expand, guide = "none")
    sc$super <- ScaleDiscretePosition
    class(sc) <- class(ScaleDiscretePosition)
    sc$range_c <- continuous_range()
    sc
}

Just to demonstrate that the change I refer to above is relevant, you can install r-lib/scales#70 to see that the example now throws an error.

[hadley]

Ah, that makes sense. But I'd rather fix it by handing the poor performance with many levels

[hadley]

Hmmm, something is seriously messed up with code that handles the combination of discrete and continuous ranges. I need to think through exactly how this should behave - the reason why discrete scale also have a continuous component is so that this sort of plot can work:

ggplot(mtcars, aes(factor(cyl), mpg)) + geom_jitter()