grouped_df should not require a ".group" variable for data on the other layers

[yutannihilation]

Due to the change in response to #2351, a grouped_df now creates a variable called .group. This may be problematic because ggplot2 treats the first layer so special as all the following layers are required to contain all mapped variables except when another mapping is specified individually. Here is an example:

reprex::reprex_info()
#> Created by the reprex package v0.1.1.9000 on 2017-12-22

library(dplyr, warn.conflicts = FALSE)
library(ggplot2)

df <- data.frame(
  x = 1:10,
  y = 1:10,
  type = rep(c("a", "b"), 5),
  stringsAsFactors = FALSE
) %>%
  group_by(type)

df_labels <- data.frame(
  x = 1,
  y = 2,
  text = "foo!"
)

ggplot(df, aes(x, y)) +
  geom_point(aes(colour = type)) +
  geom_label(data = df_labels, aes(label = text))
#> Error in FUN(X[[i]], ...): object '.group' not found

ggplot(ungroup(df), aes(x, y)) +
  geom_point(aes(colour = type)) +
  geom_label(data = df_labels, aes(label = text))

Though I'm not sure this is just a bug, I'm afraid #2351 needs a bit more considerations (better error messages, at least).

[mjskay]

What about some opt-in syntax for this? Since it implicitly adds a .group column now, one could simply not also set the group aesthetic and allow users to do this to "opt-in":

ggplot(some_grouped_df, aes(..., group = .group)) +
  ...

Which would avoid breaking existing code that has been passing grouped data frames into ggplot, and make the use of the grouping explicit rather than implicit. (I'll admit bias towards the explicit approach as this issue came to my attention by simultaneously breaking a bunch of my plots).

[hadley]

Yeah, this has a much bigger impact that I had thought. I like @mjskay's suggestion to make it more opt-in.

[yutannihilation]

I like the idea! For the notation, I feel it's good if more explicit one like calc() is introduced. .group seems a new ...var..., I'm a bit afraid it would be hard to find what it is and how it is special.

[yutannihilation]

But, it's far better than the status quo. I'm OK with explicit .group (at least, for now).

[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/