Undocumented change: preservation of class attribute

[heavywatal]

The class attribute is lost in the released 0.7.8, and preserved in the dev version:

x = as_tibble(iris)
class(x) = c("myclass", class(x))
x %>% dplyr::filter(Sepal.Length < 5) %>% class()

# 0.7.8
# [1] "tbl_df"     "tbl"        "data.frame"

# 0.7.99.9000
# [1] "myclass"    "tbl_df"     "tbl"        "data.frame"

It can be a side effect of the changes in grouped_df (or slice()), and happens with the stable tibble 1.4.2 and other verbs like mutate(). Anyway it is favorable to me and seems to be consistent with the upcoming tibble with subclass support. I just hope it is documented more explicitly so that I can give a clear direction to issues like GuangchuangYu/tidytree#7.

[elinw]

This is a breaking change for skimr because it means that the print function no longer is print.data.frame() but rather print.myclass() so the output needs to be piped to as.data.frame() in order to avoid an error. Thanks @romainfrancois for alerting us to the breakage. I feel this is an extremely significant change that, if kept, should be documented in the list of major breaking changes. I'm actually not sure what to do that would work with both versions except for overloading the impacted dplyr verbs.

[elinw]

Based on https://github.com/tidyverse/dplyr/issues/4051 I guess we should do mutate also; it's just not in any of our examples or vignettes. Are you aware of this being an issue for other verbs?

[yutannihilation]

Is the preservation of class really an intended change? If so, I'll file the below as a new issue; for regrouping purpose, group_by(..., add = TRUE) might be better replaced because group_by() doesn't respect the original classes.

library(dplyr, warn.conflicts = FALSE)

x <- as_tibble(iris)
class(x) <- c("myclass", class(x))
x %>%
  add_count(Species) %>%
  class()
#> [1] "tbl_df"     "tbl"        "data.frame"

^{Created on 2019-01-07 by the reprex package (v0.2.1)}

[hadley]

Yes, it's definitely intended although it won't be until dplyr 0.9.0 (where we have full vctrs integration) that we expect every function to do the right thing. The behaviour of add_count() etc is going to need quite a lot of thought, so an issue would be helpful.

[yutannihilation]

Thanks, it's really a good news to me then :) I'll file a new issue.