FWIW this happens in dplyr 1.0.10 and looks to have happened since at least 2020, so I don't consider this an immediate regression.

And I'm not 100% it is a bug. ungroup() is essentially the same as calling as_tibble(<grouped_df>), and we haven't decided yet if that is an operation that should drop "unknown" attributes or not. See also tidyverse/tibble#769

# pak::pak("tidyverse/[email protected]")
library(dplyr, warn.conflicts = FALSE)

d = mtcars
attr(d, "foo") = "bar"
attr(d, "foo")
#> [1] "bar"

# kept after group-by
d |> group_by(cyl) |> attr("foo")
#> [1] "bar"

# dropped here
d |> group_by(cyl) |> ungroup() |> attr("foo")
#> NULL

# really the `ungroup()` case is because of
d |> group_by(cyl) |> as_tibble() |> attr("foo")
#> NULL

# which looks like
dplyr:::as_tibble.grouped_df
#> function (x, ...) 
#> {
#>     new_tibble(dplyr_vec_data(x), nrow = nrow(x))
#> }
#> <bytecode: 0x7fa953f20e68>
#> <environment: namespace:dplyr>

^{Created on 2023-03-16 with reprex v2.0.2.9000}

Interesting, I would have thought count(d, ...) would also drop the attributes then too. I guess because it is using dplyr::reconstruct() instead of ungroup()?

even though tally(d) drops the attributes.

Looking at tidyverse/tibble#769, while I understand the reasoning I don't think it's intuitive for ungroup() to drop attributes, because the ungroup() operation does not necessarily mean a structural change to the tibble or accompanying attributes. group()/ungroup() operations can be used for mutate() statements to do within/group calculations without changing the row structure of the tibble. So regardless of the the discussion in tidyverse/tibble#769, I don't think that d |> group_by(cyl) |> as_tibble() is conceptually equivalent to d |> group_by(cyl) |> ungroup() (even if the code is equivalent right now) as I'm not necessarily trying to return to a "bare" tibble when ungrouping.

Another test case for consideration: partial ungrouping does not drop the attribute:

library(dplyr)

d = mtcars
attr(d, "foo") = "bar"

d |> group_by(cyl, am) |> ungroup(cyl) |>attr("foo")
#> [1] "bar"

d |> group_by(cyl, am) |> ungroup(am) |>attr("foo")
#> [1] "bar"

d |> group_by(cyl, am) |> ungroup(am, cyl) |> attr("foo")
#> NULL

d |> group_by(cyl, am) |> ungroup(am) |> ungroup(cyl) |> attr("foo")
#> NULL

nor does ungrouping an ungrouped data frame:

d |> ungroup() |> attr("foo")
#> [1] "bar"

Which feels funny to me, since the treatment of attributes differs based on if I'm dropping some groups or all groups. The core of the matter is that as a user I would generally expect the following to be TRUE:

identical(
    d |> group_by(...) |> ungroup(),
    d
)

Relacing the as_tibble() call with new_tibble(groups = NULL) maintains attributes and kicks the can down the road until a decision is made on tidyverse/tibble#769, so

becomes

ungroup.grouped_df <- function(x, ...) { 
   if (missing(...)) { 
     new_tibble(x, groups = NULL)

(and the same change would be needed for ungroup.rowwise_df()).

Test:

d = as_tibble(mtcars)
attr(d, "foo") = "bar"

d |> group_by(cyl, am) |>tibble::new_tibble(groups = NULL) |> attr("foo")
#> [1] "bar"

identical(
  d |> group_by(cyl, am) |>tibble::new_tibble(groups = NULL),
  d
)
#> [1] TRUE