Use vctrs::vec_restore to preserve data frame subclass and attributes

[krlmlr]

to get rid of redundant tbl_df methods which complicate navigating the code. When sloop is ready, we'll be ready to switch.

[markvanderloo]

It would be great if dplyr (and other tidyverse) verbs can keep user-defined attributes. Obviously, (as @hadley noted in #3558) you can't keep row.names in a filter. Let me explain one use case.

In my lumberjack package, I define a new pipe operator that can track changes in data in a custom way. The pseudocode is like this:

`%>>%` <- function(x,f){
  old <- x
  new <- f(x)
  if ( x has a logger attribute ) attr(x, "logger")(old, new)

  return(new)
}

I have other use cases as well. What I'm saying is that attributes are not necessarily row-wise or column wise.

I understand that the whole concept of the tidyverse is to be a type-safe zone within R. Keeping user-added attributes does not make the output of tidyverse verbs inconsistent with the tidyverse, except in case of name collisions but those can be handled e.g. with a warning or message. Being consistent with the tidyverse is the responsibilty of package authors on the fringe of it (like me).

At the moment, you are creating an 'inconsistent' object, in interaction with at least one package outside of the tidyverse (mine). I don't think that's what you'd want either.

[romainfrancois]

Also relevant: #1064. Not sure why we keep attributes for some cases.

test_that("filter, slice and arrange preserves attributes (#1064)", {
  df <- structure(
    data.frame(x = 1:10, g1 = rep(1:2, each = 5), g2 = rep(1:5, 2)),
    meta = "this is important"
  )
  res <- filter(df, x < 5) %>% attr("meta")
  expect_equal(res, "this is important")

  res <- filter(df, x < 5, x > 4) %>% attr("meta")
  expect_equal(res, "this is important")

  res <- df %>% slice(1:50) %>% attr("meta")
  expect_equal(res, "this is important")

  res <- df %>% arrange(x) %>% attr("meta")
  expect_equal(res, "this is important")

  res <- df %>% summarise(n()) %>% attr("meta")
  expect_equal(res, "this is important")

  res <- df %>% group_by(g1) %>% summarise(n()) %>% attr("meta")
  expect_equal(res, "this is important")

  res <- df %>% group_by(g1, g2) %>% summarise(n()) %>% attr("meta")
  expect_equal(res, "this is important")
})

[markvanderloo]

I just figured out, while walking to the eRum conference venue, that in my use case, I can just re-add the attribute I need in the %L>% operator :-). Anyway, it would still be useful to be able to somehow keep attributes when passing data through dplyr verbs, but I understand something is in te works. @romainfrancois: cheers for the (live) discussions 👍

[hadley]

See discussion of problem, and solution, in https://adv-r.hadley.nz/s3.html#allowing-subclassing, and examples at https://vctrs.r-lib.org/articles/s3-vector.html.

[hadley]

Now tracking in r-lib/vctrs#703