separate drops attributes of data frames

[CameronBieganek]

This is similar to the closed dplyr issue tidyverse/dplyr#1064. Here's a minimal example:

data_frame(
   x = c('blue:circle', 'blue:circle', 'orange:square', 'orange:square')
) %>%
   `attr<-`('val', 101) %>%
   separate(x, c('color', 'shape'), sep = ':') %>%
   str
# Classes ‘tbl_df’, ‘tbl’ and 'data.frame': 4 obs. of  2 variables:
#  $ color: chr  "blue" "blue" "orange" "orange"
#  $ shape: chr  "circle" "circle" "square" "square"

I haven't tested any of the other tidyr functions to see if they keep or drop attributes.

[hadley]

The problem is that it's difficult in general to copy attributes, because you don't know if the attributes are global, per-row, or per-column.

The root cause is append_df():

append_df <- function(x, values, after = length(x)) {
  y <- append(x, values, after = after)
  class(y) <- class(x)
  attr(y, "row.names") <- attr(x, "row.names")

  y
}

which presumably loses attributes because of the subsetting + c() inside append() (and you can see I've already done a little to repair them).

I experimented with an alternative implementation:

append_df <- function(x, values, after = length(x)) {
  UseMethod("append_df")
}

#' @export
append_df.data.frame <- function(x, values, after = length(x)) {
  n <- length(x)
  if (after <= 0) {
    dplyr::bind_cols(values, x)
  } else if (after >= n) {
    dplyr::bind_cols(x, values)
  } else {
    dplyr::bind_cols(x[1L:after], values, x[(after + 1L):n])
  }
}
#' @export
append_df.tbl_dt <- function(x, values, after = length(x)) {
  tbl_dt(NextMethod())
}

but that doesn't preserve attributes, presumably because the rules for how you should combine when binding columns are non-obvious.

I also tried:

append_df <- function(x, values, after = length(x)) {
  n <- length(x)
  if (after <= 0) {
    x[] <- c(values, x)
  } else if (after >= n) {
    x[] <- c(values, x)
  } else {
    x[] <- c(x[1L:after], values, x[(after + 1L):n])
  }
  x
}

But that doesn't work because you can't use sub-assignment to increase the number of variables.

If you can think of a way to insert a column in an arbitrary position without losing attributes, I'd be happy to incorporate it

[CameronBieganek]

I agree that it's not always clear which attributes ought to be copied. However, for the specific case of separate, I think that it's more often correct to copy all the attributes than it is to drop them. Working under that assumption, the following seems like it should work:

append_df <- function(x, values, after = length(x)) {
   y <- append(x, values, after = after)
   ynames <- names(y)
   attributes(y) <- attributes(x)
   names(y) <- ynames    # Since the above line overwrites the correct names

   y
}

Unfortunately, I don't remember my original use case when I opened the issue. Perhaps I wanted to attach some meta-information to a data frame and then thought better of it.