as_tibble_row() doesn't work with date vectors

[FrancoisR95]

In tibble 3.0.1, as_tibble_row(x) returns an error when x is a vector of dates.

The expected behavior would be to return a tibble row containing dates, exactly like as_tibble_col() or tibble() are doing.

1. Example of the issue with as_tibble_row:

      x <- c(col1 = as.Date("2012-04-04"), col2 = as.Date("2012-05-04"))
      as_tibble_row(x)

This returns the following error:

Error: `x` must be a bare vector in `as_tibble_row()`, not Date.

2. On the contrary, the same code is working fine with as_tibble_col:

      x <- c(col1 = as.Date("2012-04-04"), col2 = as.Date("2012-05-04"))
      as_tibble_col(x)

Result:

# A tibble: 2 x 1
  value     
  <date>    
1 2012-04-04
2 2012-05-04

The same is also obtained with tibble(x) instead of as_tibble_col(x).

3. Also a code for creating a 2-column tibble of dates is handled perfectly by tibble():

tibble(col1 = as.Date("2012-02-12"), col2 = as.Date(as.Date("2012-04-01"):as.Date("2012-04-05"), origin="1970-01-01"))
tibble(x)

Result:

# A tibble: 5 x 2
  col1       col2      
  <date>     <date>    
1 2012-02-12 2012-04-01
2 2012-02-12 2012-04-02
3 2012-02-12 2012-04-03
4 2012-02-12 2012-04-04
5 2012-02-12 2012-04-05

Conclusion
The behavior of as_tibble_row() with date vectors seems to be a bug, or at least is very inconsistent and unexpected. It's also really a pity when the code of the as_tibble_row(x) function (new_tibble(as.list(x), nrow = 1)) would work just fine if there wasn't, in the function, an unfortunate test making this unnecessary error... This test was introduced recently (6f572f5#diff-cb6cd20025e0469d14d1ce6271cbf679) in order to try to fix #739 but the fix is much worse than the problem (which in my opinion was not even a real problem, basically it was said that an error message, which was "All elements must be size one, use list() to wrap.", was not clear enough but I think it was rather clear, and better and clearer at least than now getting this unexpected error "x must be a bare vector in as_tibble_row(), not Date" for something which should work).

[lionel-]

It's also really a pity when the code of the as_tibble_row(x) function (new_tibble(as.list(x), nrow = 1)) would work just fine if there wasn't, in the function, an unfortunate test making this unnecessary error...

new_tibble() is a bare bones function. The implementation you're suggesting creates corrupt tibbles.

To handle arbitrary vectors, vec_chop2() would be useful but it isn't implemented yet.

library(vctrs)
library(rlang)

vec_chop2 <- function(x) {
  stopifnot(
    vec_is(x),
    # Unimplemented:
    !is.data.frame(x),
    is_null(dim(x))
  )

  names <- names(x)
  if (is_null(names)) {
    vec_chop2_unnamed(x)
  } else {
    vec_chop2_named(x, names)
  }
}
vec_chop2_unnamed <- function(x) {
  n <- vec_size(x)
  out <- vector("list", n)

  for (i in seq_len(n)) {
    out[[i]] <- x[[i]]
  }

  out
}
vec_chop2_named <- function(x, names) {
  n <- vec_size(x)
  out <- vector("list", n)
  out_names <- character(n)

  for (i in seq_len(n)) {
    out[[i]] <- x[[i]]
    out_names[[i]] <- names[[i]]
  }

  names(out) <- out_names
  out
}

vec_chop2() has more consistent handling than as.list(), for instance with factors:

fct <- set_names(factor(c("a", "b")), c("foo", "bar"))

as.list(fct)
#> [[1]]
#> foo
#>   a
#> Levels: a b
#>
#> [[2]]
#> bar
#>   b
#> Levels: a b

vec_chop2(fct)
#> $foo
#> [1] a
#> Levels: a b
#>
#> $bar
#> [1] b
#> Levels: a b

date <- c(as.Date("2020-01-01"), as.Date("2020-01-02"))
date <- set_names(date, c("foo", "bar"))

vec_chop2(date)
#> $foo
#> [1] "2020-01-01"
#>
#> $bar
#> [1] "2020-01-02"

as_tibble_row2 <- function(x) {
  if (!vec_is_list(x)) {
    x <- vec_chop2(x)
  }

  tibble::as_tibble_row(x)
}

as_tibble_row2(fct)
#> # A tibble: 1 x 2
#>   foo   bar
#>   <fct> <fct>
#> 1 a     b

as_tibble_row2(date)
#> # A tibble: 1 x 2
#>   foo        bar
#>   <date>     <date>
#> 1 2020-01-01 2020-01-02

[krlmlr]

as_tibble_row() doesn't document that it requires bare vectors.

We could do new_tibble(map(seq_along(x)), ~ vec_slice(x, .), nrow = 1) . However, just like #740, this requires vec_names(). Any chance we can make this available in vctrs?

[FrancoisR95]

Many thanks @lionel- and @krlmlr for your quick replies and explanations.
I still don't understand, however, what is the problem of new_tibble(as.list(x), nrow = 1) with date vectors (while I understand the problem with factors).

In your example of date vector, as.list(x) seems to have exactly the same result as vec_chop2(x).

date <- c(as.Date("2020-01-01"), as.Date("2020-01-02"))
date <- set_names(date, c("foo", "bar"))

vec_chop2(date)
#> $foo
#> [1] "2020-01-01"
#>
#> $bar
#> [1] "2020-01-02"

as.list(date)
#> $foo
#> [1] "2020-01-01"
#>
#> $bar
#> [1] "2020-01-02"

Why not, for the as_tibble_row() function, write a test like:

if (!is_bare_vector(x)) {
  if (class(x) == "Date") {
    if (!is_vector(x)) {
      cnd_signal(error_as_tibble_row_date_vector(x))
    }
  } else {  
    cnd_signal(error_as_tibble_row_bare(x))
  }
}

instead of:

if (!is_bare_vector(x)) {
  cnd_signal(error_as_tibble_row_bare(x))
}

[lionel-]

@FrancoisR95 We're looking into a more general solution that would work with any vector.

@krlmlr

We could do new_tibble(map(seq_along(x)), ~ vec_slice(x, .), nrow = 1) . However, just like #740, this requires vec_names(). Any chance we can make this available in vctrs?

Yes I think we could now expose vec_names() and vec_set_names(). How do you plan to deal with names exactly?

[krlmlr]

The names of the vector become column names here?

[lionel-]

Yup so we need to remove the inner names and use them as outer names. Also data frames should be special-cased here so I'm not sure where vec-names would come into play. That's why I thought vec-chop2 as outlined above would be appropriate for non-list and non-df inputs.

[krlmlr]

I don't follow. Yes, I can use inner names as outer names with vec_names() (forgot to actually submit that PR). Why would we special-case data frames? Wouldn't vec_slice() take care of that?

If we can make vec_chop2() available in vctrs -- fine. Maybe we can prototype here?

[lionel-]

We also need to remove the inner names otherwise they linger in the columns.

Regarding data frames, I assumed as_tibble_row(mtcars[1, ]) should be a no-op and fail if the input is not size 1? It seems strange to transform each row to a column, and would only work for data frames with row names which is strange as well.