vec_slice() cannot slice 0-column data.frames properly

[yutannihilation]

library(vctrs)
library(tibble)

vec_slice(tibble(.rows = 3), 1:3)
#> # A tibble: 0 x 0

vec_slice(tibble(x = tibble(.rows = 3)), 1:3)
#> Error: Positive column indexes in `[` must match number of columns:
#> * `.data` has 0 columns
#> * Position 1 equals 1
#> * Position 2 equals 2
#> * Position 3 equals 3
#> Backtrace:
#>     █
#>  1. └─vctrs::vec_slice(tibble(x = tibble(.rows = 3)), 1:3)
#>  2.   └─base::lapply(x, `[`, i)
#>  3.     ├─base:::FUN(X[[i]], ...)
#>  4.     └─tibble:::`[.tbl_df`(X[[i]], ...)
#>  5.       └─tibble:::check_names_df(i, x)
#>  6.         └─tibble:::check_names_df_numeric(j, x)

^{Created on 2019-01-31 by the reprex package (v0.2.1.9000)}

[yutannihilation]

It seems rownames are lost here:

vctrs/R/size.R

Lines 108 to 109 in ed62f1e

	out <- lapply(x, vec_slice, i)
	vec_restore(out, x)

[yutannihilation]

To fix this, I feel we have to provide the expected length to vec_restore(). Otherwise it cannot know what length the result is supposed to be.

library(vctrs)

x <- tibble::tibble(.rows = 3)
out <- lapply(x, vec_slice, 1)

# the expected length is 1, but neither of out nor x know it here
vec_restore(out, x)
#> # A tibble: 0 x 0

^{Created on 2019-02-07 by the reprex package (v0.2.1)}

Does this look OK...? if so, I'll make a PR.

vec_restore <- function (x, to, ...) {
    UseMethod("vec_restore", to)
}

vec_restore.data.frame <- function (x, to, n = NULL, ...) {
  ...
  attr_to[["row.names"]] <- .set_row_names(n %||% df_length(x))
  ...
}

[hadley]

I think if you have to supply arguments, it's a sign that you shouldn't be using vec_restore() — I think I may have used vec_restore() in places (like here) where it's not actually suitable.

[yutannihilation]

Hmm..., thanks, I agree it's a bad sign. But, I don't think I "shouldn't be using vec_restore()," because vec_restore() is still useful (as long as it doesn't raise errors) for restoring classes and attributes except for rownames.

Probably, we can add rownames again after vec_restore(). Or, do you think it's better to remove vec_restore() here?

[hadley]

The more I think about it, the more I think this use of vec_restore() is inappropriate — this clause is just a hack around the slowness of [.data.frame and there's no guarantee that it will work with other data frame subclasses.

I'm not sure what we should do here given that [.tibble is much more efficient, so there's no need for the hack. I suspect we should be testing identical(class(x), "data.frame") and then calling new_data_frame() on the output of the lapply().

Alternatively, this might be resolved by #124, where we'll make vec_slice generic, so it would make more sense to provide our own specialised implementation of row-subsetting a data frame.

[yutannihilation]

Thanks, given the current vec_slice() is just a hack for speed, that's convincing. Then I'll just wait for other discussions.

[hadley]

Oh please leave this one open; it was way more info.

[yutannihilation]

Ah, sorry. Sure.

[hadley]

It turns out that the fixes needed for the newly focussed #199, also ended up fixing this issue. I think this is correct because I was just too aggressively throwing away the row names attribute (throwing the baby out with the bath water), but I'd appreciate it if you'd take a look at the patch and see if it makes sense to you too.

[yutannihilation]

@hadley Sorry, I don't think this is "fixed." The second one below no more raises an error, but the result seems still incorrect; the dimensions should be 3 x 0, and 3 x 1.

library(vctrs)
library(tibble)

packageDescription("vctrs")["GithubSHA1"]
#> $GithubSHA1
#> [1] "3b80a81418d0f8037f86e875e80e5a73b2744485"

vec_slice(tibble(.rows = 3), 1:3)
#> # A tibble: 0 x 0

vec_slice(tibble(x = tibble(.rows = 3)), 1:3)
#> # A tibble: 0 x 1
#> # ... with 1 variable: x <lgl>

^{Created on 2019-02-27 by the reprex package (v0.2.1)}