pivot_wider does not work well with data.table

[ColeMiller1]

Per this StackOverflow Question, it appears that there can be errors when data.table is loaded and data.table methods are dispatched on data.table objects (i.e., [.data.frame vs. [.data.table.

library(data.table)
library(tidyr)
library(dplyr)

df <- data.frame(names=letters[1:3],values=1:3)
df %>% pivot_wider(names_from=names,values_from=values)
#> # A tibble: 1 x 3
#>       a     b     c
#>   <int> <int> <int>
#> 1     1     2     3

dt <- data.table(names1=letters[1:3],values1=1:3)
pivot_wider(dt, names_from=names1,values_from=values1)
#> Error in data.frame(row = row_id, col = col_id): arguments imply differing number of rows: 0, 3

This is the traceback:

5: stop(gettextf("arguments imply differing number of rows: %s",
paste(unique(nrows), collapse = ", ")), domain = NA)
4: data.frame(row = row_id, col = col_id)
3: pivot_wider_spec(data, spec, !!id_cols, names_repair = names_repair,
values_fill = values_fill, values_fn = values_fn)
2: pivot_wider.data.frame(dt, names_from = names1, values_from = values1)
1: pivot_wider(dt, names_from = names1, values_from = values1)

I believe this is the problem:

tidyr/R/pivot-wide.R

Line 224 in 501a757

df_rows <- data[key_vars]

To make it agnostic between data.table and data.frame methods, this should work:

df_rows <- data[, key_vars, drop = FALSE]

As is, dt[cols] would be equivalent to df[cols, ]. I can make a PR if you'd like.

[ColeMiller1]

While looking into it, my proposed solution does not work. This particular example ended up being df[character(0L)] which returns a zero-column data.frame. The proposed dt[, character(0L)] returns character(0L) because that is how j was designed in dt[i, j, by].

An alternate solution would be to use length() instead of ncol() would provide a way forward with data.table. This would be the line to change:

tidyr/R/pivot-wide.R

Line 225 in 501a757

if (ncol(df_rows) == 0) {

There are actually some other things that need changed. This would address the problem with all tests passing. I will plan to do a PR.

  # Figure out rows in output
  df_rows <- data[, key_vars, drop = FALSE] ##changed
  if (length(df_rows) == 0) { ##changed from ncol(df_rows) == 0
    rows <- tibble(.rows = 1)
    nrow <- 1L
    row_id <- rep(1L, nrow(data)) ##changed from nrow(df_rows); it is strange to have a 0 column n row data.frame IMO
  } else {
    row_id <- vec_group_id(df_rows)
    nrow <- attr(row_id, "n")
    rows <- vec_slice(df_rows, vec_unique_loc(row_id))
  }