Referencing other columns inside mutate/summarise across is broken in 1.0.4

[juancamilog]

In dplyr 1.0.3 you can reference other columns in the same data-frame/tibble group by name. This functionality is broken in 1.0.4

To reproduce, the following example works in 1.0.3

storms %>% mutate(across(c('wind', 'pressure'), function(x) {
  return(x/lat)
}))

In 1.0.4, running the above example results in the following error message

>rlang::last_error()
<error/dplyr:::mutate_error>
Problem with mutate() input ..1.
x object 'lat' not found

i.e. we can't reference other columns by name. A possible workaround is to use cur_data()$name_of_column but this is slower as the following benchmark demonstrates:

library(dplyr, warn.conflicts = F)

df <- tibble(cbind.data.frame(
  grp_1 = sort(rep(1:250, 4)),
  grp_2 = rep(1:4, 250), 
  matrix(rnorm(1000 * 100), nrow = 1000)))

bench::mark(iterations = 100,
            filter_gc = FALSE,
            use_cur_data = df %>% summarise(across(is.numeric, function(x) {
              rows = cur_data()
              mask = (rows$grp_1 %% 2) == 0
              return(mean(x[mask] / rows$grp_2[mask]))
            })),
            direct_reference = df %>% summarise(across(is.numeric, function(x) {
              mask = (grp_1 %% 2) == 0
              return(mean(x[mask] / grp_2[mask]))
            }))) %>%
  select(expression, min, median, `itr/sec`, `gc/sec`)

which results in the following output

# A tibble: 2 x 5
  expression            min   median `itr/sec` `gc/sec`
  <bch:expr>       <bch:tm> <bch:tm>     <dbl>    <dbl>
1 use_cur_data       17.2ms  18.29ms      51.1    10.7 
2 direct_reference   8.12ms   8.59ms     108.      9.76

TL;DR In dply 1.0.3, using cur_data()$column_name to reference columns instead of directly using the column names can be considerably slower. In 1.0.4 referencing columns by name, not using cur_data, is currently broken.

[romainfrancois]

I'm not sure why this worked previously. It seems odd that a function would be instrumented like this.

OTOH, functions created from formulas are:

library(dplyr, warn.conflicts = FALSE)

storms %>% 
  mutate(across(c(wind, pressure), ~ . / lat))
#> # A tibble: 10,010 x 13
#>    name   year month   day  hour   lat  long status category  wind pressure
#>    <chr> <dbl> <dbl> <int> <dbl> <dbl> <dbl> <chr>  <ord>    <dbl>    <dbl>
#>  1 Amy    1975     6    27     0  27.5 -79   tropi… -1       0.909     36.8
#>  2 Amy    1975     6    27     6  28.5 -79   tropi… -1       0.877     35.5
#>  3 Amy    1975     6    27    12  29.5 -79   tropi… -1       0.847     34.3
#>  4 Amy    1975     6    27    18  30.5 -79   tropi… -1       0.820     33.2
#>  5 Amy    1975     6    28     0  31.5 -78.8 tropi… -1       0.794     32.1
#>  6 Amy    1975     6    28     6  32.4 -78.7 tropi… -1       0.772     31.2
#>  7 Amy    1975     6    28    12  33.3 -78   tropi… -1       0.751     30.4
#>  8 Amy    1975     6    28    18  34   -77   tropi… -1       0.882     29.6
#>  9 Amy    1975     6    29     0  34.4 -75.8 tropi… 0        1.02      29.2
#> 10 Amy    1975     6    29     6  34   -74.8 tropi… 0        1.18      29.5
#> # … with 10,000 more rows, and 2 more variables: ts_diameter <dbl>,
#> #   hu_diameter <dbl>

^{Created on 2021-02-04 by the reprex package (v0.3.0)}

[juancamilog]

Sometimes it is useful to change the value of a set of columns (the columns inside the across statement), depending on the value of other columns for the corresponding rows.

[juancamilog]

Actually, this can be dealt with in 1.0.4 by doing the following

storms %>% mutate(
  across(c('wind', 'pressure'),
            function(x, lat_) {
              return(x/lat_) 
            },
            lat
  )
)

Now lat_ is available inside the function scope and corresponds to what I want.

This issue can be closed

[juancamilog]

Reopening because the example in the comment above does not work in 1.0.4