`names_glue` evaluation issues

[lionel-]

Environment

Currently no environment is passed to glue_data() so the current environment is used as default. This means that the glue string can access internal tidyr data:

us_rent_income %>%
  pivot_wider(
    names_from = variable,
    names_glue = "{variable}_{.value}_{names_vary}",
    values_from = c(estimate, moe)
  ) %>%
  names()
#> [1] "GEOID"                   "NAME"
#> [3] "income_estimate_fastest" "rent_estimate_fastest"
#> [5] "income_moe_fastest"      "rent_moe_fastest"

And conversely can't access user data that's not reachable through the global env:

local({
  my_string <- "foo"

  us_rent_income %>%
    pivot_wider(
      names_from = variable,
      names_glue = "{variable}_{.value}_{my_string}",
      values_from = c(estimate, moe)
    )
})
#> Error:
#> ! object 'my_string' not found

Englueing

@charliejhadley brought up the following problem a few months ago: https://gist.github.com/charliejhadley/47f123256d392b569fc3c33ec3d436ac

It could be solved by using rlang::englue() instead of glue::glue_data(). This would require r-lib/rlang#1565. Then the solution would be:

fn <- function(data, target_column){
  data %>%
    pivot_wider(
      names_from = {{ target_column }},
      names_glue = "{{ target_column }}_{.value}",
      values_from = c(estimate, moe)
    )
}

us_rent_income %>% fn(variable)

[DavisVaughan]

In unnest_longer() I remember making a conscious decision to evaluate indices_to and values_to in an empty environment that just knows about {col}

tidyr/R/unnest-longer.R

Lines 290 to 295 in 0764e65

	glue_col_names <- function(string, col_names) {
	data <- list(col = col_names)
	out <- glue::glue_data(data, string, .envir = NULL)
	out <- as.character(out)
	out
	}

I'm not saying that is 100% correct, just that we should consider updating here too if we decide to change anything

[lionel-]

That's a valid approach as well. It makes it easier to implement but removes some flexibility for users (e.g. varying suffixes/prefixes).

If we want the englue syntax though, we necessarily need the user environment because that's where the embraced function arguments live.

[lionel-]

bummer, I forgot to relax this:

Error in `rlang::englue()`:
#> ! Must use `{{`.
#> ℹ Use `glue::glue()` for interpolation with `{`.

This prevents englue() from being used in tidyr. I've opened an issue to remember: r-lib/rlang#1601