Feature request: `fill(.direction = "sandwich")`

[jl5000]

It would be useful if there was an option for fill to only populate missing values if they are sandwiched by identical values before and after (with an option to decide the maximum number of sequential NAs for the filling to happen) - this is a common use case. Current solutions I've found seem to require the zoo package.

It might also be worth considering some kind of tolerance for numerics, or being able to provide a callback as a function of the two sandwich values for more niche use cases.

---

Reprex courtesy of Mario Reyes: https://stackoverflow.com/questions/51704128/na-filling-only-if-sandwiched-by-the-same-value-using-dplyr

tibble::tibble(id= 1:12,
       var1 = letters[1:12],
       var2 = c(NA,rep("A",2),rep(NA,2),rep("A",2),rep(NA,2),rep("B",2),NA))
#> # A tibble: 12 x 3
#>       id var1  var2 
#>    <int> <chr> <chr>
#>  1     1 a     <NA> 
#>  2     2 b     A    
#>  3     3 c     A    
#>  4     4 d     <NA> 
#>  5     5 e     <NA> 
#>  6     6 f     A    
#>  7     7 g     A    
#>  8     8 h     <NA> 
#>  9     9 i     <NA> 
#> 10    10 j     B    
#> 11    11 k     B    
#> 12    12 l     <NA>

^{Created on 2024-07-24 with reprex v2.0.2}

[DavisVaughan]

I think this is probably getting a bit too specific for tidyr, but you can dip down and use vctrs functions to recreate the zoo example

library(vctrs)
library(dplyr)

df_in <- tibble(id= 1:12,
  var1 = letters[1:12],
  var2 = c(NA,rep("A",2),rep(NA,2),rep("A",2),rep(NA,2),rep("B",2),NA))

df_in
#> # A tibble: 12 × 3
#>       id var1  var2 
#>    <int> <chr> <chr>
#>  1     1 a     <NA> 
#>  2     2 b     A    
#>  3     3 c     A    
#>  4     4 d     <NA> 
#>  5     5 e     <NA> 
#>  6     6 f     A    
#>  7     7 g     A    
#>  8     8 h     <NA> 
#>  9     9 i     <NA> 
#> 10    10 j     B    
#> 11    11 k     B    
#> 12    12 l     <NA>

dplyr::mutate(
  df_in,
  var_new = if_else(
    vec_fill_missing(var2, "down") == vec_fill_missing(var2, "up"),
    vec_fill_missing(var2, "down"),
    NA_character_
  )
)
#> # A tibble: 12 × 4
#>       id var1  var2  var_new
#>    <int> <chr> <chr> <chr>  
#>  1     1 a     <NA>  <NA>   
#>  2     2 b     A     A      
#>  3     3 c     A     A      
#>  4     4 d     <NA>  A      
#>  5     5 e     <NA>  A      
#>  6     6 f     A     A      
#>  7     7 g     A     A      
#>  8     8 h     <NA>  <NA>   
#>  9     9 i     <NA>  <NA>   
#> 10    10 j     B     B      
#> 11    11 k     B     B      
#> 12    12 l     <NA>  <NA>

I think an arbitrary callback is something we'd like to avoid, as it would be much too slow when there are many groups. The implementation of vec_fill_missing() (used by fill()) is in C for performance.