Add grouped-df method for complete()

[DavisVaughan]

Closes #396
Closes #966
Closes #1110
Closes #1288 (Alternate approach)

Here is CRAN complete(), which demonstrates the issue from #396.

library(tidyr)
library(dplyr, warn.conflicts = FALSE)

df <- tibble(
  g = factor(c("x", "y", "y")),
  id = c("a", "a", "b"),
  x = c(1, 2, 3)
)

gdf <- group_by(df, g)
gdf
#> # A tibble: 3 × 3
#> # Groups:   g [2]
#>   g     id        x
#>   <fct> <chr> <dbl>
#> 1 x     a         1
#> 2 y     a         2
#> 3 y     b         3

complete(gdf, g, id)
#> # A tibble: 6 × 3
#> # Groups:   g [2]
#>   g     id        x
#>   <fct> <chr> <dbl>
#> 1 x     a         1
#> 2 y     a         2 # <- This should be NA, not 2
#> 3 x     a         1 # <- This should be NA, not 1
#> 4 x     b        NA
#> 5 y     a         2
#> 6 y     b         3

complete() on a grouped data frame should work like:

• Split by group
• Complete on each group
• Combine results

In other words, we should have gotten this:

res_x <- complete(df[1,], g, id)
res_x
#> # A tibble: 2 × 3
#>   g     id        x
#>   <fct> <chr> <dbl>
#> 1 x     a         1
#> 2 y     a        NA

res_y <- complete(df[2:3,], g, id)
res_y
#> # A tibble: 4 × 3
#>   g     id        x
#>   <fct> <chr> <dbl>
#> 1 x     a        NA
#> 2 x     b        NA
#> 3 y     a         2
#> 4 y     b         3

group_by(vec_rbind(res_x, res_y), g)
#> # A tibble: 6 × 3
#> # Groups:   g [2]
#>   g     id        x
#>   <fct> <chr> <dbl>
#> 1 x     a         1
#> 2 y     a        NA
#> 3 x     a        NA
#> 4 x     b        NA
#> 5 y     a         2
#> 6 y     b         3

The issue is that there is no grouped-df method for complete(), so we end up correctly doing the expansion by group using the grouped-df part of expand(), but we don't do the join right. Rather than joining each expansion to each individual group it was expanded from, we join the total expanded result to the original data, which results in the wrongly filled values from above.

I feel that we probably didn't ever need a grouped-df method for expand() for a few reasons:

• By definition, it generates a "new" data frame and is really just a crossing() wrapper! So it should return a bare tibble.
• I think an invariant of expand() should be that it never returns duplicated rows. But that can happen through the grouped-df method (see below)
• We could have instead fixed this with the grouped-df method for complete(), which is added here

That said, I think removing the grouped-df method for expand() would probably break at least one revdep (I haven't checked), so I am not doing that here.

library(tidyr)
library(dplyr, warn.conflicts = FALSE)

df <- tibble(
  g = factor(c("x", "y", "y")),
  id = c("a", "a", "b")
)

gdf <- group_by(df, g)

# (x, a) is duplicated, for example
expand(gdf, g, id)
#> # A tibble: 6 × 2
#> # Groups:   g [2]
#>   g     id   
#>   <fct> <chr>
#> 1 x     a    
#> 2 y     a    
#> 3 x     a    
#> 4 x     b    
#> 5 y     a    
#> 6 y     b

# that should probably just return this:
expand(df, g, id)
#> # A tibble: 4 × 2
#>   g     id   
#>   <fct> <chr>
#> 1 x     a    
#> 2 x     b    
#> 3 y     a    
#> 4 y     b