`summarise()` edge case recycling bug

[DavisVaughan]

library(dplyr)
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union

df <- tibble(x = 1:3)

# Why doesn't this recycle to size 0?
summarise(df, sum = sum(x), empty = tibble())
#> Error in `dplyr_col_modify()`:
#> ! Can't recycle `empty` (size 0) to size 1.

#> Backtrace:
#>     ▆
#>  1. ├─dplyr::summarise(df, sum = sum(x), empty = tibble())
#>  2. ├─dplyr:::summarise.data.frame(df, sum = sum(x), empty = tibble())
#>  3. │ └─dplyr:::summarise_build(.data, cols)
#>  4. │   ├─dplyr::dplyr_col_modify(out, cols$new)
#>  5. │   └─dplyr:::dplyr_col_modify.data.frame(out, cols$new)
#>  6. │     └─vctrs::vec_recycle_common(!!!cols, .size = nrow(data))
#>  7. └─vctrs:::stop_recycle_incompatible_size(...)
#>  8.   └─vctrs:::stop_vctrs(...)
#>  9.     └─rlang::abort(message, class = c(class, "vctrs_error"), ..., call = vctrs_error_call(call))

# This does
summarise(df, sum = sum(x), empty = tibble(a = integer()))
#> # A tibble: 0 × 2
#> # … with 2 variables: sum <int>, empty <tibble[,1]>

# And these work
summarise(df, sum = sum(x), tibble())
#> # A tibble: 1 × 1
#>     sum
#>   <int>
#> 1     6

summarise(df, sum = sum(x), tibble(a = integer()))
#> # A tibble: 0 × 2
#> # … with 2 variables: sum <int>, a <int>

^{Created on 2022-10-18 with reprex v2.0.2.9000}

[DavisVaughan]

Something seems wrong with the summarise recycling implementation because it is treating data frames with 0 cols as "not useful" chunks, when they definitely are

dplyr/src/summarise.cpp

Line 60 in 9473bec

return !Rf_inherits(ptype, "data.frame") || XLENGTH(ptype) > 0;

[kcarnold]

Probably related bug:

library(dplyr)
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union
starwars %>% summarize(sum_height = sum(height, na.rm = TRUE))
#> # A tibble: 1 × 1
#>   sum_height
#>        <int>
#> 1      14123
starwars %>% summarize(sum_height = sum(height, na.rm = TRUE), mean_height = sum(height, na.rm = TRUE) / nrow(height))
#> # A tibble: 0 × 2
#> # … with 2 variables: sum_height <int>, mean_height <dbl>

^{Created on 2022-11-14 by the reprex package (v2.0.1)}

I would have expected the second dataframe to either have one row with the sum and some sort of missing value, or error.

[DavisVaughan]

nrow(height) gives NULL, which is size 0, and <dbl> / NULL in the division gives numeric(0), also size 0, so the whole summarise() call is recycled to size 0.

This is more related to #6382 where some people disagree with summarise() being able to return multi-row (or zero row in this case) results