Support a 0-length array result in summarize, when working on an empty DataFrame

[machow]

Currently, afaict dplyr handles summarizing an empty dataframe as follows:

• for each argument:
• apply the given calculation, using the empty dataframe
• use the result type to produce a new 0-length column, but discard its value (edit: grouped summarize discards the value, ungrouped summarize keeps it)
• if the result is length 0 or > 1, print a deprecation warning

This results in a sort of funky situation where operations like + raise a warning, because integer() + 1 results in a 0-length array. I wonder if a 0-length result when summarizing an empty data.frame should not be deprecated? It seems like there is already some special behavior around empty frames, and some operations returning 0-length results is likely in this case :/.

library(tidyverse)

df <- tibble(a = integer())

integer() + 1
#> numeric(0)

df %>% summarize(b = sum(a))
#> # A tibble: 1 × 1
#>       b
#>   <int>
#> 1     0

df %>% summarize(b = a + 1)
#> Warning: Returning more (or less) than 1 row per `summarise()` group was deprecated in
#> dplyr 1.1.0.
#> ℹ Please use `reframe()` instead.
#> ℹ When switching from `summarise()` to `reframe()`, remember that `reframe()`
#>   always returns an ungrouped data frame and adjust accordingly.
#> # A tibble: 0 × 1
#> # … with 1 variable: b <dbl>

^{Created on 2023-01-10 by the reprex package (v2.0.1)}

related to machow/siuba#467

edit: wait -- I just noticed ungrouped summarize keeps the value, but grouped summarize discards it...

library(tidyverse)

df <- tibble(a = integer())

df %>% group_by(a) %>% summarize(b = sum(a))
#> # A tibble: 0 × 2
#> # … with 2 variables: a <int>, b <int>

^{Created on 2023-01-10 by the reprex package (v2.0.1)}

[hadley]

I think those results are consistent:

• An ungrouped summarise always returns 1 row
• A grouped summarise always return 1 row per group

+ isn't a summarise operations so I'm not surprised we get some weirdness here.

[machow]

That's fair--I'm still trying to wrap my head around this, but something feels a bit inconsistent. For example, mutate on an empty frame allows either 0 or 1 row...

library(tidyverse)

df <- tibble(a = integer())

df %>% mutate(b = integer())
#> # A tibble: 0 × 2
#> # … with 2 variables: a <int>, b <int>

df %>% mutate(b = 1)
#> # A tibble: 0 × 2
#> # … with 2 variables: a <int>, b <dbl>

^{Created on 2023-01-11 by the reprex package (v2.0.1)}

As long as aggregation functions always return a single value, when they're given empty data, then it seems like the current behavior shouldn't be a problem for summarize?

[hadley]

Yeah, because mutate() obeys the recycling rules, which allows a length-1 vector to be expanded or shrunk to any size.

[machow]

Okay, thanks, this is all super helpful. Knowing that aggregate functions should always return some 1-length value, even when given empty data was the missing piece! (and that recycling can also reduce a 1-length value to 0-length!)

[DavisVaughan]

and that recycling can also reduce a 1-length value to 0-length

@machow these recycling rules are actually the same as the broadcasting rules used by numpy, if you want a Python connection https://numpy.org/doc/stable/user/basics.broadcasting.html#general-broadcasting-rules