[R] Let `na.rm` of `mean()` support removing `NaN` as in base R

[psychelzh]

Describe the enhancement requested

See the following example. In base R na.rm will also remove NaN values, because of the fact that is.na(NaN) == TRUE.

suppressPackageStartupMessages(library(dplyr))
data.frame(x = c(1, 2, NaN)) |> 
  arrow::as_arrow_table() |>
  summarise(x_avg = mean(x, na.rm = TRUE)) |> 
  collect()
#> # A tibble: 1 × 1
#>   x_avg
#>   <dbl>
#> 1   NaN

data.frame(x = c(1, 2, NaN)) |> 
  summarise(x_avg = mean(x, na.rm = TRUE))
#>   x_avg
#> 1   1.5

^{Created on 2024-09-12 with reprex v2.1.0}

Component(s)

R

[amoeba]

Hi @psychelzh, a change here might require some discussion. My understanding is that different types of systems treat NaNs differently and what you're seeing is a result of arrow having different semantics than R. R and the pandas Python package treat NaNs as NA/None whereas database systems do not and I think arrow is emulating the database system semantics here.

For comparison, Python has a similar situation as R: pandas treats NaN as NA/null (and actually ignores them by default so this is slightly different from R):

In [1]: pd.Series([1, 2, np.nan]).mean()
Out[1]: 1.5

While PyArrow does not:

In [2]: pc.mean(pa.array([1, 2, np.nan]))
Out[2]: <pyarrow.DoubleScalar: nan>

However, because compatibility is important, PyArrow supports a from_pandas flag which makes PyArrow match pandas semantics:

In [3]: pc.mean(pa.array([1, 2, np.nan], from_pandas=True))
Out[3]: <pyarrow.DoubleScalar: 1.5>

There are a few ways this could be addressed and I think this type of issue has come up before in places like #29060. The R package could add something similar to PyArrow's from_pandas flag or an option could be added to individual kernels (mean in this case). Thoughts @ianmcook @nealrichardson @jonkeane @thisisnic?

[thisisnic]

I have mixed feelings - I'm generally opposed to changing how something has worked for a long time given that other users may be depending on it or working around it , but in favour of matching base R when we are emulating their functions.

@amoeba I'm a bit rusty with the codebase - if we were to do this, how complex would it be to implement?

Note: we allow NaN to be treated as NA on Arrow Datum objects, e.g.

library(arrow)

my_vec <- c(42, NA, NaN)
is.na(my_vec)
#> [1] FALSE  TRUE  TRUE

my_arrow_array <- Array$create(my_vec)
is.na(my_arrow_array)
#> Array
#> <bool>
#> [
#>   false,
#>   true,
#>   true
#> ]

I think, given the above behaviour that we have chosen to implement, if this isn't too complex and someone is willing to implement it, we could totally look into supporting it. Could we perhaps have it as an option? Or something else? Keen to hear what others think though.

Some tangentially related PRs:

• [C++] Should NaN comparison return false or NaN/NA? #29038
• [R] Cast of NaN to integer should return NA_integer_ #29626
• [R] is.na() evaluates to FALSE on Arrow NaN values #27883
• ARROW-12964: [R] Add bindings for ifelse() and if_else() #10724

[thisisnic]

This issue has been marked as stale because it has had no activity in the past 365 days. Please remove the stale label or comment below, or this issue will be closed in 14 days. If this improvement is still desired but has no current owner, please add the 'Status: needs champion' label.