Should constructors initialized with `NA` set all elements to `NA`?

[DavisVaughan]

I find it a little strange that if you add an NA, or initialize with NA, that it doesn't set all of the slots to NA. I feel like this implies that there are multiple "types" of missing Period objects. It might help internal consistency if NA propagated to all slots. This would formalize a missing Period object as one where all slots are NA

library(lubridate)

days(NA_real_)
#> [1] NA

str(days(NA_real_))
#> Formal class 'Period' [package "lubridate"] with 6 slots
#>   ..@ .Data : num NA
#>   ..@ year  : num 0
#>   ..@ month : num 0
#>   ..@ day   : num NA
#>   ..@ hour  : num 0
#>   ..@ minute: num 0

days(NA_real_) + years(1)
#> [1] NA

days(1) + years(NA_real_)
#> [1] NA

str(days(NA_real_) + years(1))
#> Formal class 'Period' [package "lubridate"] with 6 slots
#>   ..@ .Data : num NA
#>   ..@ year  : num 1
#>   ..@ month : num 0
#>   ..@ day   : num NA
#>   ..@ hour  : num 0
#>   ..@ minute: num 0

str(days(1) + years(NA_real_))
#> Formal class 'Period' [package "lubridate"] with 6 slots
#>   ..@ .Data : num NA
#>   ..@ year  : num NA
#>   ..@ month : num 0
#>   ..@ day   : num 1
#>   ..@ hour  : num 0
#>   ..@ minute: num 0

This is what I expect internally:

library(lubridate)
#> 
#> Attaching package: 'lubridate'
#> The following object is masked from 'package:base':
#> 
#>     date

period()[NA_real_]
#> [1] NA

str(period()[NA_real_])
#> Formal class 'Period' [package "lubridate"] with 6 slots
#>   ..@ .Data : num NA
#>   ..@ year  : num NA
#>   ..@ month : num NA
#>   ..@ day   : num NA
#>   ..@ hour  : num NA
#>   ..@ minute: num NA

[vspinu]

I thought about this briefly recently but didn't implement because of the efficiency and simplicity reasons. But I agree, for the sake of things like minute(days(NA)) it might be worth implementing it.

What do you need this internal "consistency" for? I am just afraid that syncronizing them might be a rabbit hole and might not be done properly in all situations.

[DavisVaughan]

I'm working on vctrs methods for lubridate, and was adding a test that implicitly required period()[NA_integer_] to be identical to period(NA_integer_)

[vspinu]

Hmm, interesting corner case. Ok, let's add this in the constructor then. It should not harm, and it's the right thing to do anyhow.