withr::with_preserve_seed() does not preserve is.null(.Random.seed)

[HenrikBengtsson]

Example:

$ Rscript --vanilla -e "str(.GlobalEnv$.Random.seed); withr::with_preserve_seed(x <- sample.int(1L)); str(.GlobalEnv$.Random.seed)"
 NULL
int [1:626] 10403 1 1586168631 257966637 -1833262268 2037450858 175878869 786533487 163480950 -2036568416 ...

> packageVersion("withr")
[1] '2.1.2'

[krlmlr]

I'm not sure it's even possible to preserve a NULL random seed. Can we solve this by documentation?

[HenrikBengtsson]

Sorry, I wasn't clear. The value is not NULL - that's just a side effect of <env>$ - its that it does not exist at all. To solve it, first check exists(".Random.seed", envir = .GlobalEnv) (or is.null(oseed)). Then, reset .Random.seed either to the old value or remove it using rm(".Random.seed", envir = .GlobalEnv). This approach is used in several packages.

BTW, .GlobalEnv$.Random.seed <- NULL is not a valid seed. It'll trigger warnings by R itself.

I'd say it's not sufficient to document. A real-world problem where withr::with_preserve_seed() fail because of this is:

$ R --quiet --vanilla
> exists(".Random.seed", envir = .GlobalEnv)
[1] FALSE
> library(ggplot2)
> exists(".Random.seed", envir = .GlobalEnv)
[1] TRUE

It's an unfortunate side effect that loading/attaching a package causes an update to the RNG state. It prevents certain type of RNG validations. (This is actually how I tracked it down).

[jimhester]

#125 should fix this, after that change we have the following behavior (run in an interactive session so that the ggplot onAttach code would trigger)

exists(".Random.seed", envir = .GlobalEnv)
#> [1] FALSE
library(ggplot2)
exists(".Random.seed", envir = .GlobalEnv)
#> [1] FALSE

[HenrikBengtsson]

Yes, LGTM. Thxs. Not sure when you want to close this issue, so I leave that to you.