Ensure reproducible if using model cache

[gowerc]

Closes #469

FYI @luwidmer

[luwidmer]

@gowerc doesn't .Random.seed always exist? At least I get

Restarting R session...

> exists(".Random.seed")
[1] TRUE
> ls()
character(0)
> .Random.seed <<- .Random.seed
> ls()
character(0)
> exists(".Random.seed")
[1] TRUE

[gravesti]

I think it doesn't necessarily exist. eg if I start a new R session from the terminal

gravesti$ R

R version 4.4.1 (2024-06-14) -- "Race for Your Life"
Copyright (C) 2024 The R Foundation for Statistical Computing
Platform: aarch64-apple-darwin20
... snip ...

> .Random.seed
Error: object '.Random.seed' not found
> rnorm(1)
[1] -0.6881252
> .Random.seed
  [1]       10403           2 -1831660992  -462780898 -1349857047   996995611
  [7]   676991530   594638156 -1127753969  1257593413  1702644732  1213026898
 [13]  -333959219  -261150329 -1072259986   635315272  1119354443    93778857
 [19]   864664312   193269286  -542546111 -1416680237  1130410114   418352820
 [25]   -98325961  -999378643 -1509490108  1850580842   640379861   544927087

[gravesti]

@gowerc
To be a bit more robust, eg in case the function has an error, you can do something like

Suggested change

	current_seed_state <- .Random.seed
	current_seed_state <- .Random.seed
	on.exit(suspendInterrupts(
	assign(".Random.seed", value = current_seed_state, envir = globalenv(), inherits = FALSE)
	))

In another package we have

    # Make sure to leave '.Random.seed' as-is on exit
    old_seed <- globalenv()$.Random.seed
    on.exit(suspendInterrupts(set_random_seed(old_seed)))

with the function

# Restore the RNG back to a previous state using the global .Random.seed
set_random_seed <- function(old_seed) {
  if (is.null(old_seed)) {
    rm(".Random.seed", envir = globalenv(), inherits = FALSE)
  } else {
    assign(".Random.seed", value = old_seed, envir = globalenv(), inherits = FALSE)
  }

[gowerc]

yer good idea will update !

[gowerc]

@luwidmer / @gravesti - Sorry if a stupid question but what exactly is suspendInterrupts supposed to be doing here ? My guess based on the sparse docs would be that it block SIG* interrupts however SIGTERM still seems to kill the program regardless:

I then wondered if it meant R based signals but it has no affect on that either:

> tryCatch(
+   {
+     suspendInterrupts({
+       warning("hi")
+       print("hello")
+     })
+   },
+   warning = \(e) print("Caught warning")
+ )
[1] "Caught warning"

[gowerc]

Looks like it just doesn't work properly on MacOS / Linux:
HenrikBengtsson/Wishlist-for-R#82

[gravesti]

These changes wouldn't be needed if you use on.exit()

[luwidmer]

Yup, turns out I didn't have a completely new R session even after restarting! Thank you for testing @gowerc & @gravesti ! IMO the on.exit() solution (when .Random.seed exists) would be a pretty robust one.

[gowerc]

Yup, turns out I didn't have a completely new R session even after restarting! Thank you for testing @gowerc !

In fairness I had no idea it wasn't created on new R sessions until all the CICD tests failed unexpectedly 😓