Seed setting for 0 and N daemons

[eli-daniels]

Hello,

I'm writing a function using purrr::in_parallel() and randomness. Essentially, I'd like my function to produce the same results whether daemons exist or not. Here is an example:

library(mirai)
library(purrr)
library(testthat)

mirai::daemons(0)
daemons_0 <- purrr::map(
  1:4,
  purrr::in_parallel(function(i) {
    rnorm(3)
  })
)

mirai::daemons(2)
daemons_2 <- purrr::map(
  1:4,
  purrr::in_parallel(function(i) {
    rnorm(3)
  })
)

expect_equal(daemons_0, daemons_2)

I'd love to know if it's possible to have daemons_0 and daemons_2 be equal, I've tried a few things with mirai::daemons(seed = 1) but not sure what the equivalent would be for when there are no daemons.

(And thanks for this wonderful extension to purrr)

[shikokuchuo]

Thanks for asking - and doubly so as it led me to find a subtle bug (#509).

So try this after installing the dev version of mirai using e.g.:

pak::pak("r-lib/mirai")

As parallel seeds work fundamentally differently to normal seeds, I suggest setting synchronous daemons with the same seed value rather than try to replicate what happens on a daemon in the main session (although that is also possible).

library(mirai)
library(purrr)
library(testthat)

seed <- 12

daemons(sync = TRUE, seed = seed)
daemons_0 <- map(1:4, in_parallel(function(i) { rnorm(3) }))

daemons(2, seed = seed)
daemons_2 <- map(1:4, in_parallel(function(i) { rnorm(3) }))

expect_equal(daemons_0, daemons_2)

When you specify seed in a daemons() call, then the results will be reproducible regardless of the number of daemons used.

For completeness, if this usage is contemplated within a CRAN package then the guidance at https://mirai.r-lib.org/articles/v06-packages.html#guidance does allow you to use daemons(sync = TRUE) as a fallback, but only after checking that !daemons_set().

[eli-daniels]

Thanks Charlie, very helpful 😄