Why does setting the seed in one mirai_map() use the same seed in another one?

[etiennebacher]

Hi, I came across some weird behavior related to randomness and I didn’t find related mentions in the docs (I looked for “random” and “seed”). In the example below, it appears that setting the seed in a mirai_map() keeps using the same seed in subsequent mirai_map() calls (the random draw is different across mirai_map() calls, but in a single mirai_map() the same seed is used everywhere).

library(mirai)
library(parallel)
packageVersion("mirai")
#> [1] '2.2.0.9001'

cl <- make_cluster(4)
daemons(4)
#> [1] 4
vec <- 1:3
vec2 <- 4:6

# Returns different values: good
mirai_map(list(vec, vec2), \(x) rnorm(x))[]
#> [[1]]
#> [1]  0.07040429  0.40971925 -0.39779374
#> 
#> [[2]]
#> [1] -0.5323828  0.1015743 -0.5020980

# Set the seed in the function
mirai_map(list(vec, vec2), \(x) {
  set.seed(123)
  rnorm(x)
})[]
#> [[1]]
#> [1] -0.9685927  0.7061091  1.4890213
#> 
#> [[2]]
#> [1] -0.9685927  0.7061091  1.4890213

# Do not set the seed in the function: still identical results?
mirai_map(list(vec, vec2), \(x) rnorm(x))[]
#> [[1]]
#> [1] -1.8150926  0.3304096 -1.1421557
#> 
#> [[2]]
#> [1] -1.8150926  0.3304096 -1.1421557

Maybe I come with the wrong priors but I though that “what happens in the mirai_map() stays in the mirai_map()” and doesn’t contaminate others. I see that setting dispatcher = FALSE in daemons() solves this:

daemons(0)
#> [1] 0
daemons(4, dispatcher = FALSE)
#> [1] 4

mirai_map(list(vec, vec2), \(x) {
  set.seed(123)
  rnorm(x)
})[]
#> [[1]]
#> [1] -0.9685927  0.7061091  1.4890213
#> 
#> [[2]]
#> [1] -0.9685927  0.7061091  1.4890213

mirai_map(list(vec, vec2), \(x) rnorm(x))[]
#> [[1]]
#> [1] -1.7628219  0.1110283  0.3758775
#> 
#> [[2]]
#> [1]  0.1192312 -0.4474403 -1.9538161

Should this be mentioned more prominently in the docs (unless I missed it)?

In any case, thanks for this package!

[shikokuchuo]

Attempting to set the seed in each iteration of a mirai_map() call is probably not what you'd want, as you get the same random results (as you found).

A bit more about why the behaviour is like this: mirai uses L'Ecuyer CMRG to provide parallel-safe random streams. That means random results in each daemon are independent from one another. The random seed is not reset after each mirai call by design. This is to ensure that however many random draws are made in any mirai call, the next random draw follows on in the stream, and hence have the desired statistical properties.

Once you set the seed in a mirai call, then subsequent calls will follow on, hence your findings.

It's nothing to do with dispatcher btw, if you set daemons(n = 2), you'll see that you get the same behaviour (you were simply hitting the unaffected daemons).

If you're able to give some context about what you're attempting, and what you end up doing, that would help me add something in the docs to help others. Thanks.

[etiennebacher]

Many thanks for the answer.

I started writing a vignette on using purrr and mirai with polars objects (eitsupi/neo-r-polars#321) and used rnorm() as an example of function that is in R but not in Polars. I thought I'd need to use a seed so that the output doesn't change every time (I then realized that this was pointless, first because the output would always be the same no matter the input so I would end up with a list full of identical elements, and also because the vignette is only rendered in CI so there's no risk that git detects changes in output every time we re-document the package). So this lead me to this rabbit hole, but the original rationale was wrong anyway.

If it helps, I remember that furrr has a vignette on "Common gotchas" and mentions the seed issues and reproducibility at some point. Maybe it would be useful to have such a page if you have questions that often come up in the issue tracker or in the discussion?