Computation continues after interruption of parallelised `future` workers in R

[jfsalzmann]

Describe the bug I noticed that when doing parallel sampling from multiple (possibly different) STAN models within future workers and cancelling the execution of the main R script (the R parent process - in my case, using Ctrl+C in CLI mode), the STAN models and the children R processes will continue running as ghost processes. Appearently, something in the signal chain does not work. My guess is that although the signalling chain works well on the future side (this doesn't occur when no STAN model is running), the children R processes cannot stop as long as its children processes (= STAN process) are not stopped. I may be wrong though and this might be a bug in the future or foreach package, too.

To Reproduce

library(cmdstanr)
library(foreach)
library(doFuture)
plan(multisession,workers = 2)

foreach(i=1:2, .options.future = list(seed = TRUE), .inorder = F) %dofuture% {
    // Add code for starting the sampling of a (fairly long-running) STAN model
    // .. e.g. different(ly parameterised) models or data depending on i
}

Expected behavior Same as when running STAN models directly: Interrupting the R script execution should immediately interrupt the sampling process.

Operating system Ubuntu 22.04 LTS

CmdStanR version number 0.9.0

[jgabry]

Thank you for reporting this.

I haven't had a chance to test this for myself yet. I guess there are a bunch of possibilities here since this involves future, foreach, cmdstanr, CmdStan itself, and also processx, which is the package cmdstanr uses to run and control system processes.