Problems collecting child processes when using the multicore backend alongside mclapply()

[wlandau]

I see warnings like "In selectChildren(ac[!fin], -1): cannot wait for child 16782 as it does not exist" when I invoke mclapply() before cleaning up the workers. In drake, I was hoping to be able to use mclapply() while clustermq workers are still up, but it is not such a loss. Does it have something to do with the way clustermq needs to change the SIGCHLD handler?

options(clustermq.scheduler = "multicore")
w <- clustermq::workers(8)
out <- clustermq::Q(1:2, fun = function(x) x * 2, workers = w)
parallel::mclapply(1:2, sqrt, mc.cores = 2)
w$cleanup() # Warnings go away if I clean up before calling mclapply().

[mschubert]

This seems to be independent of clustermq:

p = parallel::mcparallel({ Sys.sleep(1) })
# parallel::mccollect(p) # uncomment and warnings go away
parallel::mclapply(1:2, sqrt, mc.cores = 2)
# Warning messages:
# 1: In selectChildren(ac[!fin], -1) :
#   cannot wait for child 18305 as it does not exist
#   (repeats a couple of times...)

Related: HenrikBengtsson/future#218 (comment), #92

As far as I can tell, this may be a bug in R which was introduced in 3.5.0 (where CI also started reporting these warnings with no changes to clustermq). They do not affect function, so you should be safe to suppress them; or wait for the supposed fix to be merged.

[wlandau]

Glad the bug was already discovered elsewhere. I am noticing those same warnings in R 3.5.1, so perhaps I need to wait a couple more versions.