Q() segfaults on Windows & Linux with multiprocess scheduler with clustermq 0.9.0

[luwidmer]

I also get segfaults with the following even simpler Q() example (simpler than #306), both on Linux and Windows:

options(clustermq.scheduler = "multiprocess")
library(clustermq)
fun <- function(x) {x}

fun(1)
Q(fun = fun, x = 1:1000, n_jobs = 2)
Q(fun = fun, x = 1:1000, n_jobs = 2)

On Linux with R 4.1.0, this results in

Starting 2 processes ...
Running 1,000 calculations (5 objs/19.3 Kb common; 1 calls/chunk) ...
[===================================================>] 100% (2/2 wrk) eta:  0sAssertion failed: check () (src/msg.cpp:387)
Aborted

On Windows with R 4.3.0 this results in the same error as for @wlandau's example in #306:

Starting 2 processes ...
Running 1,000 calculations (5 objs/19.3 Kb common; 1 calls/chunk) ...
[===================================================>] 100% (2/2 wrk) eta:  0sAssertion failed: check () (../zeromq-4.3.4/src/msg.cpp:414)

[luwidmer]

Seems to be related to shutdown of workers given that it always happens after all jobs have completed. Also, can provoke this with LSF (but it is harder to reproduce, needs Q() in a loop)

[luwidmer]

@mschubert are you able to reproduce this as well?

[mschubert]

Yes: I can (occasionally) reproduce, and I'll try to track it down as soon as possible.

I'm also happy to report that I've got internet again at the place I moved to 😅

[mschubert]

@luwidmer Can you check if it still occurs with the current git version?

remotes::install_github("mschubert/clustermq@master")

[luwidmer]

Starting 2 processes ...
Running 1,000 calculations (5 objs/19.3 Kb common; 1 calls/chunk) ...
[===================================================>] 100% (2/2 wrk) eta:  0sAssertion failed: check () (../zeromq-4.3.4/src/msg.cpp:414)

Unfortunately yes (I modified the version number in DESCRIPTION to be 0.9.0.12345 and that version indeed got loaded)

[mschubert]

I fixed another bug in 5612364, which may be the cause of this crash as well. Can you confirm if this now works? (same git install command as above)

[luwidmer]

I just really tried to provoke it with 1000s of Q() calls, that seems to have done it, superb @mschubert ! Might make sense to push this as 0.9.1 if no other big issues pop up?

[mschubert]

Great, thanks!

Yes, plan is to push 0.9.1 within the next few days, there are still some other issues to fix.