Skip to contents

Stop a PSOCK cluster and reap workers that outlive the request

Source: R/cluster_teardown.R
dot-mosaic_stop_cluster.Rd

parallel::stopCluster() shuts a worker down by writing a shutdown message to its socket. A worker that is not currently reading that socket never receives it, so stopCluster() returns cleanly while the worker survives its own cluster. That happens whenever a run is interrupted mid-task, or a task stalls inside run_simulation() and the gather gives up on it.

Usage

.mosaic_stop_cluster(cl, pids = NULL)

.mosaic_cluster_worker_pids(cl)

.mosaic_proc_cmdline(pid)

.mosaic_psock_alive(pids)

Arguments

cl

A cluster object from parallel::makeCluster() or make_mosaic_cluster().

pids

Integer worker PIDs. Defaults to the "mosaic_worker_pids" attribute of cl, else a live query.

Value

NULL, invisibly.

Details

A leaked worker is not merely idle: it inherited the parent's stdout, so the pipe never reaches EOF and a finished run looks like it is hanging (no output from tail, no R master in the process table). It also holds ~1 GB of RSS and one of R's 128 connection slots, which is enough to make a later cluster creation in the same session fail.

.mosaic_stop_cluster() therefore calls stopCluster() and then SIGKILLs any recorded worker PID that is still a live PSOCK worker.

Why the PIDs are collected up front

The PIDs cannot be asked for at teardown time in the case that matters – clusterEvalQ(cl, Sys.getpid()) would queue behind the very task that is stuck. make_mosaic_cluster() therefore records them on the cluster object as the "mosaic_worker_pids" attribute at creation, while every worker is known to be idle, and .mosaic_stop_cluster() reads that attribute. For a cluster built elsewhere (no attribute) it falls back to querying, which works unless a worker is already wedged – the same behaviour as before this function existed, never worse.

Scope

The kill is Linux-only (it reads /proc/<pid>/cmdline) and fires only for processes whose command line still contains RSOCK. On any other platform, and for FORK workers (whose command line is the parent's), this degrades to plain stopCluster().