Misbehaving processes aren't killed with --race

[whymarrh]

Given the following package.json file (cruft removed):

{
  "scripts": {
    "a": "test/programs/a",
    "b": "test/programs/b",
    "sleep10s": "sleep 10"
  },
  "devDependencies": {
    "npm-run-all": "^4.1.3"
  }
}

Racing a, b, and sleep10s should terminate a and b when sleep 10 exits after 10 seconds but instead run-p hangs:

$ ./node_modules/.bin/run-p --race a b sleep10s

> example@1.0.0 a /Users/jim/example
> test/programs/a


> example@1.0.0 b /Users/jim/example
> test/programs/b


> example@1.0.0 sleep10s /Users/jim/example
> sleep 10

GRACEFUL
IGNORING SIGNAL

This is because b doesn't correctly handle SIGTERM and needs to be sent SIGKILL.

Racing tasks that do correctly respond to SIGTERM works fine:

$ ./node_modules/.bin/run-p --race a a a sleep10s

> example@1.0.0 sleep10s /Users/jim/example
> sleep 10


> example@1.0.0 a /Users/jim/example
> test/programs/a


> example@1.0.0 a /Users/jim/example
> test/programs/a


> example@1.0.0 a /Users/jim/example
> test/programs/a

GRACEFUL
GRACEFUL
GRACEFUL

What's worse is that further sending SIGINT to run-p will cause it to exit (with 130) and orphan b. The process tree while hanging:

-+= 74470 jim node ./node_modules/.bin/run-p --race a a a b sleep10s
 \-+- 74474 jim npm   
   \--- 74480 jim /usr/local/Cellar/python@2/2.7.15_1/Frameworks/Python.framework/Versions/2.7/Resources/Python.app/Contents/MacOS/Python -u test/programs/b

The process tree after SIGINT:

-+= 00001 root /sbin/launchd
 \-+- 74474 jim npm
   \--- 74480 jim /usr/local/Cellar/python@2/2.7.15_1/Frameworks/Python.framework/Versions/2.7/Resources/Python.app/Contents/MacOS/Python -u test/programs/b

Example programs

test/programs/a (correctly responds to SIGTERM):

#!/usr/bin/env python -u

from __future__ import print_function
import signal
import sys

def graceful_exit(sig, frame):
    print('GRACEFUL')
    print('GRACEFUL', file=open("was-graceful.log", "wt"))
    sys.exit(0)

signal.signal(signal.SIGINT,  graceful_exit)
signal.signal(signal.SIGTERM, graceful_exit)
signal.pause()

test/programs/b (incorrectly ignores SIGTERM):

#!/usr/bin/env python -u

from __future__ import print_function
import signal
import sys
import time

def ignore_signal(sig, frame):
    print('IGNORING SIGNAL')
    while True:
        time.sleep(42)

signal.signal(signal.SIGINT,  ignore_signal)
signal.signal(signal.SIGTERM, ignore_signal)
signal.pause()

Possible solution

Based on a bit of research, I think a more correct way to kill child processes would be to send SIGTERM to the children (not the the whole trees, just immediate children) and after some timeout send SIGKILL to the tree. This is what a few init systems will do. It should also be possible to handle signals and ensure that all child processes are cleaned up before exiting.

Related issues

• If one task crashes, others keep running #6 included a discussion of what parallel task termination should look like. SIGTERM is the correct signal to start with but it does need to be followed up with.

[whymarrh]

https://github.com/kimmobrunfeldt/concurrently/issues/155 is a similar issue with concurrently

[webartoli]

Solved installing package procps.
I have figured out that on *nix package an additional package procps is required and can be installed with:

# debian/ubuntu based distro
$ [sudo] apt install -y procps

# alpine based distro
$ apk add -y procps