werkzeug.serving.run_simple does not handle SIGTERM correctly

[masklinn]

When the process receives SIGTERM, it should shut down and exit, with as few operations as possible and without printing anything.

A werkzeug.serving.run_simple process receiving SIGTERM generally result in a return code of 141 (symptom of an un/mishandled SIGPIPE), and when using the reloader the process goes zombie (it has to be killed manually, as the port stays bound).

Adding a signal handler for SIGTERM which simply invokes sys.exit(0) is sufficient to fix the issue (in that there is no more misbehavior of the process), but I am not sure it's the actually correct fix.

[mitsuhiko]

I bind a signal handler now when run with reloader. Hope that helps.

[theduderog]

In what version is this fix? This is still a problem in Flask 0.8.

[orf]

This is still an issue, its quite annoying when using Flask with an IDE - whenever you stop debugging the process persists and continues to serve requests.

[DasIch]

I'm reopening this issue as it seems to persist, see the following discussion from IRC today.

20:20 < mcdonc> can somebody fix flask's reloader so when you send the process a sigint it actually stops the child process
20:20 < untitaker> mcdonc: it seems to work for me
20:21 < untitaker> mcdonc: it used to cause problems but for me it's fixed in latest master
20:21 < mcdonc> ah good.  i just got some number of complaints from people who run it under supervisor.
20:22 < untitaker> mcdonc: you are talking about the one from the Py3 port?
20:22 < untitaker> released versions should work properly
20:22 < mcdonc> no.. i am talking about.. well.. yes, i dont actually know what i'm talking about ;-)  i dont use it, i just get people telling me they need to send a stop signal to the entire process group instead of to the process to make sure its killed.
20:23 < mcdonc> this is not recent.. for the last year or so
20:23 < mcdonc> why people run the reloader under supervisor (presumably in production) i cannot fathom
20:23 < mcdonc> but they do
20:24 < Alex_Gaynor> mcdonc: I've toyed with using supervisord in dev, FWIW
20:24 < Alex_Gaynor> mcdonc: for cases where you don't just have web proc, you've also got background daemons and such, it could be nice
[...]
20:32 < DasIch> untitaker: the supervisor issue is independent from the threading/thread issue
20:32 < untitaker> DasIch: ah okay
20:32 < untitaker> didn't know that
20:32 < untitaker> DasIch: is the reloader behaving weird in supervisor?
20:33 < DasIch> untitaker: I guess what happens if you run the reloader in supervisor is that supervisor kill the reloading process but that doesn't kill the process started by the reloader
20:34 < untitaker> DasIch: couldn't one write a wrapper shell script that kills both?
20:34 < untitaker> at least for now
20:34 < DasIch> untitaker: I think you shouldn't use the reloader in production
20:35 < untitaker> well yeah
20:35 < asdf`> (supervisord has a 'kill as group' thing)
20:35 < DasIch> right there is that as well
20:35 < asdf`> (it even mentions the werkzeug reloader in the docs paragraph about it!)
20:36 < mcdonc> yes i put it there
20:37 < asdf`> (then you might want to fix it, because AFAIR it actually says 'flask', while the reloader is part of werkzeug. But i admit 'flask' is something more people will know)
20:37 < mcdonc> nobody reads docs anyway ;)
20:38 < DasIch> I just wanted to mention I don't care unless someone creates an issue with a valid use case for that but apparently this seems to be it https://github.com/mitsuhiko/werkzeug/issues/58
20:38 < mcdonc> like alex said, it's not entirely crazy to want to use the reloader under supervisor in dev, esp. if your app is reliant on other processes being started
20:39 < mcdonc> i actually dont run my own apps under supervisor, but that's because i don't use a reloader, i just press ctrl-c.. because i'm a savage
20:40 < DasIch> I do use the reloader but I tend to save so often with bad syntax that I end up restarting manually all the time

[daGrevis]

I think it's still relevant.

Doing os.kill(parent_id, signal.SIGTERM) doesn't kill children processes.

[untitaker]

I've encountered this issue too while reworking the testsuite for werkzeug.serving. I worked around it by killing the whole process group: https://github.com/mitsuhiko/werkzeug/blob/a00377315bbf02ec48fdad22c6bb08433fc1e9c1/tests/conftest.py#L158

[drewdogg]

I ran into this same problem in Flask with debug mode (use_debugger=True). However, I see a return code of 0 on the "parent" process. Without debug mode enabled, SIGTERM works fine and the process exits with code 143. Python 2.7.5.

[zyv]

Got hit by this bug as well, in my case docker stop was sending SIGTERM to werkzeug-powered server (moto), but server ignored and then docker killed it with SIGKILL resulting into a non-zero exit code.

The workaround was to specify SIGINT (Ctrl+C) as a preferred stop signal in Dockerfile (STOPSIGNAL SIGINT), after that the containers shut down cleanly.

[agausmann]

I have the same problem while running a Flask app inside of Docker; however, STOPSIGNAL SIGINT is still not enough to stop the container. I have to use SIGKILL.

[aenglander]

I cannot recreate the issue. I have tried this with the official Python container with the 2.7 and 3.7 tag. I used the following Dockerfile:

FROM python:2.7

WORKDIR /usr/src/app

RUN pip install click \
                werkzeug \
                sqlalchemy \
                jinja2

COPY . .

RUN python manage-shorty.py initdb

ENTRYPOINT ["python"]

CMD ["manage-shorty.py", "runserver"]

And built a container from the Dockerfile in the examples directory with the command:

 docker build -t werkzeug-examples .

I would then run the container in interactive mode and cancel with:

$ docker run -it --name werkzeug-example werkzeug-examples
 * Running on http://localhost:5000/ (Press CTRL+C to quit)
 * Restarting with stat
^C

Running docker ps showed it exited with 0:

$ docker ps -a
CONTAINER ID        IMAGE               COMMAND                  CREATED              STATUS                      PORTS               NAMES
8c708ea4ef77        werkzeug-examples   "python manage-short…"   About a minute ago   Exited (0) 58 seconds ago                       werkzeug-example

Running the container and stopping with docker stop werkzeug-example also exits with 0.

Here is the result of Docker Version on the computer I ran these commands:

Client: Docker Engine - Community
 Version:           18.09.2
 API version:       1.39
 Go version:        go1.10.8
 Git commit:        6247962
 Built:             Sun Feb 10 04:12:39 2019
 OS/Arch:           darwin/amd64
 Experimental:      false

Server: Docker Engine - Community
 Engine:
  Version:          18.09.2
  API version:      1.39 (minimum version 1.12)
  Go version:       go1.10.6
  Git commit:       6247962
  Built:            Sun Feb 10 04:13:06 2019
  OS/Arch:          linux/amd64
  Experimental:     false

Can you provide an example which can reproduce the issue you are experiencing?

[aenglander]

Until we can get a reproducible scenario, I will close this as it cannot be reproduced in the latest version of Docker and Werkzeug.