Gunicorn: Missing unix socket after restart

[si14]

(can be related to #1523 or #1298)

Unix domain socket gets deleted when Gunicorn is restarted as a service:

ubuntu@host:~$ sudo service gunicorn --full-restart
ubuntu@host:~$ ls /run/gunicorn/
pid
ubuntu@host:~$ sudo service gunicorn stop
ubuntu@host:~$ sudo service gunicorn start
ubuntu@host:~$ ls /run/gunicorn/
pid  socket
ubuntu@host:~$ sudo service gunicorn restart
ubuntu@host:~$ ls /run/gunicorn/
pid
ubuntu@host:~$ sudo service gunicorn stop
ubuntu@host:~$ sudo service gunicorn start
ubuntu@host:~$ ls /run/gunicorn/
pid  socket

Gunicorn is runned by systemd, the service is configured as recommended in the docs

Full service config (slightly redacted):

[Unit]
Description=gunicorn daemon
Requires=gunicorn.socket
After=network.target

[Service]
PIDFile=/run/gunicorn/pid
User=some_user
Group=some_group
RuntimeDirectory=gunicorn
WorkingDirectory=some_path
ExecStart=venv_path/bin/gunicorn --pid /run/gunicorn/pid   \
          --workers 6  --max-requests 1000  \
          --name some_gunicorn  --statsd-host localhost:8125  \
          --bind unix:/run/gunicorn/socket app.wsgi
ExecReload=/bin/kill -s HUP $MAINPID
ExecStop=/bin/kill -s TERM $MAINPID
PrivateTmp=true

Environment=DATABASE_URL=...
Environment=DJANGO_SETTINGS_MODULE=...

[Install]
WantedBy=multi-user.target

Gunicorn version is 19.7.1

[valq7711]

Hi, I had the same problem and after having 2 days reading/studying systemd doc I got why this happens. The core problem is that unix.socket (don't confuse with gunicorn.socket) placed in the runtimedirectory that is spiced up by PrivateTmp=true, that means killing it while daemon is stoping (restart=stop & start). My solution - just place socket any where else. i.e. change in gunicorn.socket ListenStream=/run/gunicorn/socket to ListenStream=/run/gunicorn.sock (for example) and --bind in gunicorn.service (--bind=unix:/run/gunicorn.sock ).
And I added 2 extra lines to [Socket] section (just in case):

SocketUser=someuser
SocketGroup=somegroup 

[alizain]

@valq7711, thanks for the response. I spent quite some time debugging this before I found this issue.

@benoitc, is this the expected solution to working with systemd? If so, the docs should be updated.

[benoitc]

On 7 Sep 2017, at 12:45, Alizain Feerasta [e-mail] wrote:

@valq7711 https://github.com/valq7711, thanks for the response. I spent quite some time debugging this before I found this issue.

@benoitc https://github.com/benoitc, is this the expected solution to working with systemd? If so, the docs should be updated.

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub https://github.com/benoitc/gunicorn/issues/1524#issuecomment-327763919, or mute the thread https://github.com/notifications/unsubscribe-auth/AAA4okrJNK-sGMLZUn911gPDkDVsUja6ks5sf8lggaJpZM4Ny1-8.

i think @tilgovi fixed it in latest master. Can you test it?

• benoit

[alizain]

I'll take a look on the weekend, for sure!

[tilgovi]

I do not know of any change I made that would fix this that is not already on 19.7.1.

[benoitc]

@tilgovi I was thinking to https://github.com/benoitc/gunicorn/pull/1310 but that different. I misread the ticket..

Anyway I think it's expected that a socket created by gunicorn is deleted when gunicorn is stopped. Why should it survive in such case? fd inheritance is different and I'm not sure that require is actually creating that socket .in any case the fd inherited should be given differently.

[alizain]

The problem is:

• the gunicorn.socket unit creates a socket in the /run/gunicorn directory.
• when the first request comes into the socket, systemd starts up the gunicorn.service unit to respond, which wipes the /run/gunicorn directory.
• gunicorn.service complains about not having a socket file, and dies.

This happens when gunicorn is configured exactly the way it is explained in gunicorn's documentation about deploying with systemd.

I don't know enough about systemd to say that it is a functionality issue. I'm guessing it's expected behaviour with systemd. If so, gunicorn's documentation should be updated.

[benoitc]

@alizain thanks for the problem description.

I'm not a user of linux/systemd myself, I will need to check as well. Reading the current doc it seems that the socket is created via /etc/systemd/system/gunicorn.socket.

At this point I need to think what is the correct behaviour :) I will need to make some research

[benoitc]

if i'm correct systemd is the one listening on the socket so we should probably onlyuse the FD given by stemd but I maybe wrong.

[valq7711]

Hi!
http://0pointer.de/blog/projects/socket-activation.html:

Socket activation makes it possible to start all four services completely simultaneously, without any kind of ordering. Since the creation of the listening sockets is moved outside of the daemons themselves

Thus, a socket isn't private part of any service (it's only a bus that provides communication), so private dir of any service is illogical place for it (by my opinion)

[ErnstHaagsman]

I recently encountered this bug as well. Here's my project, which reproduces the issue in a Vagrant box: https://github.com/ErnstHaagsman/grouporder/tree/a14dd4078d9fd97f6a9d8350a49da309b107180a

By running:

sudo systemctl restart gunicorn
systemctl status gunicorn

It looks like gunicorn thinks the socket works ("Feb 23 18:14:25 vagrant gunicorn[22503]: [2018-02-23 18:14:25 +0000] [22503] [INFO] Listening at: unix:/run/gunicorn/gunicorn.sock (22503)")

However, ls /run/gunicorn shows that there's only a pid file, and no socket file there.

[pe224]

Can confirm that deployment with systemd as described in the current docs doesn't work.
Moving the socket as suggested by @valq7711 works for me.

[valintepes]

Ran into the same issue. In my case I set PrivateTmp=false, and was able to route to the python app on the first request each time I kill the service and restart the socket. Subsequent curls fail with nginx saying no such file or directory.

I can also confirm @valq7711 's suggestion worked.

[ykeyani]

I spent quite a few hours narrowing it down to the socket aswell, using the systemd example in the docs. It's definitely not obvious and a note should probably be added to the example.

@valq7711 's solution works.

[arianitu]

Confirmed changing path from /run/gunicorn/socket to /run/gunicorn.sock fixes the problem.

Can we get the docs updated for systemd?

[tilgovi]

@arianitu I would be happy to review and merge a PR! Look in docs/source/deploy.rst.

[arianitu]

@tilgovi opened https://github.com/benoitc/gunicorn/pull/1895