logrotate (USR1 signal) restarts uvicorn worker

[euri10]

Discussed in #1558

^{Originally posted by wonjoonSeol-WS July 7, 2022}
I Use FastAPI+Uvicorn+Gunicorn and issuing USR1 signal to uvicorn worker restarts the worker.

I believe the issue is to do with uvicorn worker not implementing signal handler - github.com/encode/uvicorn/blob/….

    def init_signals(self) -> None:
        # Reset signals so Gunicorn doesn't swallow subprocess return codes
        # other signals are set up by Server.install_signal_handlers()
        # See: https://github.com/encode/uvicorn/issues/894
        for s in self.SIGNALS:
            signal.signal(s, signal.SIG_DFL)

It converts it to signal.SIG_DFL.

SIGUSR1 and SIGUSR2 both have the default action Term -- the process is terminated. (unix.stackexchange.com/questions/38589/…)

According to a similar issue page, #896 (comment)
yinkh changed init_signals to gunicorn worker's original will not restart the worker.

I also notice that Uvicorn only handles two signals:

HANDLED_SIGNALS = (
    signal.SIGINT,  # Unix signal 2. Sent by Ctrl+C.
    signal.SIGTERM,  # Unix signal 15. Sent by `kill <pid>`.
)

I would be more than happy to contribute to uvicorn, if I can change the worker init_signals function to do the same with gunicorn worker?

    def init_signals(self):
        # reset signaling
        for s in self.SIGNALS:
            signal.signal(s, signal.SIG_DFL)
        # init new signaling
        signal.signal(signal.SIGQUIT, self.handle_quit)
        signal.signal(signal.SIGTERM, self.handle_exit)
        signal.signal(signal.SIGINT, self.handle_quit)
        signal.signal(signal.SIGWINCH, self.handle_winch)
        signal.signal(signal.SIGUSR1, self.handle_usr1)
        signal.signal(signal.SIGABRT, self.handle_abort)

        # Don't let SIGTERM and SIGUSR1 disturb active requests
        # by interrupting system calls
        signal.siginterrupt(signal.SIGTERM, False)
        signal.siginterrupt(signal.SIGUSR1, False)

        if hasattr(signal, 'set_wakeup_fd'):
            signal.set_wakeup_fd(self.PIPE[1])
```</div>

[Kludex]

Uvicorn's reload strategy is different from gunicorn. Something about this is written on #1205.

[euri10]

it's not about uvicorn reload strategy, it's about the fact our UvicornWorker is disabling a signal that gunicorn uses to reload its workers and trigger logrotate, see https://docs.gunicorn.org/en/latest/deploy.html?highlight=logrotate#logging

[Kludex]

Sorry. Wrong signal 😅 HUP is the one for reload.

I'm going to check this. 👍

[wonjoonSeol-WS]

Steps to reproduce

Minimal Fastapi code

import logging
import signal
import time
from logging import handlers

from fastapi import FastAPI
from starlette.responses import PlainTextResponse

logging.basicConfig(level=logging.DEBUG, filename='./log.txt')
logger = logging.getLogger()
log_handler = logging.handlers.WatchedFileHandler('./log.txt')
logger.addHandler(log_handler)

app = FastAPI(title="uvicorn-test")

logger = logging.getLogger(__name__)


@app.on_event("startup")
async def on_startup():
    # signal.signal(signal.SIGUSR1, signal.SIG_IGN)
    ctime = time.ctime()
    logger.debug(f"uvicorn-test start up {ctime}")


@app.on_event("shutdown")
async def on_shutdown():
    ctime = time.ctime()
    logger.debug(f"uvicorn-test shutdown {ctime}")


@app.get("/health_check", response_class=PlainTextResponse)
def health_check():
    ctime = time.ctime()
    logger.debug(f"logging test {ctime}")

Run with gunicorn

gunicorn -w 1 -k uvicorn.workers.UvicornWorker --pid /tmp/log/gunicorn.pid main:app

Set up logging

while [ 0 ]; do
    curl 'http://localhost:8000/health_check'
    echo
    sleep 1
done

Set up logrotate

/etc/logrotate.d/uvicorn-test

{file_location.txt}
{
    daily
    dateext
    rotate 30
    missingok
    notifempty
    create 0644 user group
    compress
    delaycompress
    postrotate
        kill -USR1 $(cat /tmp/log/gunicorn.pid)
    endscript
}

Send USR1 signal

kill -USR1 $(cat /tmp/log/gunicorn.pid)
or force logrotate

sudo logrotate --force /etc/logrotate.d/uvicorn-test

We can see the process is terminated:

~/w/uvicorn_mvp ❯ gunicorn -w 1 -k uvicorn.workers.UvicornWorker --pid /tmp/log/gunicorn.pid main:app
[2022-07-11 02:41:14 +0900] [22563] [INFO] Starting gunicorn 20.1.0
[2022-07-11 02:41:14 +0900] [22563] [INFO] Listening at: http://127.0.0.1:8000 (22563)
[2022-07-11 02:41:14 +0900] [22563] [INFO] Using worker: uvicorn.workers.UvicornWorker 
[2022-07-11 02:41:14 +0900] [22565] [INFO] Booting worker with pid: 22565
[2022-07-11 02:41:15 +0900] [22565] [INFO] Started server process [22565]
[2022-07-11 02:41:15 +0900] [22565] [INFO] Waiting for application startup.
[2022-07-11 02:41:15 +0900] [22565] [INFO] Application startup complete.


[2022-07-11 02:45:32 +0900] [22563] [INFO] Handling signal: usr1
[2022-07-11 02:45:32 +0900] [22563] [WARNING] Worker with pid 22565 was terminated due to signal 30 <<-
[2022-07-11 02:45:32 +0900] [23506] [INFO] Booting worker with pid: 23506
[2022-07-11 02:45:32 +0900] [23506] [INFO] Started server process [23506]
[2022-07-11 02:45:32 +0900] [23506] [INFO] Waiting for application startup.
[2022-07-11 02:45:32 +0900] [23506] [INFO] Application startup complete.

Proposal: Change uvicorn/worker.py to the following code:

    def init_signals(self) -> None:
        # Reset signals so Gunicorn doesn't swallow subprocess return codes
        # other signals are set up by Server.install_signal_handlers()
        # See: https://github.com/encode/uvicorn/issues/894
        for s in self.SIGNALS:
            signal.signal(s, signal.SIG_DFL)

        # Remove me if SIGUSR1 is handled in server.py
        signal.signal(signal.SIGUSR1, self.handle_usr1)
        # Don't let SIGUSR1 disturb active requests by interrupting system calls
        signal.siginterrupt(signal.SIGUSR1, False)

---

Re-run test

~/w/uvicorn_mvp ❯ gunicorn -w 1 -k uvicorn.workers.UvicornWorker --pid /tmp/log/gunicorn.pid main:app
[2022-07-11 02:46:42 +0900] [24005] [INFO] Starting gunicorn 20.1.0
[2022-07-11 02:46:42 +0900] [24005] [INFO] Listening at: http://127.0.0.1:8000 (24005)
[2022-07-11 02:46:42 +0900] [24005] [INFO] Using worker: uvicorn.workers.UvicornWorker
[2022-07-11 02:46:42 +0900] [24006] [INFO] Booting worker with pid: 24006
[2022-07-11 02:46:42 +0900] [24006] [INFO] Started server process [24006]
[2022-07-11 02:46:42 +0900] [24006] [INFO] Waiting for application startup.
[2022-07-11 02:46:42 +0900] [24006] [INFO] Application startup complete.
[2022-07-11 02:46:46 +0900] [24005] [INFO] Handling signal: usr1

Process is no longer terminated, logrotates fine and log.txt is still appended after the file rotation.

tail -F log.txt
logging test Mon Jul 11 02:46:42 2022
logging test Mon Jul 11 02:46:43 2022
logging test Mon Jul 11 02:46:44 2022
logging test Mon Jul 11 02:46:45 2022
logging test Mon Jul 11 02:46:46 2022 <<- SIGUSR1
logging test Mon Jul 11 02:46:47 2022
logging test Mon Jul 11 02:46:48 2022

Clearly this change only impacts Uvicorn's gunicorn worker class. i.e. kill -USR1 pid after running it with uvicorn only
(uvicorn --host 0.0.0.0 --port 8000 main:app), the worker is still terminated even with the work above.

But USR1: Reopen the log files behaviour is only documented on Gunicorn and undocumented on Uvicorn master (yet) so this could be expected behaviour(?). Otherwise we would need to create an async version of handle_usr1 method in server.py.

Btw, we can't simply fully copy gunicorn's init_signals method as suggested by yinkh in #896 (comment)

As the below signals are not compatible with asyncio.

        signal.signal(signal.SIGQUIT, self.handle_quit)
        signal.signal(signal.SIGWINCH, self.handle_winch)
        signal.signal(signal.SIGABRT, self.handle_abort)

We would end up with errors like below.

~/w/uvicorn_mvp ❯ gunicorn -w 1 -k uvicorn.workers.UvicornWorker main:app
[2022-07-10 23:49:12 +0900] [63819] [INFO] Starting gunicorn 20.1.0
[2022-07-10 23:49:12 +0900] [63819] [INFO] Listening at: http://127.0.0.1:8000 (63819)
[2022-07-10 23:49:12 +0900] [63819] [INFO] Using worker: uvicorn.workers.UvicornWorker
[2022-07-10 23:49:12 +0900] [63820] [INFO] Booting worker with pid: 63820
[2022-07-10 23:49:12 +0900] [63820] [INFO] Started server process [63820]
[2022-07-10 23:49:12 +0900] [63820] [INFO] Waiting for application startup.
[2022-07-10 23:49:12 +0900] [63820] [INFO] Application startup complete.

[2022-07-10 23:49:59 +0900] [63820] [ERROR] Traceback (most recent call last):
  File "/Users/-/.pyenv/versions/3.10.2/lib/python3.10/asyncio/runners.py", line 44, in run
    return loop.run_until_complete(main)
  File "/Users/-/.pyenv/versions/3.10.2/lib/python3.10/asyncio/base_events.py", line 628, in run_until_complete
    self.run_forever()
  File "/Users/-/.pyenv/versions/3.10.2/lib/python3.10/asyncio/base_events.py", line 595, in run_forever
    self._run_once()
  File "/Users/-/.pyenv/versions/3.10.2/lib/python3.10/asyncio/base_events.py", line 1845, in _run_once
    event_list = self._selector.select(timeout)
  File "/Users/-/.pyenv/versions/3.10.2/lib/python3.10/selectors.py", line 562, in select
    kev_list = self._selector.control(None, max_ev, timeout)
  File "/Users/-/workspace/uvicorn_mvp/.venv/lib/python3.10/site-packages/gunicorn/workers/base.py", line 198, in handle_quit
    sys.exit(0)
SystemExit: 0

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/Users/-/workspace/uvicorn_mvp/.venv/lib/python3.10/site-packages/starlette/routing.py", line 638, in lifespan
    await receive()
  File "/Users/-/workspace/uvicorn_mvp/.venv/lib/python3.10/site-packages/uvicorn/lifespan/on.py", line 137, in receive
    return await self.receive_queue.get()
  File "/Users/-/.pyenv/versions/3.10.2/lib/python3.10/asyncio/queues.py", line 159, in get
    await getter
asyncio.exceptions.CancelledError

[2022-07-10 23:49:59 +0900] [63820] [INFO] Worker exiting (pid: 63820)
[2022-07-10 23:49:59 +0900] [64039] [INFO] Booting worker with pid: 64039
[2022-07-10 23:49:59 +0900] [64039] [INFO] Started server process [64039]
[2022-07-10 23:49:59 +0900] [64039] [INFO] Waiting for application startup.
[2022-07-10 23:49:59 +0900] [64039] [INFO] Application startup complete.
^C[2022-07-10 23:50:49 +0900] [63819] [INFO] Handling signal: int
[2022-07-10 23:50:49 +0900] [64039] [ERROR] Traceback (most recent call last):
  File "/Users/-/.pyenv/versions/3.10.2/lib/python3.10/asyncio/runners.py", line 44, in run
    return loop.run_until_complete(main)
  File "/Users/-/.pyenv/versions/3.10.2/lib/python3.10/asyncio/base_events.py", line 628, in run_until_complete
    self.run_forever()
  File "/Users/-/.pyenv/versions/3.10.2/lib/python3.10/asyncio/base_events.py", line 595, in run_forever
    self._run_once()
  File "/Users/-/.pyenv/versions/3.10.2/lib/python3.10/asyncio/base_events.py", line 1845, in _run_once
    event_list = self._selector.select(timeout)
  File "/Users/-/.pyenv/versions/3.10.2/lib/python3.10/selectors.py", line 562, in select
    kev_list = self._selector.control(None, max_ev, timeout)
  File "/Users/-/workspace/uvicorn_mvp/.venv/lib/python3.10/site-packages/gunicorn/workers/base.py", line 198, in handle_quit
    sys.exit(0)
SystemExit: 0
...

Testing other signals such as TTIN, TTOU, QUIT, INT, TERM on gunicorn's signal handling page (https://docs.gunicorn.org/en/stable/signals.html) seem to work fine.

I have no idea if these two lines from gunicorn worker base class https://github.com/benoitc/gunicorn/blob/027f04b4b4aee4f50b980a7158add0feaf4c1b29/gunicorn/workers/base.py#L185
are needed though. It seem to work fine without it. So I excluded it for now:

        if hasattr(signal, 'set_wakeup_fd'):
            signal.set_wakeup_fd(self.PIPE[1])

If there is no problem with the above change. I'd like to make a PR. I don't seem to have permission to push a branch now though.

Thanks.

[Kludex]

You can fork the project and create a branch there. When the PR creation window opens, it will target the main project.

[Kludex]

Closed by #1565.

It will be available on the next uvicorn release. We are on 0.18.3 at the moment of this writing.