Unhandled exception in celery freezes workers

I am running celery in docker via a redis backend. I have

• celery-beat container
• celery-worker container
• redis container

The celery worker container spawns 6 worker processes. If a celery task encounters an exception, the workers (all of them) stop consuming jobs. I tried to debug the processes a bit and it appear a single process will get stuck on pipe read and the rest on a futex call.

Debugging info:

# Worker 1
$ sudo strace -p 15959 -s 10000
strace: Process 15959 attached
read(4, ^Cstrace: Process 15959 detached
 <detached ...>

# Worker 2 through N
$ sudo strace -p 15960 -s 10000
strace: Process 15960 attached
futex(0x7f95c3f94000, FUTEX_WAIT_BITSET|FUTEX_CLOCK_REALTIME, 0, NULL, 0xffffffff^Cstrace: Process 15960 detached
 <detached ...>

$ sudo lsof -p 15958
COMMAND   PID USER   FD      TYPE DEVICE SIZE/OFF   NODE NAME
celery  15958 root  txt       REG  0,197    32248 264184 /usr/local/bin/python3.5
...
celery  15958 root    4r     FIFO   0,12      0t0 348559 pipe  # frozen here

Strangely, the workers will stay in a frozen state permanently, unless 1 of 2 things happen:

1. Restart the workers (docker restart celery-worker)
2. Jump start celery.

The "jump start celery" I find amusing. By issuing this command all the workers "wake up" and spring back to life and start consuming tasks until the next exception.

docker exec -it celery-worker celery -A CELERY_APP inspect active

Here the worker sprang back to life.

$ sudo strace -p 15958 -s 10000
strace: Process 15958 attached
read(4, "\0\0\3\36", 4)                 = 4
read(4, "\200\3K\2(Mj\nNccelery.app.trace\n_fast_trace_task\nq\0(X\"\0\0\0tasks.status_taskq\1X$\0\0\00071bf9972-cf5b-4a20-a8b7-ce4d7921fe0dq\2}q\3(X\t\0\0\0parent_idq\4NX\4\0\0\0langq\5X\2\0\0\0pyq\6X\3\0\0\0etaq\7NX\5\0\0\0groupq\10NX\7\0\0\0expiresq\tNX\t\0\0\0timelimitq\n]q\v(NNeX\6\0\0\0originq\fX\21\0\0\0gen1@03e7668436e5q\rX\10\0\0\0argsreprq\16X\2\0\0\0()q\17X\n\0\0\0kwargsreprq\20X\2\0\0\0{}q\21X\10\0\0\0reply_toq\22X$\0\0\0005ad0db0b-a759-375c-b173-07598914633eq\23X\4\0\0\0taskq\24h\1X\16\0\0\0correlation_idq\25X$\0\0\00071bf9972-cf5b-4a20-a8b7-ce4d7921fe0dq\26X\7\0\0\0root_idq\27X$\0\0\00071bf9972-cf5b-4a20-a8b7-ce4d7921fe0dq\30X\7\0\0\0retriesq\31K\0X\r\0\0\0delivery_infoq\32}q\33(X\10\0\0\0priorityq\34K\0X\10\0\0\0exchangeq\35X\0\0\0\0q\36X\v\0\0\0redeliveredq\37NX\v\0\0\0routing_keyq X\6\0\0\0celeryq!uX\6\0\0\0shadowq\"NX\2\0\0\0idq#h\2uCM[[], {}, {\"chord\": null, \"chain\": null, \"errbacks\": null, \"callbacks\": null}]q$X\20\0\0\0application/jsonq%X\5\0\0\0utf-8q&tq'}q(tq)\206q*.", 798) = 798
futex(0x7f95c3f94000, FUTEX_WAKE, 1)    = 1
write(7, "\0\0\0\34\200\3K\0(Mj\nNG@\327\204T\213\21\21\\K\nNtq\0\206q\1.", 32) = 32
getpid()                                = 10

Any idea why this is? Is this a bug? Is there something I can configure so celery does not hang on task exception?

I am using eventlet and I was running the workers with the default pre-fork pool. Switching to the eventlet pool seems to have fixed the problem.

Ie

celery worker -A CELERY_APP --pool eventlet