#celery

only linux can set the buffer size, so not really a solution, could be interesting to tweak for performance

and that won't help if the buffer is cleared

asksol: what if you set the buffer to 0

or some small value

for linux there's F_SETPIPE_SZ

still would not help unless the parent was constantly reading

asksol: is it hard to switch to dgram ?

asksol: how about this idea: before death, child send this "exit sentinel" and then waits on a read for something around 1 second . when the parent gets the 'exit sentinel' he sends to the children an "ok to die sentiel".

that's probably easier to do than switching to dgram

don't think it needs to wait for one second, it can just wait and exit if the fd is closed

but hmm

if the exitcode is 155

can you not be sure that the result was written?

the result is only used for logs

and that will even change in 3.2

where the logging will happen in the child

so it doesn't really need the return value of the task, it only needs to know that it completed

so it could simply log a success without an actual return value until 3.2

as in, it suceeded but we don't know what it returned, the state is still updated

rarely do you need to see the return value in logs

well it does help sometimes

asksol: hmmm

asksol: the child needs to wait on something otherwise we're back to the original race condition

what is that?

the close from child/read from parent race

what is the synq for?

we don't need the return value, only need to know if the task completed (which 155 implies)

synq is not completed, it's intended for safe termination

of tasks

yeah but i don't like to discard results

i want to explore alternatives a bit

what's the 'safe termination'

it's so that revoke+terminate means terminate the task, not the process

terminate is for manual administration, if you used improperly it may very well terminate a different task

so synq would be a feature that can be enabled

asksol: it could terminate a different task? both celery 3.0 and 3.1 ?

it's for terminating the process, not the task

not to be used programatically

but if you see a process that is stuck it's fine

asksol: about the sentinel

so i'm looking at state_handlers in pool.py:ResultHandler

i would add a new state handler for DEATH and send that token from the child

right before exit

and then i wait for confirmation (get another token from the queue)

that will work, but another vector for a deadlock

now i want to send that token from the handler in state_handlers

hmmmm, how?

how long does the child wait for confirmation?

there is no safe timeout

i suppose we could have some timeout mechanism

can't we?

what does a timeout mean?

recv should have a timeout

i mean, we can set the timeout on the pipe

and then switch the pipe to blocking

oh wait, the timeouts in python sockets aren't actual timeouts, python just selects for a while on the socket right ?

it's already blocking in the child, so I guess the timeout will result in an error for the task

well no, the result would be already be sent by then

no way to have an error that way

that's how all socket timeouts are implemented I believe

well i guess we could do the same

no big deal

now i can't figure how to put something in the queue from the state handler

i've tried to pass _quick_put to the ResultHandler constructor but turns out it's None - what the hell is going on ?

ionelmc: result handler cannot do that you need to use _write_job

or more exactly, you need to use pool.apply_async()

sending a different payload than TASK would be way a lot of work for a temporary workaround till 3.2

need to take the train home :/

well, another day then

asksol: still around?

asksol: i have found a workaround: https://github.com/celery/billiard/pull/87

please tell me there's a better way :)

hi people! does anyone have a hint about celeryd not starting from a fabric task but starting ok from "normal' cmdline?

the fab tasks executes "sudo /etc/init.d/celeryd restart" and fails - it says it can't import the module

but if I ssh into the host and run the same, it all goes well

it says "ImportError: cannot import name celeryd_multi"

I ran the same command from bash and it executes well

anyone there?

rodrigo__: must be some environment difference

maybe you have different PYTHONPATH when you run it manually from shell

ionelmc, thanks for answering

ionelmc, it's weird because it started ok before upgrading celery

rodrigo__: wouldn't know - i hate init scripts, they are hard to debug

so i use supervisord instead

ionelmc, yup, me too

but the init script works if it is called from a "regular bash"

it doesn't when it is called from fabic

this seems to be the same error: http://openhatch.org/bugs/issue613

rodrigo__: in that module (djcelery.management.commands.celeryd)

rodrigo__: maybe you can add a print statement before it imports celery.bin