Monday, 26 November 2018

Accessing celery worker instance inside the task

I want to use jupyter kernels in side the celery worker. There will be one Jupyter Kernel for each Celery Worker.

To achieve it I am overriding the default Worker class of the celery, at the initialisation of the worker I am starting the jupyter kernel and with the stop method I am shutting down the jupyter kernel.

The current problem I am facing is how can I access that kernel instance inside the task while the task is running ?

Is there any better way to override the Worker class definition for the celery application than app.Worker = CustomWorker ?

Here is the celery config with the Custom Worker.

from __future__ import absolute_import, unicode_literals
from celery import Celery
from jupyter_client import MultiKernelManager

app = Celery('proj',
    broker='redis://',
    backend='redis://',
    include=['tasks'])

app.conf.update(
    result_expires=3600
)

class CustomWorker(app.Worker):
    def __init__(self, *args, **kwargs):
        self.km = MultiKernelManager()
        self.kernel_id = self.km.start_kernel()
        print("Custom initializing")
        self.kernel_client = km.get_kernel(kernel_id).client()
        super(CustomWorker, self).__init__(*args, **kwargs)

    def on_close(self):
        self.km.shutdown_kernel(self.kernel_id)
        super(CustomWorker, self).on_close()

app.Worker = CustomWorker

if __name__ == '__main__':
    app.start()

Here is a skeleton of tasks.py

from __future__ import absolute_import, unicode_literals
from celery import app

from celery import Task
from tornado import gen
from jupyter_client import MultiKernelManager
from zmq.eventloop import ioloop
from zmq.eventloop.zmqstream import ZMQStream
ioloop.install()

reply_futures = {}

# This is my celery task where I pass the arbitary python code to execute on
# some celery worker(actually to the corresponding kernel)
@app.task
def pythontask(code):
    # I don't know how to get the kernel_client for current celery worker !!?
    kernel_client = self.get_current_worker().kernel_client
    mid = kernel_client.execute(code)

    # defining the callback which will be executed when message arrives on
    # zmq stream
    def reply_callback(session, stream, msg_list):
        idents, msg_parts = session.feed_identities(msg_list)
        reply = session.deserialize(msg_parts)
        parent_id = reply['parent_header'].get('msg_id')
        reply_future = reply_futures.get(parent_id)
        if reply_future:
            reply_future.set_result(reply)

    @gen.coroutine
    def execute(kernel_client, code):
        msg_id = kernel_client.execute(code)
        f = reply_futures[msg_id] = Future()
        yield f
        raise gen.Return(msg_id)

    # initializing the zmq streams and attaching the callback to receive message
    # from the kernel
    shell_stream = ZMQStream(kernel_client.shell_channel.socket)
    iopub_stream = ZMQStream(kernel_client.iopub_channel.socket)
    shell_stream.on_recv_stream(partial(reply_callback, kernel_client.session))
    iopub_stream.on_recv_stream(partial(reply_callback, kernel_client.session))

    # create a IOLoop
    loop = ioloop.IOLoop.current()
    # listen on the streams
    msg_id = loop.run_sync(lambda: execute(kernel_client,code))
    print(reply_msgs[msg_id])
    reply_msgs[msg_id] = []

    # Disable callback and automatic receiving.
    shell_stream.on_recv_stream(None)
    iopub_stream.on_recv_stream(None)



from Accessing celery worker instance inside the task

No comments:

Post a Comment