GearmanWorker的多进程实现

前言

因为项目原因选择了gearman作为任务委派的中间件，但原生的python拓展包只支持单进程，期间为了将gearman改造成自适应多进程的方式在实现方式上走进了些误区，故在此记录这些误区的坑以及目前的最优解决方案。

实现思路

实现方式

1. 主进程接收任务，子进程处理任务。以一个主进程作为任务委派的接收进程，接收到任务后将任务分派给子进程进行处理，处理完成后由该子进程直接返回任务结果给gearman。
2. 多进程接收并处理任务。批量fork多个子进程注册任务，子进程间互不影响，各自完成接收、处理任务的过程。

先说说第一种实现方式的优缺点

优点：

• 由于worker较多的时间是消耗在等待接收请求上，因此主进程只单一的进行轮训任务接收可以提高单条gearman请求通道的利用率。
• 由子进程直接返回任务结果可分离主进程与子进程的作业，主进程无需关心任务的结果而只专注于接收任务。

缺点：

• 主进程接收到任务请求后将请求转发给子进程处理任务，由于子进程处理任务完成后需要将任务结果返回给gearman，因此子进程需要将该任务请求对应的gearman socket传递给子进程，而该过程实现起来过于复杂。（通常具有socket的实例无法通过pickle传递给子进程，虽然Unix的sendmsg可以用来传递socket，但将传递的socket构造成一个GearmanWorker又是另外一件痛苦的事情）
• 子进程通过传递的socket构造出GearmanWorker后，由于原socket的句柄仍被父进程持有，所以在等待结果的任务请求方无法收到子进程所返回的处理结果。

再来说说第二种实现方式的优缺点

优点：

• 等价于fork多个原进程，逻辑、作业方式均无改变。
• 可在fork子进程之前完成公有资源的加载而无需每个GearmanWorker匀加载一次。

缺点：

• 子进程异常退出后主进程无法正确感知，虽然主进程会维持相同的子进程数，但是异常退出所重启的子进程没有正确注册到gearman接收任务。
• 主进程异常退出后子进程无法感知，将导致出现僵尸进程。

选择第二种方案，针对缺点的解决办法

1. 利用PID文件记录每个子进程的pid，确保主进程退出后仍能通过PID文件退出子进程。
2. 利用Redis的发布订阅模式实现GearmanWorker的正常退出。

client 端代码： test_multi_gearman_worker.py

#!/usr/bin/env python
from gearman import GearmanClient

gearman_client = GearmanClient(['10.10.13.5:4730'])


new_jobs = [
    dict(task='test_multi_gearman_worker', data='job-0.mp4'),
    dict(task='test_multi_gearman_worker', data='job-1.mp4'),
    dict(task='test_multi_gearman_worker', data='job-2.mp4'),
    dict(task='test_multi_gearman_worker', data='job-3.mp4'),
    dict(task='test_multi_gearman_worker', data='job-4.mp4'),
    dict(task='test_multi_gearman_worker', data='job-5.mp4'),
    dict(task='test_multi_gearman_worker', data='job-6.mp4'),
    dict(task='test_multi_gearman_worker', data='job-7.mp4'),
    dict(task='test_multi_gearman_worker', data='job-8.mp4'),
    dict(task='test_multi_gearman_worker', data='job-9.mp4'),
    dict(task='test_multi_gearman_worker', data='job-10.mp4'),
]


completed_requests = gearman_client.submit_multiple_jobs(new_jobs)
for current_request in completed_requests:
    print (current_request.result)


print ("Game over!")

worker端代码： multi-gearman-worker.py

#!/usr/bin/env python

import os
import string
import signal
import threading
import subprocess
import multiprocessing

import redis
from gearman.worker import GearmanWorker, POLL_TIMEOUT_IN_SECONDS

WORKER_PROCESS_PID = '/tmp/multi_gearman_worker.pid'


class MultiGearmanWorker(GearmanWorker):
    """ multi-process gearman worker """
    def __init__(self, host_list=None, redis_host=None, redis_port=None, pid=WORKER_PROCESS_PID):
        super(MultiGearmanWorker, self).__init__(host_list=host_list)
        self.redis_host = redis_host
        self.redis_port = redis_port
        self.pid = pid

    def work(self, poll_timeout=POLL_TIMEOUT_IN_SECONDS, process=multiprocessing.cpu_count()):
        """
        start working, blocking process first
        :param poll_timeout: int ,  the connection time of gearman
        :param process: int , the number of process for working , the default is the number of cpu-core
        :return:
        """
        print('Clear last process.')
        self.gearman_worker_exit()
        print('Ready to start %d process for work.' % process)
        gm_poll = multiprocessing.Pool(process)
        for x in range(0, process):
            print('start %d child process', x)
            gm_poll.apply_async(gearman_work, (self, poll_timeout, self.pid))

        gm_poll.close()
        gm_poll.join()

        # Delete the PID file if all sub-process exit normally
        if os.path.isfile(self.pid):
            os.remove(self.pid)

        print('Multi gearman worker exit.')


    def gearman_worker_exit(self):
        """ Terminate sub-process """
        if not os.path.isfile(self.pid):
            return True

        with open(self.pid, 'r+') as f:
            for pid in f.readlines():
                pid = int(pid)
                try:
                    os.kill(pid, signal.SIGKILL)
                    print('Kill process %d.' % pid)
                except OSError:
                    print('Process %d not exists' % pid)
                    continue

        os.remove(self.pid)
        print('Remove process pid file.')
        return True


#The gearman job switch identifier used by the child process
GEARMAN_CONTINUE_WORK = True


def gearman_work(gm_worker, poll_timeout=POLL_TIMEOUT_IN_SECONDS, pid=WORKER_PROCESS_PID):
    """ Open gearman;s worker in multiple processes  """
    try:
        """ Record the child process pid,
              so that the main process is cleared by the supervisor
        to clear the child process did not exit the last time
        """
        with open(pid, 'a+') as f:
            f.write("%d%s" % (os.getpid(), os.linesep))

        print('Child process start for work.')
        continue_working = True
        worker_connections = []
        d = threading.Thread(name='monitor', target=gearman_monitor,
                             args=(gm_worker.redis_host, gm_worker.redis_port))
        d.start()

        def continue_while_connections_alive(any_activity):
            return gm_worker.after_poll(any_activity)

        # Shuffle our connections after the poll timeout
        while continue_working and GEARMAN_CONTINUE_WORK:
            worker_connections = gm_worker.establish_worker_connections()
            continue_working = gm_worker.poll_connections_until_stopped(
                worker_connections, continue_while_connections_alive, timeout=poll_timeout)

        # If we were kicked out of the worker loop, we should shutdown all our connections
        for current_connection in worker_connections:
            current_connection.close()

        print('Gearman worker closed')
        return None
    except Exception as e:
        print(e)


def gearman_monitor(redis_host, redis_port):
    """ Listen to dynamic update instructions """
    global GEARMAN_CONTINUE_WORK
    print('Start gearman monitor.')
    while GEARMAN_CONTINUE_WORK:
        """To prevent abnormal operation caused by the thread is not monitoring the redis response after hanging up,
       exception handling on here, after an exception re-listen
        """
        try:
            sub = redis.StrictRedis(redis_host, redis_port).pubsub()
            sub.subscribe('hot')
            for i in sub.listen():
                if isinstance(i.get('data'), str):
                    if i.get('data') == 'exit':
                        # worker???????????????????
                        print('Gearman monitor receive restart signal.')
                        GEARMAN_CONTINUE_WORK = False
                        sub.unsubscribe('hot')
                        break
                # ????????,?????????gearman worker????????

        except Exception as e:
            print(e)
            try:
                sub.unsubscribe('hot')
            except Exception:
                pass

    print('Gearman monitor closed')


if __name__ == '__main__':
    def test_multi_gearman_worker(worker, job):
        print('who ', worker)
        print('do what ', job.data)
        print('Game over! ', job.data)
        return job.data




    gearman_worker = MultiGearmanWorker(('10.10.13.5:4730', ), '10.10.13.8', 6379)
    gearman_worker.register_task('test_multi_gearman_worker', test_multi_gearman_worker)
    gearman_worker.work(POLL_TIMEOUT_IN_SECONDS, 5)

注：
1. 需要先有安装好的redis; 
2. 需要先有安装好的gearman