研究了下multiprocessing.managers,略有收获,随笔一篇;
核心思路是构造一个manager进程,这个进程可以通过unix socket或tcp socket与其它进程通信;因为利用了socket,所以通信的进程间不要求具备父子关系,甚至可以跨主机(通过tcp socket);
通过manager进行数据结构共享,可以应用于很多的IPC场景;这里只做一个示例,在manager内维护一个队列,做为生产者消费者模型中的消息队列;
# coding:utf-8 import os import time import subprocess from multiprocessing import Process #from multiprocessing.managers import SyncManager from multiprocessing.managers import BaseManager from multiprocessing import JoinableQueue #父进程 pid = os.getpid() print("parent pid %d" % pid) class TestManager(BaseManager): pass #用于共享的队列 jqueue = JoinableQueue() #用于获取队列的方法 def _get_queue(): return jqueue #启动server(manager) def make_server(port, authkey): # 注册rpc,获取共享的队列(的代理,这段代码的实现很有意思,建议看看源码) TestManager.register("get_queue",callable=lambda : _get_queue()) # 如果要走tcp socket,就用这一行 #manager = TestManager(address=('', port), authkey=authkey) # 如果要走unix socket,就用这一行 manager = TestManager(authkey=authkey) # 启动server(manager)进程 manager.start() return manager # consumer进程的入口 def do_consume(manager): print ("consumer pid %d" % os.getpid()) queue=manager.get_queue() count = 0 while True: item = queue.get(block=True) #print(item) #time.sleep(1) queue.task_done() if item is None: conn = queue._tls.connection break count += 1 print ("done consuming %d" % count) #构造新的manager实例做为client连接manager server def make_client(address, authkey): # 注册rpc,用于非父子进程环境时,新构造的manager识别rpc方法 TestManager.register("get_queue") manager = TestManager(address=address, authkey=authkey) manager.connect() return manager # producer进程的入口 def do_produce(address, authkey): print ("producer pid %d" % os.getpid()) client=make_client(address, authkey) queue=client.get_queue() for i in range(10000): queue.put(i, block=True) print ("done producing") # terminator进程的入口 def do_terminate(address, authkey): client=make_client(address, authkey) queue=client.get_queue() queue.put(None, block=True) authkey = b'foo' manager=make_server(6666, authkey) address = manager._address # 查看manager的进程号 print ("manager pid %d" % manager._process.ident) # 通过父子进程变量传递的方式,向consumer进程传递manager consumer = Process(target=do_consume, args=(manager, )) consumer.start() # 伪造非父子进程传递address和authkey的方式,向producer进程传递连接manager需要的信息 producer = Process(target=do_produce, args=(address, authkey)) producer.start() # 查看当前的进程树 status, output = subprocess.getstatusoutput('pstree -p %d' % pid) print (output) producer.join() # 伪造非父子进程传递address和authkey的方式,再启动一个terminator进程结束通信 terminator = Process(target=do_terminate, args=(address, authkey)) terminator.start() terminator.join() consumer.join()
以上示例代码的过程如下:
- 构造manager server;
- 构造一个consumer进程,直接从父进程获取到manager对象;
- 再构造一个producer进程,通过传递address和authkey,新构造manager client,并连接manager server;
- producer获取到共享队列,生产消息;
- consumer获取到共享队列,消费消息;
- terminator(producer)生产一个空消息;
- consumer获取到空消息,消费结束;
结果:
parent pid 30460 manager pid 30461 consumer pid 30463 producer pid 30464 python(30460)-+-pstree(30465) |-python(30461)-+-{python}(30462) | |-{python}(30472) | |-{python}(30474) | `-{python}(30475) |-python(30463) `-python(30464) done producing done consuming 10000
可以看到共生成4个子进程,一个manager server、一个consumer、一个producer、还有一个pstree查看进程树;