day11

Day11 - 异步IO\数据库\队列\缓存

 

本节内容html

  1. Gevent协程
  2. Select\Poll\Epoll异步IO与事件驱动
  3. Python链接Mysql数据库操做
  4. RabbitMQ队列
  5. Redis\Memcached缓存
  6. Paramiko SSH
  7. Twsited网络框架

 

 

引子

到目前为止,咱们已经学了网络并发编程的2个套路, 多进程,多线程,这哥俩的优点和劣势都很是的明显,咱们一块儿来回顾下python

 

 

 

协程

协程,又称微线程,纤程。英文名Coroutine。一句话说明什么是线程:协程是一种用户态的轻量级线程git

协程拥有本身的寄存器上下文和栈。协程调度切换时,将寄存器上下文和栈保存到其余地方,在切回来的时候,恢复先前保存的寄存器上下文和栈。所以:程序员

协程能保留上一次调用时的状态(即全部局部状态的一个特定组合),每次过程重入时,就至关于进入上一次调用的状态,换种说法:进入上一次离开时所处逻辑流的位置。github

 

协程的好处:sql

  • 无需线程上下文切换的开销
  • 无需原子操做锁定及同步的开销
    •   "原子操做(atomic operation)是不须要synchronized",所谓原子操做是指不会被线程调度机制打断的操做;这种操做一旦开始,就一直运行到结束,中间不会有任何 context switch (切换到另外一个线程)。原子操做能够是一个步骤,也能够是多个操做步骤,可是其顺序是不能够被打乱,或者切割掉只执行部分。视做总体是原子性的核心。
  • 方便切换控制流,简化编程模型
  • 高并发+高扩展性+低成本:一个CPU支持上万的协程都不是问题。因此很适合用于高并发处理。

 

缺点:数据库

  • 没法利用多核资源:协程的本质是个单线程,它不能同时将 单个CPU 的多个核用上,协程须要和进程配合才能运行在多CPU上.固然咱们平常所编写的绝大部分应用都没有这个必要,除非是cpu密集型应用。
  • 进行阻塞(Blocking)操做(如IO时)会阻塞掉整个程序

使用yield实现协程操做例子    编程

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
import  time
import  queue
def  consumer(name):
     print ( "--->starting eating baozi..." )
     while  True :
         new_baozi  =  yield
         print ( "[%s] is eating baozi %s"  %  (name,new_baozi))
         #time.sleep(1)
 
def  producer():
 
     =  con.__next__()
     =  con2.__next__()
     =  0
     while  n <  5 :
         + = 1
         con.send(n)
         con2.send(n)
         print ( "\033[32;1m[producer]\033[0m is making baozi %s"  % n )
 
 
if  __name__  = =  '__main__' :
     con  =  consumer( "c1" )
     con2  =  consumer( "c2" )
     =  producer()

看楼上的例子,我问你这算不算作是协程呢?你说,我他妈哪知道呀,你前面说了一堆废话,可是并没告诉我协程的标准形态呀,我腚眼一想,以为你说也对,那好,咱们先给协程一个标准定义,即符合什么条件就能称之为协程:数组

  1. 必须在只有一个单线程里实现并发
  2. 修改共享数据不需加锁
  3. 用户程序里本身保存多个控制流的上下文栈
  4. 一个协程遇到IO操做自动切换到其它协程

基于上面这4点定义,咱们刚才用yield实现的程并不能算是合格的线程,由于它有一点功能没实现,哪一点呢?缓存

 

 

Gevent 

Gevent 是一个第三方库,能够轻松经过gevent实现并发同步或异步编程,在gevent中用到的主要模式是Greenlet, 它是以C扩展模块形式接入Python的轻量级协程。 Greenlet所有运行在主程序操做系统进程的内部,但它们被协做式地调度。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
import  gevent
 
def  func1():
     print ( '\033[31;1m小闯在跟小涛搞...\033[0m' )
     gevent.sleep( 2 )
     print ( '\033[31;1m小闯又回去跟继续跟小涛搞...\033[0m' )
 
def  func2():
     print ( '\033[32;1m小闯切换到了跟小龙搞...\033[0m' )
     gevent.sleep( 1 )
     print ( '\033[32;1m小闯搞完了小涛,回来继续跟小龙搞...\033[0m' )
 
 
gevent.joinall([
     gevent.spawn(func1),
     gevent.spawn(func2),
     #gevent.spawn(func3),
])

  

 

输出:

小闯在跟小涛搞...
小闯切换到了跟小龙搞...
小闯搞完了小涛,回来继续跟小龙搞...
小闯又回去跟继续跟小涛搞...

 

 

 

同步与异步的性能区别 

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
import  gevent
 
def  task(pid):
     """
     Some non-deterministic task
     """
     gevent.sleep( 0.5 )
     print ( 'Task %s done'  %  pid)
 
def  synchronous():
     for  in  range ( 1 , 10 ):
         task(i)
 
def  asynchronous():
     threads  =  [gevent.spawn(task, i)  for  in  range ( 10 )]
     gevent.joinall(threads)
 
print ( 'Synchronous:' )
synchronous()
 
print ( 'Asynchronous:' )
asynchronous()

上面程序的重要部分是将task函数封装到Greenlet内部线程的gevent.spawn。 初始化的greenlet列表存放在数组threads中,此数组被传给gevent.joinall 函数,后者阻塞当前流程,并执行全部给定的greenlet。执行流程只会在 全部greenlet执行完后才会继续向下走。  

遇到IO阻塞时会自动切换任务

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
from  gevent  import  monkey; monkey.patch_all()
import  gevent
from   urllib.request  import  urlopen
 
def  f(url):
     print ( 'GET: %s'  %  url)
     resp  =  urlopen(url)
     data  =  resp.read()
     print ( '%d bytes received from %s.'  %  ( len (data), url))
 
gevent.joinall([
         gevent.spawn(f,  'https://www.python.org/' ),
         gevent.spawn(f,  'https://www.yahoo.com/' ),
         gevent.spawn(f,  'https://github.com/' ),
])

 

经过gevent实现单线程下的多socket并发

server side 

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
import  sys
import  socket
import  time
import  gevent
 
from  gevent  import  socket,monkey
monkey.patch_all()
 
 
def  server(port):
     =  socket.socket()
     s.bind(( '0.0.0.0' , port))
     s.listen( 500 )
     while  True :
         cli, addr  =  s.accept()
         gevent.spawn(handle_request, cli)
 
 
 
def  handle_request(conn):
     try :
         while  True :
             data  =  conn.recv( 1024 )
             print ( "recv:" , data)
             conn.send(data)
             if  not  data:
                 conn.shutdown(socket.SHUT_WR)
 
     except  Exception as  ex:
         print (ex)
     finally :
         conn.close()
if  __name__  = =  '__main__' :
     server( 8001 )

  

client side   

1
2
3
4
5
6
7
8
9
10
11
12
13
14
import  socket
 
HOST  =  'localhost'     # The remote host
PORT  =  8001            # The same port as used by the server
=  socket.socket(socket.AF_INET, socket.SOCK_STREAM)
s.connect((HOST, PORT))
while  True :
     msg  =  bytes( input ( ">>:" ),encoding = "utf8" )
     s.sendall(msg)
     data  =  s.recv( 1024 )
     #print(data)
 
     print ( 'Received' repr (data))
s.close()
复制代码
import socket
import threading

def sock_conn():

    client = socket.socket()

    client.connect(("localhost",8001))
    count = 0
    while True:
        #msg = input(">>:").strip()
        #if len(msg) == 0:continue
        client.send( ("hello %s" %count).encode("utf-8"))

        data = client.recv(1024)

        print("[%s]recv from server:" % threading.get_ident(),data.decode()) #结果
        count +=1
    client.close()


for i in range(100):
    t = threading.Thread(target=sock_conn)
    t.start()
复制代码

 

 

  

论事件驱动与异步IO

一般,咱们写服务器处理模型的程序时,有如下几种模型:
(1)每收到一个请求,建立一个新的进程,来处理该请求;
(2)每收到一个请求,建立一个新的线程,来处理该请求;
(3)每收到一个请求,放入一个事件列表,让主进程经过非阻塞I/O方式来处理请求
上面的几种方式,各有千秋,
第(1)中方法,因为建立新的进程的开销比较大,因此,会致使服务器性能比较差,但实现比较简单。
第(2)种方式,因为要涉及到线程的同步,有可能会面临 死锁等问题。
第(3)种方式,在写应用程序代码时,逻辑比前面两种都复杂。
综合考虑各方面因素,通常广泛认为第(3)种方式是大多数 网络服务器采用的方式
 

看图说话讲事件驱动模型

在UI编程中,经常要对鼠标点击进行相应,首先如何得到鼠标点击呢?
方式一:建立一个线程,该线程一直循环检测是否有鼠标点击,那么这个方式有如下几个缺点
1. CPU资源浪费,可能鼠标点击的频率很是小,可是扫描线程仍是会一直循环检测,这会形成不少的CPU资源浪费;若是扫描鼠标点击的接口是阻塞的呢?
2. 若是是堵塞的,又会出现下面这样的问题,若是咱们不但要扫描鼠标点击,还要扫描键盘是否按下,因为扫描鼠标时被堵塞了,那么可能永远不会去扫描键盘;
3. 若是一个循环须要扫描的设备很是多,这又会引来响应时间的问题;
因此,该方式是很是很差的。

方式二:就是事件驱动模型
目前大部分的UI编程都是事件驱动模型,如不少UI平台都会提供onClick()事件,这个事件就表明鼠标按下事件。事件驱动模型大致思路以下:
1. 有一个事件(消息)队列;
2. 鼠标按下时,往这个队列中增长一个点击事件(消息);
3. 有个循环,不断从队列取出事件,根据不一样的事件,调用不一样的函数,如onClick()、onKeyDown()等;
4. 事件(消息)通常都各自保存各自的处理函数指针,这样,每一个消息都有独立的处理函数;

 

 

 

事件驱动编程是一种编程范式,这里程序的执行流由外部事件来决定。它的特色是包含一个事件循环,当外部事件发生时使用回调机制来触发相应的处理。另外两种常见的编程范式是(单线程)同步以及多线程编程。

让咱们用例子来比较和对比一下单线程、多线程以及事件驱动编程模型。下图展现了随着时间的推移,这三种模式下程序所作的工做。这个程序有3个任务须要完成,每一个任务都在等待I/O操做时阻塞自身。阻塞在I/O操做上所花费的时间已经用灰色框标示出来了。

 

在单线程同步模型中,任务按照顺序执行。若是某个任务由于I/O而阻塞,其余全部的任务都必须等待,直到它完成以后它们才能依次执行。这种明确的执行顺序和串行化处理的行为是很容易推断得出的。若是任务之间并无互相依赖的关系,但仍然须要互相等待的话这就使得程序没必要要的下降了运行速度。

在多线程版本中,这3个任务分别在独立的线程中执行。这些线程由操做系统来管理,在多处理器系统上能够并行处理,或者在单处理器系统上交错执行。这使得当某个线程阻塞在某个资源的同时其余线程得以继续执行。与完成相似功能的同步程序相比,这种方式更有效率,但程序员必须写代码来保护共享资源,防止其被多个线程同时访问。多线程程序更加难以推断,由于这类程序不得不经过线程同步机制如锁、可重入函数、线程局部存储或者其余机制来处理线程安全问题,若是实现不当就会致使出现微妙且使人痛不欲生的bug。

在事件驱动版本的程序中,3个任务交错执行,但仍然在一个单独的线程控制中。当处理I/O或者其余昂贵的操做时,注册一个回调到事件循环中,而后当I/O操做完成时继续执行。回调描述了该如何处理某个事件。事件循环轮询全部的事件,当事件到来时将它们分配给等待处理事件的回调函数。这种方式让程序尽量的得以执行而不须要用到额外的线程。事件驱动型程序比多线程程序更容易推断出行为,由于程序员不须要关心线程安全问题。

当咱们面对以下的环境时,事件驱动模型一般是一个好的选择:

  1. 程序中有许多任务,并且…
  2. 任务之间高度独立(所以它们不须要互相通讯,或者等待彼此)并且…
  3. 在等待事件到来时,某些任务会阻塞。

当应用程序须要在任务间共享可变的数据时,这也是一个不错的选择,由于这里不须要采用同步处理。

网络应用程序一般都有上述这些特色,这使得它们可以很好的契合事件驱动编程模型。

 

此处要提出一个问题,就是,上面的事件驱动模型中,只要一遇到IO就注册一个事件,而后主程序就能够继续干其它的事情了,只到io处理完毕后,继续恢复以前中断的任务,这本质上是怎么实现的呢?下面咱们就来一块儿揭开这神秘的面纱。。。。

 

Select\Poll\Epoll异步IO 

http://www.cnblogs.com/alex3714/p/4372426.html 

番外篇 http://www.cnblogs.com/alex3714/articles/5876749.html 

select 多并发socket 例子

复制代码
#_*_coding:utf-8_*_
__author__ = 'Alex Li'

import select
import socket
import sys
import queue


server = socket.socket()
server.setblocking(0)

server_addr = ('localhost',10000)

print('starting up on %s port %s' % server_addr)
server.bind(server_addr)

server.listen(5)


inputs = [server, ] #本身也要监测呀,由于server自己也是个fd
outputs = []

message_queues = {}

while True:
    print("waiting for next event...")

    readable, writeable, exeptional = select.select(inputs,outputs,inputs) #若是没有任何fd就绪,那程序就会一直阻塞在这里

    for s in readable: #每一个s就是一个socket

        if s is server: #别忘记,上面咱们server本身也当作一个fd放在了inputs列表里,传给了select,若是这个s是server,表明server这个fd就绪了,
            #就是有活动了, 什么状况下它才有活动? 固然 是有新链接进来的时候 呀
            #新链接进来了,接受这个链接
            conn, client_addr = s.accept()
            print("new connection from",client_addr)
            conn.setblocking(0)
            inputs.append(conn) #为了避免阻塞整个程序,咱们不会马上在这里开始接收客户端发来的数据, 把它放到inputs里, 下一次loop时,这个新链接
            #就会被交给select去监听,若是这个链接的客户端发来了数据 ,那这个链接的fd在server端就会变成就续的,select就会把这个链接返回,返回到
            #readable 列表里,而后你就能够loop readable列表,取出这个链接,开始接收数据了, 下面就是这么干 的

            message_queues[conn] = queue.Queue() #接收到客户端的数据后,不马上返回 ,暂存在队列里,之后发送

        else: #s不是server的话,那就只能是一个 与客户端创建的链接的fd了
            #客户端的数据过来了,在这接收
            data = s.recv(1024)
            if data:
                print("收到来自[%s]的数据:" % s.getpeername()[0], data)
                message_queues[s].put(data) #收到的数据先放到queue里,一会返回给客户端
                if s not  in outputs:
                    outputs.append(s) #为了避免影响处理与其它客户端的链接 , 这里不马上返回数据给客户端


            else:#若是收不到data表明什么呢? 表明客户端断开了呀
                print("客户端断开了",s)

                if s in outputs:
                    outputs.remove(s) #清理已断开的链接

                inputs.remove(s) #清理已断开的链接

                del message_queues[s] ##清理已断开的链接


    for s in writeable:
        try :
            next_msg = message_queues[s].get_nowait()

        except queue.Empty:
            print("client [%s]" %s.getpeername()[0], "queue is empty..")
            outputs.remove(s)

        else:
            print("sending msg to [%s]"%s.getpeername()[0], next_msg)
            s.send(next_msg.upper())


    for s in exeptional:
        print("handling exception for ",s.getpeername())
        inputs.remove(s)
        if s in outputs:
            outputs.remove(s)
        s.close()

        del message_queues[s]
复制代码
复制代码
#_*_coding:utf-8_*_
__author__ = 'Alex Li'


import socket
import sys

messages = [ b'This is the message. ',
             b'It will be sent ',
             b'in parts.',
             ]
server_address = ('localhost', 10000)

# Create a TCP/IP socket
socks = [ socket.socket(socket.AF_INET, socket.SOCK_STREAM),
          socket.socket(socket.AF_INET, socket.SOCK_STREAM),
          ]

# Connect the socket to the port where the server is listening
print('connecting to %s port %s' % server_address)
for s in socks:
    s.connect(server_address)

for message in messages:

    # Send messages on both sockets
    for s in socks:
        print('%s: sending "%s"' % (s.getsockname(), message) )
        s.send(message)

    # Read responses on both sockets
    for s in socks:
        data = s.recv(1024)
        print( '%s: received "%s"' % (s.getsockname(), data) )
        if not data:
            print(sys.stderr, 'closing socket', s.getsockname() )
复制代码

 

 

selectors模块

This module allows high-level and efficient I/O multiplexing, built upon the select module primitives. Users are encouraged to use this module instead, unless they want precise control over the OS-level primitives used.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
import  selectors
import  socket
 
sel  =  selectors.DefaultSelector()
 
def  accept(sock, mask):
     conn, addr  =  sock.accept()   # Should be ready
     print ( 'accepted' , conn,  'from' , addr)
     conn.setblocking( False )
     sel.register(conn, selectors.EVENT_READ, read)
 
def  read(conn, mask):
     data  =  conn.recv( 1000 )   # Should be ready
     if  data:
         print ( 'echoing' repr (data),  'to' , conn)
         conn.send(data)   # Hope it won't block
     else :
         print ( 'closing' , conn)
         sel.unregister(conn)
         conn.close()
 
sock  =  socket.socket()
sock.bind(( 'localhost' 10000 ))
sock.listen( 100 )
sock.setblocking( False )
sel.register(sock, selectors.EVENT_READ, accept)
 
while  True :
     events  =  sel.select()
     for  key, mask  in  events:
         callback  =  key.data
         callback(key.fileobj, mask)

  

数据库操做与Paramiko模块 

http://www.cnblogs.com/wupeiqi/articles/5095821.html 

 

 

RabbitMQ队列  

安装 http://www.rabbitmq.com/install-standalone-mac.html

安装python rabbitMQ module 

1
2
3
4
5
6
7
pip install pika
or
easy_install pika
or
源码
  
https: / / pypi.python.org / pypi / pika

实现最简单的队列通讯

 

send端

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
#!/usr/bin/env python
import  pika
 
connection  =  pika.BlockingConnection(pika.ConnectionParameters(
                'localhost' ))
channel  =  connection.channel()
 
#声明queue
channel.queue_declare(queue = 'hello' )
 
#n RabbitMQ a message can never be sent directly to the queue, it always needs to go through an exchange.
channel.basic_publish(exchange = '',
                       routing_key = 'hello' ,
                       body = 'Hello World!' )
print ( " [x] Sent 'Hello World!'" )
connection.close()

receive端

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
#_*_coding:utf-8_*_
__author__  =  'Alex Li'
import  pika
 
connection  =  pika.BlockingConnection(pika.ConnectionParameters(
                'localhost' ))
channel  =  connection.channel()
 
 
#You may ask why we declare the queue again ‒ we have already declared it in our previous code.
# We could avoid that if we were sure that the queue already exists. For example if send.py program
#was run before. But we're not yet sure which program to run first. In such cases it's a good
# practice to repeat declaring the queue in both programs.
channel.queue_declare(queue = 'hello' )
 
def  callback(ch, method, properties, body):
     print ( " [x] Received %r"  %  body)
 
channel.basic_consume(callback,
                       queue = 'hello' ,
                       no_ack = True )
 
print ( ' [*] Waiting for messages. To exit press CTRL+C' )
channel.start_consuming()

  

Work Queues

在这种模式下,RabbitMQ会默认把p发的消息依次分发给各个消费者(c),跟负载均衡差很少

消息提供者代码

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
import  pika
import  time
connection  =  pika.BlockingConnection(pika.ConnectionParameters(
     'localhost' ))
channel  =  connection.channel()
 
# 声明queue
channel.queue_declare(queue = 'task_queue' )
 
# n RabbitMQ a message can never be sent directly to the queue, it always needs to go through an exchange.
import  sys
 
message  =  ' ' .join(sys.argv[ 1 :])  or  "Hello World! %s"  %  time.time()
channel.basic_publish(exchange = '',
                       routing_key = 'task_queue' ,
                       body = message,
                       properties = pika.BasicProperties(
                           delivery_mode = 2 ,   # make message persistent
                       )
                       )
print ( " [x] Sent %r"  %  message)
connection.close()

  

 

消费者代码

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
#_*_coding:utf-8_*_
 
import  pika, time
 
connection  =  pika.BlockingConnection(pika.ConnectionParameters(
     'localhost' ))
channel  =  connection.channel()
 
 
def  callback(ch, method, properties, body):
     print ( " [x] Received %r"  %  body)
     time.sleep( 20 )
     print ( " [x] Done" )
     print ( "method.delivery_tag" ,method.delivery_tag)
     ch.basic_ack(delivery_tag = method.delivery_tag)
 
 
channel.basic_consume(callback,
                       queue = 'task_queue' ,
                       no_ack = True
                       )
 
print ( ' [*] Waiting for messages. To exit press CTRL+C' )
channel.start_consuming()

  

 

此时,先启动消息生产者,而后再分别启动3个消费者,经过生产者多发送几条消息,你会发现,这几条消息会被依次分配到各个消费者身上  

Doing a task can take a few seconds. You may wonder what happens if one of the consumers starts a long task and dies with it only partly done. With our current code once RabbitMQ delivers message to the customer it immediately removes it from memory. In this case, if you kill a worker we will lose the message it was just processing. We'll also lose all the messages that were dispatched to this particular worker but were not yet handled.

But we don't want to lose any tasks. If a worker dies, we'd like the task to be delivered to another worker.

In order to make sure a message is never lost, RabbitMQ supports message acknowledgments. An ack(nowledgement) is sent back from the consumer to tell RabbitMQ that a particular message had been received, processed and that RabbitMQ is free to delete it.

If a consumer dies (its channel is closed, connection is closed, or TCP connection is lost) without sending an ack, RabbitMQ will understand that a message wasn't processed fully and will re-queue it. If there are other consumers online at the same time, it will then quickly redeliver it to another consumer. That way you can be sure that no message is lost, even if the workers occasionally die.

There aren't any message timeouts; RabbitMQ will redeliver the message when the consumer dies. It's fine even if processing a message takes a very, very long time.

Message acknowledgments are turned on by default. In previous examples we explicitly turned them off via the no_ack=True flag. It's time to remove this flag and send a proper acknowledgment from the worker, once we're done with a task.

1
2
3
4
5
6
7
8
def  callback(ch, method, properties, body):
     print  " [x] Received %r"  %  (body,)
     time.sleep( body.count( '.' ) )
     print  " [x] Done"
     ch.basic_ack(delivery_tag  =  method.delivery_tag)
 
channel.basic_consume(callback,
                       queue = 'hello' )

  Using this code we can be sure that even if you kill a worker using CTRL+C while it was processing a message, nothing will be lost. Soon after the worker dies all unacknowledged messages will be redelivered

    

消息持久化  

We have learned how to make sure that even if the consumer dies, the task isn't lost(by default, if wanna disable  use no_ack=True). But our tasks will still be lost if RabbitMQ server stops.

When RabbitMQ quits or crashes it will forget the queues and messages unless you tell it not to. Two things are required to make sure that messages aren't lost: we need to mark both the queue and messages as durable.

First, we need to make sure that RabbitMQ will never lose our queue. In order to do so, we need to declare it as durable:

1
channel.queue_declare(queue = 'hello' , durable = True )

  

Although this command is correct by itself, it won't work in our setup. That's because we've already defined a queue called hello which is not durable. RabbitMQ doesn't allow you to redefine an existing queue with different parameters and will return an error to any program that tries to do that. But there is a quick workaround - let's declare a queue with different name, for exampletask_queue:

1
channel.queue_declare(queue = 'task_queue' , durable = True )

  

This queue_declare change needs to be applied to both the producer and consumer code.

At that point we're sure that the task_queue queue won't be lost even if RabbitMQ restarts. Now we need to mark our messages as persistent - by supplying a delivery_mode property with a value 2.

1
2
3
4
5
6
channel.basic_publish(exchange = '',
                       routing_key = "task_queue" ,
                       body = message,
                       properties = pika.BasicProperties(
                          delivery_mode  =  2 # make message persistent
                       ))

消息公平分发

若是Rabbit只管按顺序把消息发到各个消费者身上,不考虑消费者负载的话,极可能出现,一个机器配置不高的消费者那里堆积了不少消息处理不完,同时配置高的消费者却一直很轻松。为解决此问题,能够在各个消费者端,配置perfetch=1,意思就是告诉RabbitMQ在我这个消费者当前消息还没处理完的时候就不要再给我发新消息了。

 

1
channel.basic_qos(prefetch_count = 1 )

 

带消息持久化+公平分发的完整代码

相关文章
相关标签/搜索