先来看一段代码:
# ~*~ Twisted - A Python tale ~*~ from time import sleep # Hello, I'm a developer and I mainly setup Wordpress. def install_wordpress(customer): # Our hosting company Threads Ltd. is bad. I start installation and... print "Start installation for", customer # ...then wait till the installation finishes successfully. It is # boring and I'm spending most of my time waiting while consuming # resources (memory and some CPU cycles). It's because the process # is *blocking*. sleep(3) print "All done for", customer # I do this all day long for our customers def developer_day(customers): for customer in customers: install_wordpress(customer) developer_day(["Bill", "Elon", "Steve", "Mark"])
运行一下,结果如下所示:
$ ./deferreds.py 1
------ Running example 1 ------ Start installation for Bill All done for Bill Start installation ... * Elapsed time: 12.03 seconds
这是一段顺序执行的代码。四个消费者,为一个人安装需要3秒的时间,那么四个人就是12秒。这样处理不是很令人满意,所以看一下第二个使用了线程的例子:
import threading # The company grew. We now have many customers and I can't handle the # workload. We are now 5 developers doing exactly the same thing. def developers_day(customers): # But we now have to synchronize... a.k.a. bureaucracy lock = threading.Lock() # def dev_day(id): print "Goodmorning from developer", id # Yuck - I hate locks... lock.acquire() while customers: customer = customers.pop(0) lock.release() # My Python is less readable install_wordpress(customer) lock.acquire() lock.release() print "Bye from developer", id # We go to work in the morning devs = [threading.Thread(target=dev_day, args=(i,)) for i in range(5)] [dev.start() for dev in devs] # We leave for the evening [dev.join() for dev in devs] # We now get more done in the same time but our dev process got more # complex. As we grew we spend more time managing queues than doing dev # work. We even had occasional deadlocks when processes got extremely # complex. The fact is that we are still mostly pressing buttons and # waiting but now we also spend some time in meetings. developers_day(["Customer %d" % i for i in xrange(15)])
运行一下:
$ ./deferreds.py 2
------ Running example 2 ------ Goodmorning from developer 0Goodmorning from developer 1Start installation forGoodmorning from developer 2 Goodmorning from developer 3Customer 0 ... from developerCustomer 13 3Bye from developer 2 * Elapsed time: 9.02 seconds
这次是一段并行执行的代码,使用了5个工作线程。15个消费者每个花费3s意味着总共45s的时间,不过用了5个线程并行执行总共只花费了9s的时间。这段代码有点复杂,很大一部分代码是用于管理并发,而不是专注于算法或者业务逻辑。另外,程序的输出结果看起来也很混杂,可读性也天津市。即使是简单的多线程的代码同样也难以写得很好,所以我们转为使用Twisted:
# For years we thought this was all there was... We kept hiring more # developers, more managers and buying servers. We were trying harder # optimising processes and fire-fighting while getting mediocre # performance in return. Till luckily one day our hosting # company decided to increase their fees and we decided to # switch to Twisted Ltd.! from twisted.internet import reactor from twisted.internet import defer from twisted.internet import task # Twisted has a slightly different approach def schedule_install(customer): # They are calling us back when a Wordpress installation completes. # They connected the caller recognition system with our CRM and # we know exactly what a call is about and what has to be done next. # # We now design processes of what has to happen on certain events. def schedule_install_wordpress(): def on_done(): print "Callback: Finished installation for", customer print "Scheduling: Installation for", customer return task.deferLater(reactor, 3, on_done) # def all_done(_): print "All done for", customer # # For each customer, we schedule these processes on the CRM # and that # is all our chief-Twisted developer has to do d = schedule_install_wordpress() d.addCallback(all_done) # return d # Yes, we don't need many developers anymore or any synchronization. # ~~ Super-powered Twisted developer ~~ def twisted_developer_day(customers): print "Goodmorning from Twisted developer" # # Here's what has to be done today work = [schedule_install(customer) for customer in customers] # Turn off the lights when done join = defer.DeferredList(work) join.addCallback(lambda _: reactor.stop()) # print "Bye from Twisted developer!" # Even his day is particularly short! twisted_developer_day(["Customer %d" % i for i in xrange(15)]) # Reactor, our secretary uses the CRM and follows-up on events! reactor.run()
运行结果:
------ Running example 3 ------ Goodmorning from Twisted developer Scheduling: Installation for Customer 0 .... Scheduling: Installation for Customer 14 Bye from Twisted developer! Callback: Finished installation for Customer 0 All done for Customer 0 Callback: Finished installation for Customer 1 All done for Customer 1 ... All done for Customer 14 * Elapsed time: 3.18 seconds
这次我们得到了完美的执行代码和可读性强的输出结果,并且没有使用线程。我们并行地处理了15个消费者,也就是说,本来需要45s的执行时间在3s之内就已经完成。这个窍门就是我们把所有的阻塞的对sleep()的调用都换成了Twisted中对等的task.deferLater()和回调函数。由于现在处理的操作在其他地方进行,我们就可以毫不费力地同时服务于15个消费者。
前面提到处理的操作发生在其他的某个地方。现在来解释一下,算术运算仍然发生在CPU内,但是现在的CPU处理速度相比磁盘和网络操作来说非常快。所以给CPU提供数据或者从CPU向内存或另一个CPU发送数据花费了大多数时间。我们使用了非阻塞的操作节省了这方面的时间,例如,task.deferLater()使用了回调函数,当数据已经传输完成的时候会被激活。
另一个很重要的一点是输出中的Goodmorning from Twisted developer和Bye from Twisted developer!信息。在代码开始执行时就已经打印出了这两条信息。如果代码如此早地执行到了这个地方,那么我们的应用真正开始运行是在什么时候呢?答案是,对于一个Twisted应用(包括Scrapy)来说是在reactor.run()里运行的。在调用这个方法之前,必须把应用中可能用到的每个Deferred链准备就绪,然后reactor.run()方法会监视并激活回调函数。
注意,reactor的主要一条规则就是,你可以执行任何操作,只要它足够快并且是非阻塞的。
现在好了,代码中没有那么用于管理多线程的部分了,不过这些回调函数看起来还是有些杂乱。可以修改成这样:
# Twisted gave us utilities that make our code way more readable! @defer.inlineCallbacks def inline_install(customer): print "Scheduling: Installation for", customer yield task.deferLater(reactor, 3, lambda: None) print "Callback: Finished installation for", customer print "All done for", customer def twisted_developer_day(customers): ... same as previously but using inline_install() instead of schedule_install() twisted_developer_day(["Customer %d" % i for i in xrange(15)]) reactor.run()
运行的结果和前一个例子相同。这段代码的作用和上一个例子是一样的,但是看起来更加简洁明了。inlineCallbacks生成器可以使用一些一些Python的机制来使得inline_install()函数暂停或者恢复执行。inline_install()函数变成了一个Deferred对象并且并行地为每个消费者运行。每次yield的时候,运行就会中止在当前的inline_install()实例上,直到yield的Deferred对象完成后再恢复运行。
现在唯一的问题是,如果我们不止有15个消费者,而是有,比如10000个消费者时又该怎样?这段代码会同时开始10000个同时执行的序列(比如HTTP请求、数据库的写操作等等)。这样做可能没什么问题,但也可能会产生各种失败。在有巨大并发请求的应用中,例如Scrapy,我们经常需要把并发的数量限制到一个可以接受的程度上。在下面的一个例子中,我们使用task.Cooperator()来完成这样的功能。Scrapy在它的Item Pipeline中也使用了相同的机制来限制并发的数目(即CONCURRENT_ITEMS设置):
@defer.inlineCallbacks def inline_install(customer): ... same as above # The new "problem" is that we have to manage all this concurrency to # avoid causing problems to others, but this is a nice problem to have. def twisted_developer_day(customers): print "Goodmorning from Twisted developer" work = (inline_install(customer) for customer in customers) # # We use the Cooperator mechanism to make the secretary not # service more than 5 customers simultaneously. coop = task.Cooperator() join = defer.DeferredList([coop.coiterate(work) for i in xrange(5)]) # join.addCallback(lambda _: reactor.stop()) print "Bye from Twisted developer!" twisted_developer_day(["Customer %d" % i for i in xrange(15)]) reactor.run() # We are now more lean than ever, our customers happy, our hosting # bills ridiculously low and our performance stellar. # ~*~ THE END ~*~
运行结果:
$ ./deferreds.py 5 ------ Running example 5 ------ Goodmorning from Twisted developer Bye from Twisted developer! Scheduling: Installation for Customer 0 ... Callback: Finished installation for Customer 4 All done for Customer 4 Scheduling: Installation for Customer 5 ... Callback: Finished installation for Customer 14 All done for Customer 14 * Elapsed time: 9.19 seconds
从上面的输出中可以看到,程序运行时好像有5个处理消费者的槽。除非一个槽空出来,否则不会开始处理下一个消费者的请求。在本例中,处理时间都是3秒,所以看起来像是5个一批次地处理一样。最后得到的性能跟使用线程是一样的,但是这次只有一个线程,代码也更加简洁更容易写出正确的代码。
PS:deferToThread使同步函数实现非阻塞
wisted的defer.Deferred (from twisted.internet import defer)可以返回一个deferred对象.
注:deferToThread使用线程实现的,不推荐过多使用
***把同步函数变为异步(返回一个Deferred)***
twisted的deferToThread(from twisted.internet.threads import deferToThread)也返回一个deferred对象,不过回调函数在另一个线程处理,主要用于数据库/文件读取操作
.. # 代码片段 def dataReceived(self, data): now = int(time.time()) for ftype, data in self.fpcodec.feed(data): if ftype == 'oob': self.msg('OOB:', repr(data)) elif ftype == 0x81: # 对服务器请求的心跳应答(这个是解析 防疲劳驾驶仪,发给gps上位机的,然后上位机发给服务器的) self.msg('FP.PONG:', repr(data)) else: self.msg('TODO:', (ftype, data)) d = deferToThread(self.redis.zadd, "beier:fpstat:fps", now, self.devid) d.addCallback(self._doResult, extra)
下面这儿完整的例子可以给大家参考一下
# -*- coding: utf-8 -*- from twisted.internet import defer, reactor from twisted.internet.threads import deferToThread import functools import time # 耗时操作 这是一个同步阻塞函数 def mySleep(timeout): time.sleep(timeout) # 返回值相当于加进了callback里 return 3 def say(result): print "耗时操作结束了, 并把它返回的结果给我了", result # 用functools.partial包装一下, 传递参数进去 cb = functools.partial(mySleep, 3) d = deferToThread(cb) d.addCallback(say) print "你还没有结束我就执行了, 哈哈" reactor.run()
更新动态
- 凤飞飞《我们的主题曲》飞跃制作[正版原抓WAV+CUE]
- 刘嘉亮《亮情歌2》[WAV+CUE][1G]
- 红馆40·谭咏麟《歌者恋歌浓情30年演唱会》3CD[低速原抓WAV+CUE][1.8G]
- 刘纬武《睡眠宝宝竖琴童谣 吉卜力工作室 白噪音安抚》[320K/MP3][193.25MB]
- 【轻音乐】曼托凡尼乐团《精选辑》2CD.1998[FLAC+CUE整轨]
- 邝美云《心中有爱》1989年香港DMIJP版1MTO东芝首版[WAV+CUE]
- 群星《情叹-发烧女声DSD》天籁女声发烧碟[WAV+CUE]
- 刘纬武《睡眠宝宝竖琴童谣 吉卜力工作室 白噪音安抚》[FLAC/分轨][748.03MB]
- 理想混蛋《Origin Sessions》[320K/MP3][37.47MB]
- 公馆青少年《我其实一点都不酷》[320K/MP3][78.78MB]
- 群星《情叹-发烧男声DSD》最值得珍藏的完美男声[WAV+CUE]
- 群星《国韵飘香·贵妃醉酒HQCD黑胶王》2CD[WAV]
- 卫兰《DAUGHTER》【低速原抓WAV+CUE】
- 公馆青少年《我其实一点都不酷》[FLAC/分轨][398.22MB]
- ZWEI《迟暮的花 (Explicit)》[320K/MP3][57.16MB]