C[omp]ute

From: andrew cooke <andrew@...>

Date: Sun, 22 Aug 2010 17:01:28 -0400

I've been lookng at Twisted, which is a framework for cooperative
multi-tasking in Python.  I don't find that a very useful description, so here
are two alternatives:

1 - It's a way of structuring multi-threaded programs that's a lot more like
    Javascript or GUI toolkits.

2 - It's a way of writing network servers that work efficiently without using
    multiple threads.


There's a fair amount of documentation at
http://twistedmatrix.com/documents/current/core/ (and most imporantly at
http://twistedmatrix.com/documents/current/core/howto/index.html ) - I suggest
reading through that until it sticks.  It took me a while, and writing the
code below, but now it makes a lot of sense (and it seems like a very nicely
engineered system).


I structured the example below as a set of different files, which was probably
excessive, but I wanted the difference components to be as clear as possible.

Python logging can be serialised over a socket.  This code in a server that
receives seralised messages and writes them to a log.


First, the protocol:

  from cPickle import loads
  from logging import makeLogRecord, getLogger
  from struct import unpack
  from twisted.internet.protocol import Protocol, connectionDone

  '''
  The protocol for a Twisted server that receives log messages.

  See http://docs.python.org/library/logging.html#socket-handler 
  '''

  class LoggingProtocol(Protocol):

      def dataReceived(self, data):
	  self.__data += data
	  while True:
	      if not self.__message_len and len(self.__data) >= 4:
		  # unpack length prefix
		  self.__message_len = unpack(">L", self.__data[:4])[0]
		  self.__data = self.__data[4:]
	      if self.__message_len and len(self.__data) >= self.__message_len:
		  # unpack message
		  record =
	  makeLogRecord(loads(self.__data[0:self.__message_len]))
		  self.__data = self.__data[self.__message_len:]
		  self.__message_len = 0
		  logger = getLogger(record.name)
		  logger.handle(record)
	      else:
		  break

      def connectionMade(self):
	  self.__data = ''
	  self.__message_len = 0

      def connectionLost(self, reason=connectionDone):
	  self.__data = None
	  self.__message_len = None


Next, the factory (ie a protocol factory):

  from logging.config import dictConfig
  from twisted.internet.protocol import Factory

  from log.protocol import LoggingProtocol

  '''
  A factory for the remote Python logger.

  This seems to be the best location to store configuration information because
  it is accessible both in tests (using a reactor) and to an application.
  '''

  class LoggingFactory(Factory):

      protocol = LoggingProtocol

      DEFAULT_PORT = 2000
      DEFAULT_CONFIG = {'version': 1,
			'handlers':
			  {'file':
			    {'class': 'logging.FileHandler',
			     'filename': 'logging-service.log',
			     'level': 'DEBUG',
			  },},
			'root':
			  {'level': 'DEBUG',
			   'handlers': ['file']},}

      def __init__(self, config_dict=None):
	  if not config_dict:
	      config_dict = self.DEFAULT_CONFIG
	  dictConfig(config_dict)


And the service:

  from twisted.application.internet import TCPServer

  from log.factory import LoggingFactory

  '''
  A service for the remote Python logger.

  This is used by the application.
  '''

  class LoggingService(TCPServer):

      def __init__(self, port=None, config_dict=None, interface='0.0.0.0'):
	  if not port:
	      port = LoggingFactory.DEFAULT_PORT
	  # old style clases in twisted
	  TCPServer.__init__(self, port, LoggingFactory(config_dict), 
			     interface=interface)


This can then be made into an application (a daemon) that's run from the
command-line using a tool called "twistd":

  # You can run this .tac file directly with:
  #    twistd -ny service.tac

  from log.service import LoggingService
  from twisted.application import service

  application = service.Application("Logging application")
  LoggingService().setServiceParent(application)


Alternatively, for testing, the Fatcory can be used directly.  This test code
also gives a glimpse of how the reactor is used to schedule events (there's
also an abstraction for chaining callbacks called "Defered"):

  from logging.config import dictConfig
  from logging import getLogger
  from multiprocessing.process import Process
  from tempfile import mkstemp
  from twisted.internet import reactor
  from unittest import TestCase

  from log.factory import LoggingFactory


  class LoggingTest(TestCase):
      '''
      Test the logging service by starting an instance, then firing up a 
      separate process that logs to the service.
      '''

      def test_logging(self):
	  tick = Tick()
	  (_fd, self.tmp) = mkstemp()
	  process = Process(target=self.logging_process)
	  factory = LoggingFactory({'version': 1,
				    'handlers': {'file': {'class': 'logging.FileHandler',
							  'filename': self.tmp,
							  'level': 'DEBUG'}},
				    'root': {'level': 'DEBUG',
					     'handlers': ['file']}})
	  reactor.listenTCP(factory.DEFAULT_PORT, factory)
	  reactor.callLater(tick(), process.start)
	  reactor.callLater(tick(), reactor.stop)
	  reactor.run()
	  fd = open(self.tmp)
	  contents = fd.readlines()
	  assert contents == ['a warning\n'], contents
	  fd.close()

      def logging_process(self):
	  dictConfig({'version': 1,
		      'handlers':
			{'socket':
			  {'class': 'logging.handlers.SocketHandler',
			   'level': 'INFO',
			   'host': 'localhost',
			   'port': LoggingFactory.DEFAULT_PORT
			   },},
		      'root':
			{'level': 'INFO',
			 'handlers': ['socket']},
		     })
	  logger = getLogger('test')
	  logger.debug('a debug') # discarded by "level: INFO" above
	  logger.warn('a warning')


  class Tick(object):

      def __init__(self, increment=0.1):
	  self.__increment = increment
	  self.__time = 0

      def __call__(self, step=1):
	  self.__time += step * self.__increment
	  return self.__time


Andrew