Pykka is a Python implementation of the actor model. The actor model introduces some simple rules to control the sharing of state and cooperation between execution units, which makes it easier to build concurrent applications.
An actor is an execution unit that executes concurrently with other actors.
An actor does not share state with anybody else, but it can have its own state.
An actor can only communicate with other actors by sending and receiving messages. It can only send messages to actors whose address it has.
When an actor receives a message it may take actions like:
- altering its own state, e.g. so that it can react differently to a future message,
- sending messages to other actors, or
- starting new actors.
None of the actions are required, and they may be applied in any order.
An actor only processes one message at a time. In other words, a single actor does not give you any concurrency, and it does not need to use locks internally to protect its own state.
Pykka's actor API comes with the following implementations:
Threads: Each
ThreadingActor
is executed by a regular thread, i.e.threading.Thread
. As handles for future results, it usesThreadingFuture
which is a thin wrapper around aQueue.Queue
. It has no dependencies outside Python itself.ThreadingActor
plays well together with non-actor threads.Note: If you monkey patch the standard library with
gevent
oreventlet
you can still useThreadingActor
andThreadingFuture
. Python's threads will transparently use the underlying implementation provided by gevent or Eventlet.gevent: Each
GeventActor
is executed by a gevent greenlet. gevent is a coroutine-based Python networking library built on top of a libevent (in 0.13) or libev (in 1.0) event loop.GeventActor
is generally faster thanThreadingActor
, but as of gevent 0.13 it doesn't work in processes with other threads, which limits when it can be used. With gevent 1.0, which is currently available as a release candidate, this is no longer an issue. Pykka works with both gevent 0.13 and 1.0.Eventlet: Each
EventletActor
is executed by a Eventlet greenlet. Pykka is tested with Eventlet 0.12.1.
Pykka has an extensive test suite, and is tested on CPython 2.6, 2.7, and 3.2+, as well as PyPy. gevent and eventlet are currently not available for CPython 3.x or PyPy.
In its most basic form, a Pykka actor is a class with an
on_receive(message)
method:
import pykka class Greeter(pykka.ThreadingActor): def on_receive(self, message): print('Hi there!')
To start an actor, you call the class' method start()
, which starts the
actor and returns an actor reference which can be used to communicate with the
running actor:
actor_ref = Greeter.start()
If you need to pass arguments to the actor upon creation, you can pass them to
the start()
method, and receive them using the regular __init__()
method:
import pykka class Greeter(pykka.ThreadingActor): def __init__(self, greeting='Hi there!'): super(Greeter, self).__init__() self.greeting = greeting def on_receive(self, message): print(self.greeting) actor_ref = Greeter.start(greeting='Hi you!')
It can be useful to know that the init method is run in the execution context that starts the actor. There are also hooks for running code in the actor's own execution context when the actor starts, when it stops, and when an unhandled exception is raised. Check out the full API docs for the details.
To stop an actor, you can either call stop()
on the actor_ref
:
actor_ref.stop()
Or, if an actor wants to stop itself, it can simply do so:
self.stop()
Once an actor has been stopped, it cannot be restarted.
To send a message to the actor, you can either use the tell()
method or the
ask()
method on the actor_ref
object. tell()
will fire of a message
without waiting for an answer. In other words, it will never block. ask()
will by default block until an answer is returned, potentially forever. If you
provide a timeout
keyword argument to ask()
, you can specify for how
long it should wait for an answer. If you want an answer, but don't need it
right away because you have other stuff you can do first, you can pass
block=False
, and ask()
will immediately return a "future" object.
The message itself must always be a dict, but you're mostly free to use whatever dict keys you want to.
Summarized in code:
actor_ref.tell({'msg': 'Hi!'}) # => Returns nothing. Will never block. answer = actor_ref.ask({'msg': 'Hi?'}) # => May block forever waiting for an answer answer = actor_ref.ask({'msg': 'Hi?'}, timeout=3) # => May wait 3s for an answer, then raises exception if no answer. future = actor_ref.ask({'msg': 'Hi?'}, block=False) # => Will return a future object immediately. answer = future.get() # => May block forever waiting for an answer answer = future.get(timeout=0.1) # => May wait 0.1s for an answer, then raises exception if no answer.
For performance reasons, Pykka does not clone the dict you send before
delivering it to the receiver. You are yourself responsible for either using
immutable data structures or to copy.deepcopy()
the data you're sending off
to other actors.
If a message is sent using actor_ref.ask()
you can reply to the sender of
the message by simply returning a value from on_receive
method:
import pykka class Greeter(pykka.ThreadingActor): def on_receive(self, message): return 'Hi there!' actor_ref = Greeter.start() answer = actor_ref.ask({'msg': 'Hi?'}) print(answer) # => 'Hi there!'
None
is a valid response so if you return None
explicitly, or don't
return at all, a response containing None
will be returned to the sender.
From the point of view of the actor it doesn't matter whether the message was
sent using actor_ref.tell()
or actor_ref.ask()
. When the sender
doesn't expect a response the on_receive
return value will be ignored.
The situation is similar in regard to exceptions: when actor_ref.ask()
is
used and you raise an exception from within on_receive
method it will
propagate to the sender:
import pykka class Raiser(pykka.ThreadingActor): def on_receive(self, message): raise Exception('Oops') actor_ref = Raiser.start() try: actor_ref.ask({'msg': 'How are you?'}) except Exception as e: print(repr(e)) # => Exception('Oops')
With the basic building blocks provided by actors and futures, we got everything we need to build more advanced abstractions. Pykka provides a single abstraction on top of the basic actor model, named "actor proxies". You can use Pykka without proxies, but we've found it to be a very convenient abstraction when builing Mopidy.
Let's create an actor and start it:
import pykka class Calculator(pykka.ThreadingActor): def __init__(self): super(Calculator, self).__init__() self.last_result = None def add(self, a, b=None): if b is not None: self.last_result = a + b else: self.last_result += a return self.last_result def sub(self, a, b=None): if b is not None: self.last_result = a - b else: self.last_result -= a return self.last_result actor_ref = Calculator.start()
You can create a proxy from any reference to a running actor:
proxy = actor_ref.proxy()
The proxy object will use introspection to figure out what public attributes and methods the actor has, and then mirror the full API of the actor. Any attribute or method prefixed with underscore will be ignored, which is the convention for keeping stuff private in Python.
When we access attributes or call methods on the proxy, it will ask the actor
to access the given attribute or call the given method, and return the result
to us. All results are wrapped in "future" objects, so you must use the
get()
method to get the actual data:
future = proxy.add(1, 3) future.get() # => 4 proxy.last_result.get() # => 4
Since an actor only processes one message at the time and all messages are kept
in order, you don't need to add the call to get()
just to block
processing until the actor has completed processing your last message:
proxy.sub(5) proxy.add(3) proxy.last_result.get() # => 2
Since assignment doesn't return anything, it works just like on regular objects:
proxy.last_result = 17 proxy.last_result.get() # => 17
Under the hood, the proxy does everything by sending messages to the actor
using the regular actor_ref.ask()
method we talked about previously.
By doing so, it maintains the actor model restrictions. The only "magic"
happening here is some basic introspection and automatic building of three
different message types; one for method calls, one for attribute reads, and one
for attribute writes.
Sometimes you'll want to access an actor attribute's methods or attributes through a proxy. For this case, Pykka supports "traversable attributes". By marking an actor attribute as traversable, Pykka will not return the attribute when accessed, but wrap it in a new proxy which is returned instead.
To mark an attribute as traversable, simply set the pykka_traversable
attribute to True
:
import pykka class AnActor(pykka.ThreadingActor): playback = Playback() class Playback(object): pykka_traversable = True def play(self): # ... return True proxy = AnActor.start().proxy() play_success = proxy.playback.play().get()
You can access methods and attributes nested as deep as you like, as long as all attributes on the path between the actor and the method or attribute on the end is marked as traversable.
See the examples/
dir in Pykka's Git repo for some runnable examples.
Much of the naming of concepts and methods in Pykka is taken from the Akka project which implements actors on the JVM. Though, Pykka does not aim to be a Python port of Akka, and supports far fewer features.
Notably, Pykka does not support the following features:
- Supervision: Linking actors, supervisors, or supervisor groups.
- Remoting: Communicating with actors running on other hosts.
- Routers: Pykka does not come with a set of predefined message routers, though you may make your own actors for routing messages.
Install Pykka's dependencies:
- Python 2.6, 2.7, or 3.2+
- Optionally, Python 2.6/2.7 only:
To install Pykka you can use pip:
pip install pykka
To upgrade your Pykka installation to the latest released version:
pip install --upgrade pykka
To install the latest development snapshot:
pip install pykka==dev
Pykka is copyright 2010-2014 Stein Magnus Jodal and contributors. Pykka is licensed under the Apache License, Version 2.0.