Changes
v1.3.0 (UNRELEASED)
Drop support for Python 2.6, 3.2, and 3.3.
Include Python 3.5 and 3.6 in the test matrix.
v1.2.1 (2015-07-20)
Increase log level of
pykka.debug.log_thread_tracebacks()
debugging helper fromlogging.INFO
tologging.CRITICAL
.Fix typos in docs.
Various project setup and development improvements.
v1.2.0 (2013-07-15)
Enforce that multiple calls to
pykka.Future.set()
raises an exception. This was already the case for some implementations. The exception raised is not specified.Add
filter()
,join()
,map()
, andreduce()
as convenience methods using the newset_get_hook()
method.Add support for running actors based on eventlet greenlets. See
pykka.eventlet
for details. Thanks to Jakub Stasiak for the implementation.Update documentation to reflect that the
reply_to
field on the message is private to Pykka. Actors should reply to messages simply by returning the response fromon_receive()
. The internal field is renamed topykka_reply_to
a to avoid collisions with other message fields. It is also removed from the message before the message is passed toon_receive()
. Thanks to Jakub Stasiak.When messages are left in the actor inbox after the actor is stopped, those messages that are expecting a reply is now rejected by replying with an
ActorDeadError
exception. This causes other actors blocking on the returnedFuture
without a timeout to raise the exception instead of waiting forever. Thanks to Jakub Stasiak.This makes the behavior of messaging an actor around the time it is stopped more consistent:
Messaging an already dead actor immediately raises
ActorDeadError
.Messaging an alive actor that is stopped before it processes the message will cause the reply future to raise
ActorDeadError
.
Similarly, if you ask an actor to stop multiple times, and block on the responses, all the messages will now get an reply. Previously only the first message got a reply, potentially making the application wait forever on replies to the subsequent stop messages.
When
ask()
is used to asynchronously message a dead actor (e.g.block
set toFalse
), it will no longer immediately raiseActorDeadError
. Instead, it will return a future and fail the future with theActorDeadError
exception. This makes the interface more consistent, as you’ll have one instead of two ways the call can raise exceptions under normal conditions. Ifask()
is called synchronously (e.g.block
set toTrue
), the behavior is unchanged.A change to
stop()
reduces the likelyhood of a race condition when asking an actor to stop multiple times by not checking if the actor is dead before asking it to stop, but instead just go ahead and leave it totell()
to do the alive-or-dead check a single time, and as late as possible.Change
is_alive()
to check the actor’s runnable flag instead of checking if the actor is registrered in the actor registry.
v1.1.0 (2013-01-19)
An exception raised in
pykka.Actor.on_start()
didn’t stop the actor properly. Thanks to Jay Camp for finding and fixing the bug.Make sure exceptions in
pykka.Actor.on_stop()
andpykka.Actor.on_failure()
is logged.Add
pykka.ThreadingActor.use_daemon_thread
flag for optionally running an actor on a daemon thread, so that it doesn’t block the Python program from exiting. (Fixes: #14)Add
pykka.debug.log_thread_tracebacks()
debugging helper. (Fixes: #17)
v1.0.1 (2012-12-12)
Name the threads of
pykka.ThreadingActor
after the actor class name instead of “PykkaThreadingActor-N” to ease debugging. (Fixes: #12)
v1.0.0 (2012-10-26)
Backwards incompatible: Removed
pykka.VERSION
andpykka.get_version()
, which have been deprecated since v0.14. Usepykka.__version__
instead.Backwards incompatible: Removed
pykka.ActorRef.send_one_way()
andpykka.ActorRef.send_request_reply()
, which have been deprecated since v0.14. Usepykka.ActorRef.tell()
andpykka.ActorRef.ask()
instead.Backwards incompatible: Actors no longer subclass
threading.Thread
orgevent.Greenlet
. Instead they have a thread or greenlet that executes the actor’s main loop.This is backwards incompatible because you no longer have access to fields/methods of the thread/greenlet that runs the actor through fields/methods on the actor itself. This was never advertised in Pykka’s docs or examples, but the fields/methods have always been available.
As a positive side effect, this fixes an issue on Python 3.x, that was introduced in Pykka 0.16, where
pykka.ThreadingActor
would accidentally override the methodthreading.Thread._stop()
.Backwards incompatible: Actors that override
__init__()
must call the method they override. If not, the actor will no longer be properly initialized. Valid ways to call the overridden__init__()
method include:super(MyActorSubclass, self).__init__() # or pykka.ThreadingActor.__init__() # or pykka.gevent.GeventActor.__init__()
Make
pykka.Actor.__init__()
accept any arguments and keyword arguments by default. This allows you to usesuper()
in__init__()
like this:super(MyActorSubclass, self).__init__(1, 2, 3, foo='bar')
Without this fix, the above use of
super()
would cause an exception because the default implementation of__init__()
inpykka.Actor
would not accept the arguments.Allow all public classes and functions to be imported directly from the
pykka
module. E.g.from pykka.actor import ThreadingActor
can now be written asfrom pykka import ThreadingActor
. The exception ispykka.gevent
, which still needs to be imported from its own package due to its additional dependency on gevent.
v0.16 (2012-09-19)
Let actors access themselves through a proxy. See the
pykka.ActorProxy
documentation for use cases and usage examples. (Fixes: #9)Give proxies direct access to the actor instances for inspecting available attributes. This access is only used for reading, and works since both threading and gevent based actors share memory with other actors. This reduces the creation cost for proxies, which is mostly visible in test suites that are starting and stopping lots of actors. For the Mopidy test suite the run time was reduced by about 33%. This change also makes self-proxying possible.
Fix bug where
pykka.Actor.stop()
called by an actor on itself did not process the remaining messages in the inbox before the actor stopped. The behavior now matches the documentation.
v0.15 (2012-08-11)
Change the argument of
pykka.Future.set_exception()
from an exception instance to aexc_info
three-tuple. Passing just an exception instance to the method still works, but it is deprecated and may be unsupported in a future release.Due to the above change,
pykka.Future.get()
will now reraise exceptions with complete traceback from the point when the exception was first raised, and not just a traceback from when it was reraised byget()
. (Fixes: #10)
v0.14 (2012-04-22)
Add
pykka.__version__
to conform with PEP 396. This deprecatespykka.VERSION
andpykka.get_version()
.Add
pykka.ActorRef.tell()
method in favor of now deprecatedpykka.ActorRef.send_one_way()
.Add
pykka.ActorRef.ask()
method in favor of now deprecatedpykka.ActorRef.send_request_reply()
.ThreadingFuture.set()
no longer makes a copy of the object set on the future. The setter is urged to either only pass immutable objects through futures or copy the object himself before setting it on the future. This is a less safe default, but it removes unecessary overhead in speed and memory usage for users of immutable data structures. For example, the Mopidy test suite of about 1000 tests, many which are using Pykka, is still passing after this change, but the test suite runs approximately 20% faster.
v0.13 (2011-09-24)
10x speedup of traversible attribute access by reusing proxies.
1.1x speedup of callable attribute access by reusing proxies.
v0.12.4 (2011-07-30)
Change and document order in which
pykka.ActorRegistry.stop_all()
stops actors. The new order is the reverse of the order the actors were started in. This should makestop_all
work for programs with simple dependency graphs in between the actors. For applications with more complex dependency graphs, the developer still needs to pay attention to the shutdown sequence. (Fixes: #8)
v0.12.3 (2011-06-25)
If an actor that was stopped from
pykka.Actor.on_start()
, it would unregister properly, but start the receive loop and forever block on receiving incoming messages that would never arrive. This left the thread alive and isolated, ultimately blocking clean shutdown of the program. The fix ensures that the receive loop is never executed if the actor is stopped before the receive loop is started.Set the thread name of any
pykka.ThreadingActor
toPykkaActorThread-N
instead of the defaultThread-N
. This eases debugging by clearly labeling actor threads in e.g. the output ofthreading.enumerate()
.Add utility method
pykka.ActorRegistry.broadcast()
which broadcasts a message to all registered actors or to a given class of registred actors. (Fixes: #7)Allow multiple calls to
pykka.ActorRegistry.unregister()
with the samepykka.actor.ActorRef
as argument without throwing aValueError
. (Fixes: #5)Make the
pykka.ActorProxy
’s reference to itspykka.ActorRef
public aspykka.ActorProxy.actor_ref
. TheActorRef
instance was already exposed as a public field by the actor itself using the same name, but making it public directly on the proxy makes it possible to do e.g.proxy.actor_ref.is_alive()
without waiting for a potentially dead actor to return anActorRef
instance you can use. (Fixes: #3)
v0.12.2 (2011-05-05)
Actors are now registered in
pykka.registry.ActorRegistry
before they are started. This fixes a race condition where an actor tried to stop and unregister itself before it was registered, causing an exception inActorRegistry.unregister()
.
v0.12.1 (2011-04-25)
Stop all running actors on
BaseException
instead of justKeyboardInterrupt
, so thatsys.exit(1)
will work.
v0.12 (2011-03-30)
First stable release, as Pykka now is used by the Mopidy project. From now on, a changelog will be maintained and we will strive for backwards compatibility.