Preface
========
Introduction
-------------
This library provides a simple event dispatcher, similar to the
`event` construct provided by the C# language. The library has no
external dependencies.
It was intended for use cases, where the components emitting events
and the components listening for events agree about the type of
event and the semantics associated with it. This is true, for example,
for event handlers, which listen on "click" events signalled by
GUI button objects, or notifications signalled by objects, whenever
the value of some property changes. This differs from the approach
taken by, say, the `PyDispatcher`, which is more generic, and
favours communication among weakly coupled components.
Compatibility
--------------
The code was mainly written for and tested with Python 2.6. It is known
to work with Python 3.2. It should be compatible to 2.5 as well,
but you might have to insert a few `from __future__ import with_statement`
lines here and there (and this is generally untested). It should work
(this has not been tested) with alternative implementations of Python
like Jython or IronPython. Note, though, that some of the test cases
defined in this file might fail due to different garbage collection
implementations; this file was written with CPython in mind.
Documentation
==============
Basic Usage
------------
>>> from darts.lib.utils.event import Publisher, ReferenceRetention as RR
>>> some_event = Publisher()
The `Publisher` is the main component. It acts as registry for
callbacks/listeners. Let's define a listener
>>> def printer(*event_args, **event_keys):
... print event_args, event_keys
In order to receive notifications, clients must subscribe to a
publisher. This can be as simple as
>>> some_event.subscribe(printer) #doctest: +ELLIPSIS
<SFHandle ...>
The result of the call to `subscribe` is an instance of (some subclass
of) class `Subscription`. This value may be used later, in order to
cancel the subscription, when notifications are no longer desired. The
actual subclass is an implementation detail you should normally not
care about. All you need to know (and are allowed to rely on, in fact)
is, that it will be an instance of class `Subscription`, and it will
provide whatever has been documented as public API of that class (right
now: only method `cancel`).
Now, let's signal an event and see what happens:
>>> some_event.publish('an-event')
('an-event',) {}
As you can see, the `printer` has been notified of the event, and
duefully printed the its arguments to the console.
Cancelling subscriptions
-------------------------
As mentioned, the result of calling `subscribe` is a special subscription
object, which represents the registration of the listener with the
publisher.
>>> s1 = some_event.subscribe(printer)
>>> some_event.publish('another-event')
('another-event',) {}
('another-event',) {}
>>> s1.cancel()
True
>>> some_event.publish('yet-another-one')
('yet-another-one',) {}
The publisher is fully re-entrant. That means, that you can subscribe
to events from within a listener, and you can cancel subscriptions in
that context as well:
>>> def make_canceller(subs):
... def listener(*unused_1, **unused_2):
... print "Cancel", subs, subs.cancel()
... return listener
>>> s1 = some_event.subscribe(printer)
>>> s2 = some_event.subscribe(make_canceller(s1))
>>> some_event.publish('gotta-go') #doctest: +ELLIPSIS
('gotta-go',) {}
('gotta-go',) {}
Cancel <SFHandle ...> True
>>> some_event.publish('gone') #doctest: +ELLIPSIS
('gone',) {}
Cancel <SFHandle ...> False
>>> s1.cancel()
False
The result of the call to `cancel` tells us, that the subscription had
already been undone prior to the call (by our magic cancellation listener).
Generally, calling `cancel` multiple times is harmless; all but the first
call are ignored.
Let's now remove the magic I-can-cancel-stuff listener and move on:
>>> s2.cancel()
True
Using Non-Callables as callbacks
---------------------------------
Whenever we made subscriptions above, we actually simplied things a
little bit. The full signature of the method is:
def subscribe(listener[, method[, reference_retention]])
Let's explore the `method` argument first. Up to now, we only used
function objects as listeners. Basically, in fact, we might have used
any callable object. Remember, that any object is "callable" in Python,
if it provides a `__call__` method, so guess, what's the default value
of the `method` argument?
>>> s1 = some_event.subscribe(printer, method='__call__')
>>> some_event.publish('foo')
('foo',) {}
('foo',) {}
>>> s1.cancel()
True
Nothing new. So, now you might ask: when do I use a different method
name?
>>> class Target(object):
... def __init__(self, name):
... self.name = name
... def _callback(self, *args, **keys):
... print self.name, args, keys
>>> s1 = some_event.subscribe(Target('foo'))
>>> some_event.publish('Bumm!') #doctest: +ELLIPSIS
Traceback (most recent call last):
...
TypeError: 'Target' object is not callable
Oops. Let's remove the offender, before someone notices our mistake:
>>> s1.cancel()
True
>>> s1 = some_event.subscribe(Target('foo'), method='_callback')
>>> some_event.publish('works!')
('works!',) {}
foo ('works!',) {}
Reference Retention
--------------------
So, that's that. There is still an unexplored argument to `subscribe`
left, though: `reference_retention`. The name sounds dangerous, but what
does it do?
>>> listener = Target('yummy')
>>> s2 = some_event.subscribe(listener, method='_callback', reference_retention=RR.WEAK)
>>> some_event.publish('yow')
('yow',) {}
foo ('yow',) {}
yummy ('yow',) {}
Hm. So far, no differences. Let's make a simple change:
>>> listener = None
>>> some_event.publish('yow')
('yow',) {}
foo ('yow',) {}
Ah. Ok. Our `yummy` listener is gone. What happened? Well, by specifying
a reference retention policy of `WEAK`, we told the publisher, that it should
use a weak reference to the listener just installed, instead of the default
strong reference. And after we released the only other known strong
reference to the listener by setting `listener` to `None`, the listener
was actually removed from the publisher. Note, BTW., that the above example
may fail with python implementations other than CPython, due to different
policies with respect to garbage collection. The principle should remain
valid, though, in Jython as well as IronPython, but in those implementations,
there is no guarantee, that the listener is removed as soon as the last
reference to it is dropped.
Of course, this all works too, if the method to be called is the default
one: `__call__`:
>>> def make_listener(name):
... def listener(*args, **keys):
... print name, args, keys
... return listener
>>> listener = make_listener('weak')
>>> s2 = some_event.subscribe(listener, reference_retention=RR.WEAK)
>>> some_event.publish('event')
('event',) {}
foo ('event',) {}
weak ('event',) {}
>>> listener = None
>>> some_event.publish('event')
('event',) {}
foo ('event',) {}
That's about all there is to know about the library. As I said above: it
is simple, and might not be useful for all scenarioes and use cases, but
it does what it was written to.
Error handling
----------------
The `Publisher` class is not intended to be subclassed. If you need to
tailor the behaviour, you use policy objects/callbacks, which are passed
to the constructor. Right now, there is a single adjustable policy, namely,
the behaviour of the publisher in case, listeners raise exceptions:
>>> def toobad(event):
... if event == 'raise':
... raise ValueError
>>> s1 = some_event.subscribe(toobad)
>>> some_event.publish('harmless')
('harmless',) {}
foo ('harmless',) {}
>>> some_event.publish('raise')
Traceback (most recent call last):
...
ValueError
As you can see, the default behaviour is to re-raise the exception
from within `publish`. This might not be adequate depending on the use
case. In particular, it will prevent any listeners registered later
to be run. So, let's define our own error handling:
>>> def log_error(exception, value, traceback, subscription, args, keys):
... print "caught", exception
>>> publisher = Publisher(exception_handler=log_error)
>>> publisher.subscribe(toobad) #doctest: +ELLIPSIS
<SFHandle ...>
>>> publisher.subscribe(printer) #doctest: +ELLIPSIS
<SFHandle ...>
>>> publisher.publish('harmless')
('harmless',) {}
>>> publisher.publish('raise')
caught <type 'exceptions.ValueError'>
('raise',) {}
As an alternative to providing the error handler at construction time,
you may also provide an error handler when publishing an event, like
so:
>>> def log_error_2(exception, value, traceback, subscription, args, keys):
... print "caught", exception, "during publication"
>>> publisher.publish_safely(log_error_2, 'raise')
caught <type 'exceptions.ValueError'> during publication
('raise',) {}
As you can see, the per-call error handler takes precedence over the
publisher's default error handler. Note, that there is no chaining, i.e.,
if the per-call error handler raises an exception, the publisher's default
handler is *not* called, but the exception is simply propagated outwards
to the caller of `publish_safely`: the publisher has no way to distinguish
between exceptions raised because the handler wants to abort the dispatch
and exceptions raised by accident, so all exceptions raised by the handler
are simply forwarded to the client application.
Thread Safety
==============
The library is fully thread aware and thread safe. Thus, subscribing to
a listener shared across multiple threads is safe, and so is cancelling
subscriptions.
Changes
========
Version 0.4
------------
Subscription handles now provide access to their listener objects and
method names. This was added for the sake of error handling code, which
wants to log exceptions and provide a better way of identifying the
actual listener, which went rogue.
Version 0.3
------------
Fixed `setup.py` to properly proclaim the namespace packages used.
Version 0.2
------------
Error handling has been changed. Instead of subclassing the publisher,
the default exception handler is now passed as callback to the publisher
during construction. The class `Publisher` is now documented as "not
intended for being subclassed".