neutron/doc/source/devref/callbacks.rst
Tracy Jones b853e77421 Fix spelling mistakes in the docs
Change-Id: Ic01cbc0634214d3ecef2033333cd263343a78941
2016-06-30 09:38:17 -07:00

19 KiB

Neutron Callback System

In Neutron, core and service components may need to cooperate during the execution of certain operations, or they may need to react upon the occurrence of certain events. For instance, when a Neutron resource is associated to multiple services, the components in charge of these services may need to play an active role in determining what the right state of the resource needs to be.

The cooperation may be achieved by making each object aware of each other, but this leads to tight coupling, or alternatively it can be achieved by using a callback-based system, where the same objects are allowed to cooperate in a loose manner.

This is particularly important since the spin off of the advanced services like VPN, Firewall and Load Balancer, where each service's codebase lives independently from the core and from one another. This means that the tight coupling is no longer a practical solution for object cooperation. In addition to this, if more services are developed independently, there is no viable integration between them and the Neutron core. A callback system, and its registry, tries to address these issues.

In object-oriented software systems, method invocation is also known as message passing: an object passes a message to another object, and it may or may not expect a message back. This point-to-point interaction can take place between the parties directly involved in the communication, or it can happen via an intermediary. The intermediary is then in charge of keeping track of who is interested in the messages and in delivering the messages forth and back, when required. As mentioned earlier, the use of an intermediary has the benefit of decoupling the parties involved in the communications, as now they only need to know about the intermediary; the other benefit is that the use of an intermediary opens up the possibility of multiple party communication: more than one object can express interest in receiving the same message, and the same message can be delivered to more than one object. To this aim, the intermediary is the entity that exists throughout the system lifecycle, as it needs to be able to track whose interest is associated to what message.

In a design for a system that enables callback-based communication, the following aspects need to be taken into account:

  • how to become consumer of messages (i.e. how to be on the receiving end of the message);
  • how to become producer of messages (i.e. how to be on the sending end of the message);
  • how to consume/produce messages selectively;

Translate and narrow this down to Neutron needs, and this means the design of a callback system where messages are about lifecycle events (e.g. before creation, before deletion, etc.) of Neutron resources (e.g. networks, routers, ports, etc.), where the various parties can express interest in knowing when these events for a specific resources take place.

Rather than keeping the conversation abstract, let us delve into some examples, that would help understand better some of the principles behind the provided mechanism.

Subscribing to events

Imagine that you have entity A, B, and C that have some common business over router creation. A wants to tell B and C that the router has been created and that they need to get on and do whatever they are supposed to do. In a callback-less world this would work like so:

# A is done creating the resource
# A gets hold of the references of B and C
# A calls B
# A calls C
B->my_random_method_for_knowing_about_router_created()
C->my_random_very_difficult_to_remember_method_about_router_created()

If B and/or C change, things become sour. In a callback-based world, things become a lot more uniform and straightforward:

# B and C ask I to be notified when A is done creating the resource

# ...
# A is done creating the resource
# A gets hold of the reference to the intermediary I
# A calls I
I->notify()

Since B and C will have expressed interest in knowing about A's business, 'I' will deliver the messages to B and C. If B and C changes, A and 'I' do not need to change.

In practical terms this scenario would be translated in the code below:

from neutron.callbacks import events
from neutron.callbacks import resources
from neutron.callbacks import registry


def callback1(resource, event, trigger, **kwargs):
    print('Callback1 called by trigger: ', trigger)
    print('kwargs: ', kwargs)

def callback2(resource, event, trigger, **kwargs):
    print('Callback2 called by trigger: ', trigger)
    print('kwargs: ', kwargs)


# B and C express interest with I
registry.subscribe(callback1, resources.ROUTER, events.BEFORE_CREATE)
registry.subscribe(callback2, resources.ROUTER, events.BEFORE_CREATE)
print('Subscribed')


# A notifies
def do_notify():
    kwargs = {'foo': 'bar'}
    registry.notify(resources.ROUTER, events.BEFORE_CREATE, do_notify, **kwargs)


print('Notifying...')
do_notify()

The output is:

> Subscribed
> Notifying...
> Callback2 called by trigger:  <function do_notify at 0x7f2a5d663410>
> kwargs:  {'foo': 'bar'}
> Callback1 called by trigger:  <function do_notify at 0x7f2a5d663410>
> kwargs:  {'foo': 'bar'}

Thanks to the intermediary existence throughout the life of the system, A, B, and C are flexible to evolve their internals, dynamics, and lifecycles.

Subscribing and aborting events

Interestingly in Neutron, certain events may need to be forbidden from happening due to the nature of the resources involved. To this aim, the callback-based mechanism has been designed to support a use case where, when callbacks subscribe to specific events, the action that results from it, may lead to the propagation of a message back to the sender, so that it itself can be alerted and stop the execution of the activity that led to the message dispatch in the first place.

The typical example is where a resource, like a router, is used by one or more high-level service(s), like a VPN or a Firewall, and actions like interface removal or router destruction cannot not take place, because the resource is shared.

To address this scenario, special events are introduced, 'BEFORE*' events, to which callbacks can subscribe and have the opportunity to 'abort', by raising an exception when notified.

Since multiple callbacks may express an interest in the same event for a particular resource, and since callbacks are executed independently from one another, this may lead to situations where notifications that occurred before the exception must be aborted. To this aim, when an exception occurs during the notification process, an abort* event is propagated immediately after. It is up to the callback developer to determine whether subscribing to an abort notification is required in order to revert the actions performed during the initial execution of the callback (when the BEFORE* event was fired). Exceptions caused by callbacks registered to abort events are ignored. The snippet below shows this in action:

from neutron.callbacks import events
from neutron.callbacks import exceptions
from neutron.callbacks import resources
from neutron.callbacks import registry


def callback1(resource, event, trigger, **kwargs):
    raise Exception('I am failing!')

def callback2(resource, event, trigger, **kwargs):
    print('Callback2 called by %s on event  %s' % (trigger, event))


registry.subscribe(callback1, resources.ROUTER, events.BEFORE_CREATE)
registry.subscribe(callback2, resources.ROUTER, events.BEFORE_CREATE)
registry.subscribe(callback2, resources.ROUTER, events.ABORT_CREATE)
print('Subscribed')


def do_notify():
    kwargs = {'foo': 'bar'}
    registry.notify(resources.ROUTER, events.BEFORE_CREATE, do_notify, **kwargs)


print('Notifying...')
try:
    do_notify()
except exceptions.CallbackFailure as e:
    print('Error: ', e)

The output is:

> Subscribed
> Notifying...
> Callback2 called by <function do_notify at 0x7f3194c7f410> on event  before_create
> Callback2 called by <function do_notify at 0x7f3194c7f410> on event  abort_create
> Error:  Callback __main__.callback1 failed with "I am failing!"

In this case, upon the notification of the BEFORE_CREATE event, Callback1 triggers an exception that can be used to stop the action from taking place in do_notify(). On the other end, Callback2 will be executing twice, once for dealing with the BEFORE_CREATE event, and once to undo the actions during the ABORT_CREATE event. It is worth noting that it is not mandatory to have the same callback register to both BEFORE* and the respective ABORT* event; as a matter of fact, it is best to make use of different callbacks to keep the two logic separate.

As we can see from the last example, exception which is triggered in some callback will be recorded, and it will not prevent the other remaining callbacks execution. Exception triggered in callback of BEFORE_XXX will make notify process generate an ABORT_XXX event and call the related callback, while exception from PRECOMMIT_XXX will not generate ABORT_XXX event. But both of them will finally raise a unified CallbackFailure exception to the outside. For the exception triggered from other events, like AFTER_XXX and ABORT_XXX there will no exception raised to the outside.

Unsubscribing to events

There are a few options to unsubscribe registered callbacks:

  • clear(): it unsubscribes all subscribed callbacks: this can be useful especially when winding down the system, and notifications shall no longer be triggered.
  • unsubscribe(): it selectively unsubscribes a callback for a specific resource's event. Say callback C has subscribed to event A for resource R, any notification of event A for resource R will no longer be handed over to C, after the unsubscribe() invocation.
  • unsubscribe_by_resource(): say that callback C has subscribed to event A, B, and C for resource R, any notification of events related to resource R will no longer be handed over to C, after the unsubscribe_by_resource() invocation.
  • unsubscribe_all(): say that callback C has subscribed to events A, B for resource R1, and events C, D for resource R2, any notification of events pertaining resources R1 and R2 will no longer be handed over to C, after the unsubscribe_all() invocation.

The snippet below shows these concepts in action:

from neutron.callbacks import events
from neutron.callbacks import exceptions
from neutron.callbacks import resources
from neutron.callbacks import registry


def callback1(resource, event, trigger, **kwargs):
    print('Callback1 called by %s on event %s for resource %s' % (trigger, event, resource))


def callback2(resource, event, trigger, **kwargs):
    print('Callback2 called by %s on event %s for resource %s' % (trigger, event, resource))


registry.subscribe(callback1, resources.ROUTER, events.BEFORE_READ)
registry.subscribe(callback1, resources.ROUTER, events.BEFORE_CREATE)
registry.subscribe(callback1, resources.ROUTER, events.AFTER_DELETE)
registry.subscribe(callback1, resources.PORT, events.BEFORE_UPDATE)
registry.subscribe(callback2, resources.ROUTER_GATEWAY, events.BEFORE_UPDATE)
print('Subscribed')


def do_notify():
    print('Notifying...')
    kwargs = {'foo': 'bar'}
    registry.notify(resources.ROUTER, events.BEFORE_READ, do_notify, **kwargs)
    registry.notify(resources.ROUTER, events.BEFORE_CREATE, do_notify, **kwargs)
    registry.notify(resources.ROUTER, events.AFTER_DELETE, do_notify, **kwargs)
    registry.notify(resources.PORT, events.BEFORE_UPDATE, do_notify, **kwargs)
    registry.notify(resources.ROUTER_GATEWAY, events.BEFORE_UPDATE, do_notify, **kwargs)


do_notify()
registry.unsubscribe(callback1, resources.ROUTER, events.BEFORE_READ)
do_notify()
registry.unsubscribe_by_resource(callback1, resources.PORT)
do_notify()
registry.unsubscribe_all(callback1)
do_notify()
registry.clear()
do_notify()

The output is:

Subscribed
Notifying...
Callback1 called by <function do_notify at 0x7f062c8f67d0> on event before_read for resource router
Callback1 called by <function do_notify at 0x7f062c8f67d0> on event before_create for resource router
Callback1 called by <function do_notify at 0x7f062c8f67d0> on event after_delete for resource router
Callback1 called by <function do_notify at 0x7f062c8f67d0> on event before_update for resource port
Callback2 called by <function do_notify at 0x7f062c8f67d0> on event before_update for resource router_gateway
Notifying...
Callback1 called by <function do_notify at 0x7f062c8f67d0> on event before_create for resource router
Callback1 called by <function do_notify at 0x7f062c8f67d0> on event after_delete for resource router
Callback1 called by <function do_notify at 0x7f062c8f67d0> on event before_update for resource port
Callback2 called by <function do_notify at 0x7f062c8f67d0> on event before_update for resource router_gateway
Notifying...
Callback1 called by <function do_notify at 0x7f062c8f67d0> on event before_create for resource router
Callback1 called by <function do_notify at 0x7f062c8f67d0> on event after_delete for resource router
Callback2 called by <function do_notify at 0x7f062c8f67d0> on event before_update for resource router_gateway
Notifying...
Callback2 called by <function do_notify at 0x7f062c8f67d0> on event before_update for resource router_gateway
Notifying...

FAQ

Are callbacks a mechanism for remote or local communication (intra vs inter-process)?

Callbacks as described in this document are a local communication mechanism that allows multiple entities in the same process space to communicate with one another. For Neutron specific remote (IPC) mechanisms, you can see read more in RPC API <rpc_api> or Messaging callbacks <rpc_callbacks>.

Can I use the callbacks registry to subscribe and notify non-core resources and events?

Short answer is yes. The callbacks module defines literals for what are considered core Neutron resources and events. However, the ability to subscribe/notify is not limited to these as you can use your own defined resources and/or events. Just make sure you use string literals, as typos are common, and the registry does not provide any runtime validation. Therefore, make sure you test your code!

What is the relationship between Callbacks and Taskflow?

There is no overlap between Callbacks and Taskflow or mutual exclusion; as matter of fact they can be combined; You could have a callback that goes on and trigger a taskflow. It is a nice way of separating implementation from abstraction, because you can keep the callback in place and change Taskflow with something else.

Is there any ordering guarantee during notifications?

No, the ordering in which callbacks are notified is completely arbitrary by design: callbacks should know nothing about each other, and ordering should not matter; a callback will always be notified and its outcome should always be the same regardless as to in which order is it notified. Priorities can be a future extension, if a use case arises that require enforced ordering.

How is the notifying object expected to interact with the subscribing objects?

The notify method implements a one-way communication paradigm: the notifier sends a message without expecting a response back (in other words it fires and forget). However, due to the nature of Python, the payload can be mutated by the subscribing objects, and this can lead to unexpected behavior of your code, if you assume that this is the intentional design. Bear in mind, that passing-by-value using deepcopy was not chosen for efficiency reasons. Having said that, if you intend for the notifier object to expect a response, then the notifier itself would need to act as a subscriber.

Is the registry thread-safe?

Short answer is no: it is not safe to make mutations while callbacks are being called (more details as to why can be found here). A mutation could happen if a 'subscribe'/'unsubscribe' operation interleaves with the execution of the notify loop. Albeit there is a possibility that things may end up in a bad state, the registry works correctly under the assumption that subscriptions happen at the very beginning of the life of the process and that the unsubscriptions (if any) take place at the very end. In this case, chances that things do go badly may be pretty slim. Making the registry thread-safe will be considered as a future improvement.

What kind of operation I can add into callback?

For callback function of PRECOMMIT_XXX events, we can't use blocking functions or a function that would take a long time, like communicating to SDN controller over network. Callbacks for PRECOMMIT events are meant to execute DB operations in a transaction context. The errors that occur will be taken care by the context manager.

What kind of function can be a callback?

Anything you fancy: lambdas, 'closures', class, object or module methods. For instance:

from neutron.callbacks import events
from neutron.callbacks import resources
from neutron.callbacks import registry


def callback1(resource, event, trigger, **kwargs):
    print('module callback')


class MyCallback(object):

    def callback2(self, resource, event, trigger, **kwargs):
        print('object callback')

    @classmethod
    def callback3(cls, resource, event, trigger, **kwargs):
        print('class callback')


c = MyCallback()
registry.subscribe(callback1, resources.ROUTER, events.BEFORE_CREATE)
registry.subscribe(c.callback2, resources.ROUTER, events.BEFORE_CREATE)
registry.subscribe(MyCallback.callback3, resources.ROUTER, events.BEFORE_CREATE)

def do_notify():
    def nested_subscribe(resource, event, trigger, **kwargs):
        print('nested callback')

    registry.subscribe(nested_subscribe, resources.ROUTER, events.BEFORE_CREATE)

    kwargs = {'foo': 'bar'}
    registry.notify(resources.ROUTER, events.BEFORE_CREATE, do_notify, **kwargs)


print('Notifying...')
do_notify()

And the output is going to be:

Notifying...
module callback
object callback
class callback
nested callback