用在进程内部通信，Messaging Callback System是给进程间通信。为了agent不通过RPC就能得到resource的变化。

目前用在：

• QoS policies;
• Security Groups.

Using a remote publisher/subscriber pattern, the information about such resources could be published using fanout messages to all interested nodes, minimizing messaging requests from agents to server since the agents get subscribed for their whole lifecycle (unless they unsubscribe).

Within an agent, there could be multiple subscriber callbacks to the same resource events, the resources updates would be dispatched to the subscriber callbacks from a single message. Any update would come in a single message, doing only a single oslo versioned objects deserialization on each receiving agent.

message通过Oslo versioned objects 来传递以便同一消息格式。

This publishing/subscription mechanism is highly dependent on the format of the resources passed around. This is why the library only allows versioned objects to be published and subscribed. Oslo versioned objects allow object version down/up conversion.  

For the VO’s versioning schema look here: 

versioned_objects serialization/deserialization with the obj_to_primitive(target_version=..) and primitive_to_obj()  methods is used internally to convert/retrieve objects before/after messaging.

序列化Serialized versioned objects look like:

{
     'versioned_object.version': '1.0', 'versioned_object.name': 'QoSPolicy', 'versioned_object.data': {
     'rules': [                                     {
      'versioned_object.version': '1.0',                                      'versioned_object.name': 'QoSBandwidthLimitRule',                                      'versioned_object.data': {
     'name': u'a'},                                      'versioned_object.namespace': 'versionedobjects'}                                     ],                           'uuid': u'abcde',                           'name': u'aaa'}, 'versioned_object.namespace': 'versionedobjects'}

 

Rolling upgrades strategy

In this section we assume the standard Neutron upgrade process, which means upgrade the server first and then upgrade the agents:

.

The plan is to provide a semi-automatic method which avoids manual pinning and unpinning of versions by the administrator which could be prone to error.

Resource pull requests

Resource pull requests will always be ok because the underlying resource RPC does provide the version of the requested resource id / ids. The server will be upgraded first, so it will always be able to satisfy any version the agents request.

Resource push notifications

Agents will subscribe to the neutron-vo-<resource_type>-<version> fanout queue which carries updated objects for the version they know about. The versions they know about depend on the runtime Neutron versioned objects they started with.

When the server upgrades, it should be able to instantly calculate a census of agent versions per object (we will define a mechanism for this in a later section). It will use the census to send fanout messages on all the version span a resource type has.

For example, if neutron-server knew it has rpc-callback aware agents with versions 1.0, and versions 1.2 of resource type “A”, any update would be sent to neutron-vo-A_1.0 and neutron-vo-A_1.2.

TODO(mangelajo): Verify that after upgrade is finished any unused messaging resources (queues, exchanges, and so on) are released as older agents go away and neutron-server stops producing new message casts. Otherwise document the need for a neutron-server restart after rolling upgrade has finished if we want the queues cleaned up.

Leveraging agent state reports for object version discovery

We would add a row to the agent db for tracking agent known objects and version numbers. This would resemble the implementation of the configuration column.

Agents would report at start not only their configuration now, but also their subscribed object type / version pairs, that would be stored in the database and would be available to any neutron-server requesting it:

'subscribed_versions': {'QoSPolicy': '1.1',                        'SecurityGroup': '1.0',                        'Port': '1.0'}

There’s a subset of Liberty agents depending on QoSPolicy that will require ‘QoSPolicy’: ‘1.0’ if the qos plugin is installed. We will be able to identify those by the binary name (included in the report):

• ‘neutron-openvswitch-agent’
• ‘neutron-sriov-nic-agent’

Version discovery

With the above mechanism in place and considering the exception of neutron-openvswitch-agent and neutron-sriov-agent requiring QoSpolicy 1.0, we could discover the subset of versions to be sent on every push notification.

Agents that are in down state would be excluded from this calculation. We would use an extended timeout for agents in this calculation to make sure we’re on the safe side, specially if deployer marked agents with low timeouts.

Starting at Mitaka, any agent interested in versioned objects via this API should report their resource/version tuples of interest (the resource type/ version pairs they’re subscribed to).

Caching mechanism

The version subset per object will be cached to avoid DB requests on every push given that we assume that all old agents are already registered at the time of upgrade.

Cached subset will be re-evaluated (to cut down the version sets as agents upgrade) after configured TTL.

As a fast path to update this cache on all neutron-servers when upgraded agents come up (or old agents revive after a long timeout or even a downgrade) the server registering the new status update will notify the other servers about the new consumer resource versions via cast.

All notifications for all calculated version sets must be sent, as non-upgraded agents would otherwise not receive them.

It is safe to send notifications to any fanout queue as they will be discarded if no agent is listening.

Topic names for every resource type RPC endpoint

neutron-vo-<resource_class_name>-<version>

In the future, we may want to get oslo messaging to support subscribing topics dynamically, then we may want to use:

neutron-vo-<resource_class_name>-<resource_id>-<version> instead,

or something equivalent which would allow fine granularity for the receivers to only get interesting information to them.

Subscribing to resources

Imagine that you have agent A, which just got to handle a new port, which has an associated security group, and QoS policy.

The agent code processing port updates may look like:

from neutron.api.rpc.callbacks.consumer import registryfrom neutron.api.rpc.callbacks import eventsfrom neutron.api.rpc.callbacks import resourcesdef process_resource_updates(resource_type, resource, event_type):    # send to the right handler which will update any control plane    # details related to the updated resource...def subscribe_resources():    registry.subscribe(process_resource_updates, resources.SEC_GROUP)    registry.subscribe(process_resource_updates, resources.QOS_POLICY)def port_update(port):    # here we extract sg_id and qos_policy_id from port..    sec_group = registry.pull(resources.SEC_GROUP, sg_id)    qos_policy = registry.pull(resources.QOS_POLICY, qos_policy_id)

The relevant function is:

• subscribe(callback, resource_type): subscribes callback to a resource type.

The callback function will receive the following arguments:

• resource_type: the type of resource which is receiving the update.
• resource: resource of supported object
• event_type: will be one of CREATED, UPDATED, or DELETED, see neutron.api.rpc.callbacks.events for details.

With the underlaying oslo_messaging support for dynamic topics on the receiver we cannot implement a per “resource type + resource id” topic, rabbitmq seems to handle 10000’s of topics without suffering, but creating 100’s of oslo_messaging receivers on different topics seems to crash.

We may want to look into that later, to avoid agents receiving resource updates which are uninteresting to them.

Unsubscribing from resources

To unsubscribe registered callbacks:

• unsubscribe(callback, resource_type): unsubscribe from specific resource type.
• unsubscribe_all(): unsubscribe from all resources.

Sending resource events

On the server side, resource updates could come from anywhere, a service plugin, an extension, anything that updates, creates, or destroys the resource and that is of any interest to subscribed agents.

The server/publisher side may look like:

from neutron.api.rpc.callbacks.producer import registryfrom neutron.api.rpc.callbacks import events def create_qos_policy(...): policy = fetch_policy(...) update_the_db(...) registry.push(policy, events.CREATED) def update_qos_policy(...): policy = fetch_policy(...) update_the_db(...) registry.push(policy, events.UPDATED) def delete_qos_policy(...): policy = fetch_policy(...) update_the_db(...) registry.push(policy, events.DELETED)

References

http://docs.openstack.org/developer/neutron/devref/rpc_callbacks.html