Add notification support to ironic

https://bugs.launchpad.net/ironic/+bug/1526408

Near real-time inter-service notifications in OpenStack have a variety of use cases. At the Mitaka design summit, some of the use cases discussed for ironic included integration with billing systems, using notifications to support an operations panel to keep track of potential failures, and integration with TripleO. [1] Nova has supported some form of notifications since the Diablo cycle. [2] This spec proposes a design for a general notification schema that’s versioned and structured but extensible enough to support future needs.

Problem description

Ironic doesn’t currently support any type of notifications for external services to consume.

This makes it difficult for operators to monitor and keep track of events in ironic like node state changes, node power events, and successful or failed deployments. Sending notifications that contain relevant information about a node at different points in time, like state transitions, would make it easier for operators to use external tools to debug unexpected behavior.

Not having notifications also makes it difficult for external services to identify state changes in ironic. Nova and inspector have to poll ironic repeatedly to see when a state change has occurred. Adding notifications for cases like these will allow ironic to push out these event changes rather than putting the bulk of this burden on external services.

This is especially problematic if ironic were ever to be used in a standalone deployment without an external service like Nova that keeps track of resources and emits its own notifications, since the deployers of ironic wouldn’t have any way to keep track of any state changes outside of querying the ironic API periodically or looking at logs.

Proposed change

Notification support will be added to ironic using the notifier interface from oslo.messaging. [3] oslo.messaging uses the following format for notifications:

{'message_id': six.text_type(uuid.uuid4()),
 'publisher_id': 'compute.host1',
 'timestamp': timeutils.utcnow(),
 'priority': 'WARN',
 'event_type': 'compute.create_instance',
 'payload': {'instance_id': 12, ... }}

Ironic notifications will adhere to that format.

Base classes will be created to model the notification format. The fields of the notification that aren’t auto-generated (everything except message_id and timestamp) will be defined as follows for ironic:

  • publisher_id will be taken from the service emitting the notification and the hostname

  • priority will be one of DEBUG, WARN, INFO, or ERROR, depending on the notification. Whichever level a notification uses should follow the OpenStack logging standards. [4]

  • event_type will be a short string describing the notification. Each string will start with “baremetal.” to distinguish ironic notifications from other notifications on the message bus.

    This will be followed by the object that’s being acted on, a descriptor of the action being taken, and the status of the action (“start”, “end”, “error”, or “success”). These statuses will have the following semantics:

    • “start” will be used when an action that is not immediate begins.
    • “end” will be used when an action that is not immediate succeeds.
    • “error” will be used when an action fails, regardless of whether the action is immediate or not.
    • “success” will be used when an action that succeeds almost immediately, and there is no need to report the intermediate resource states.

    event_type will have a base class defining these fields, with subclasses for each type of event that occurs. These subclasses will be versioned objects. Each of these fields will be separated by a period in the event_type string sent with the notification.

    Given the above information, event_type will end up as a string on the message bus with the following format: baremetal.<object>.<action>.<status>

    A few examples of potential event_type strings:

    • “baremetal.node.power_set.start”
    • “baremetal.node.power_set.end”
    • “baremetal.node.provision_set.start”
    • “baremetal.node.provision_set.end”
    • “baremetal.conductor.init.success”
    • “baremetal.conductor.stop.success”
  • payload will be a versioned object related to the notification topic. This should include relevant information about the event being notified on. For the notifications above about a node, for example, we would send an object containing a limited subset of the fields of the Node object, removing sensitive fields like private credentials as well as fields with an unknown length, along with the version of the notification object. A separate versioned notification object must be created for each type of notification to keep notification versions separate from other object versions in ironic. This is to prevent notifications from automatically changing as new fields are added to other objects like the Node object.

The initial implementation will consist of the base classes needed for modeling the notification format and emitting notifications as well as the power state notifications described above.

Alternatives

One alternative would be to modify oslo.messaging’s notifier interface to provide the base class defining a notification as a versioned object. The advantage of this is that we wouldn’t have to rely on the payload to do versioning. The disadvantage is that it would require significant changes to oslo.messaging which affect other projects and require changes to every notification consumer as well as any downstream tooling built around the existing notification format.

We could also have a global notification version that gets incremented with each change to any notification. This doesn’t seem desirable because it could potentially miss changes if a separate ironic object included in the payload like the Node object gets changed and the notification itself doesn’t get its version bumped.

Data model impact

No database changes will be required for this feature.

The change will introduce new versioned base objects for the general notification schema and additional subclasses for individual notification types. The base classes will look similar to those in the proposal in the nova versioned notifications spec. [5]

State Machine Impact

No impact on states or transitions themselves.

Notifications will be sent out on node provision state changes if notifications are enabled. This can be achieved by sending notifications every time a TaskManager’s process_event method starts and successfully executes.

All information related to a node’s previous, current, and new target state will be included in the notifications. The .start notification will have the current provision_state before the state change and the target_provision_state. After an event is successfully processed, the .end notification will include the current provision_state (the .start notification’s target state) and the new target_provision_state. The notifications will also include the name of the event that caused the state change. This will be useful for disambiguating between cases where there are multiple potential transitions from one state to another.

REST API impact

None.

Client (CLI) impact

None.

RPC API impact

No impact from an API standpoint.

Modifications to the implementation of certain conductor RPC API methods will need to be made for notifications that are sent when an RPC is dispatched to a worker, however. See Driver API Impact for an example of how this might be done for power notifications.

Driver API impact

No impact from an API standpoint.

Notifications related to power state changes will be added, but that can be done without modifying any of the driver classes in the following manner:

  1. Send a baremetal.node.set_power_state.start notification after the ConductorManager receives the change_node_power_state call as a conductor background task.
  2. On success, after the dispatched call to node_power_action finishes without raising an exception, send a baremetal.node.set_power_state.end notification.
  3. On error, the power_state_error_handler hook will be called in the conductor manager. Send a baremetal.node.set_power_state.error notification here.

Nova driver impact

None.

Ramdisk impact

N/A

Security impact

None.

Other end user impact

None, except a message bus will have to be used if a deployer wants to use the notification system.

Scalability impact

When enabled, notifications will put additional load on whichever message bus the notifications are sent to.

Performance Impact

When enabled, code to send the notification will be called each time an event occurs that triggers a notification. This shouldn’t be much of a problem for ironic itself, but the load on whatever message bus is used should be considered (see Scalability Impact).

Other deployer impact

The following configuration options will be added:

  • The notification_transport_url option needed by oslo.messaging. [6] Defaults to None which indicates that the same configuration that’s used for RPC will be used.

  • A notification_level string parameter will be added to indicate the minimum priority level for which notifications will be sent. Available options will be DEBUG, INFO, WARN, ERROR, or None to disable notifications. None will be the default.

    An alternative to the notification_level global config option would be to create specific config options defining whether a particular notification type should be sent. This is what nova does, but summit discussions indicated that consistency is preferable.

Developer impact

Developers should adhere to proper versioning guidelines and use the notification base classes when creating new notifications.

Implementation

Assignee(s)

Primary assignee:
  • mariojv
Other contributors:
  • lucasagomes

Work Items

  • Create notification base classes and tests
  • Write documentation for how to use the base classes consistently across all ironic notifications
  • Implement an example of a notification for when a node power state is changed

Dependencies

None.

Testing

Unit tests for both the base classes and the node power state notification will be added.

Upgrades and Backwards Compatibility

No impact, but modifications to notifications created in the future must be checked for backwards compatibility.

Documentation Impact

  • Developer documentation will be added for how to add new notifications or modify existing notifications
  • Document an example of what an emitted notification will look like
  • Documentation should be added for each notification indicating when it’s expected to be emitted