1 Detailed architecture and interface specification
2 =================================================
4 This section describes a detailed implementation plan, which is based on the
5 high level architecture introduced in Section 3. Section 5.1 describes the
6 functional blocks of the Doctor architecture, which is followed by a high level
7 message flow in Section 5.2. Section 5.3 provides a mapping of selected existing
8 open source components to the building blocks of the Doctor architecture.
9 Thereby, the selection of components is based on their maturity and the gap
10 analysis executed in Section 4. Sections 5.4 and 5.5 detail the specification of
11 the related northbound interface and the related information elements. Finally,
12 Section 5.6 provides a first set of blueprints to address selected gaps required
13 for the realization functionalities of the Doctor project.
18 This section introduces the functional blocks to form the VIM. OpenStack was
19 selected as the candidate for implementation. Inside the VIM, 4 different
20 building blocks are defined (see :num:`Figure #figure6`).
24 .. figure:: images/figure6.png
32 The Monitor module has the responsibility for monitoring the virtualized
33 infrastructure. There are already many existing tools and services (e.g. Zabbix)
34 to monitor different aspects of hardware and software resources which can be
35 used for this purpose.
40 The Inspector module has the ability a) to receive various failure notifications
41 regarding physical resource(s) from Monitor module(s), b) to find the affected
42 virtual resource(s) by querying the resource map in the Controller, and c) to
43 update the state of the virtual resource (and physical resource).
45 The Inspector has drivers for different types of events and resources to
46 integrate any type of Monitor and Controller modules. It also uses a failure
47 policy database to decide on the failure selection and aggregation from raw
48 events. This failure policy database is configured by the Administrator.
50 The reason for separation of the Inspector and Controller modules is to make the
51 Controller focus on simple operations by avoiding a tight integration of various
52 health check mechanisms into the Controller.
57 The Controller is responsible for maintaining the resource map (i.e. the mapping
58 from physical resources to virtual resources), accepting update requests for the
59 resource state(s) (exposing as provider API), and sending all failure events
60 regarding virtual resources to the Notifier. Optionally, the Controller has the
61 ability to poison the state of virtual resources mapping to physical resources
62 for which it has received failure notifications from the Inspector. The
63 Controller also re-calculates the capacity of the NVFI when receiving a failure
64 notification for a physical resource.
66 In a real-world deployment, the VIM may have several controllers, one for each
67 resource type, such as Nova, Neutron and Cinder in OpenStack. Each controller
68 maintains a database of virtual and physical resources which shall be the master
69 source for resource information inside the VIM.
74 The focus of the Notifier is on selecting and aggregating failure events
75 received from the controller based on policies mandated by the Consumer.
76 Therefore, it allows the Consumer to subscribe for alarms regarding virtual
77 resources using a method such as API endpoint. After receiving a fault
78 event from a Controller, it will notify the fault to the Consumer by referring
79 to the alarm configuration which was defined by the Consumer earlier on.
81 To reduce complexity of the Controller, it is a good approach for the
82 Controllers to emit all notifications without any filtering mechanism and have
83 another service (i.e. Notifier) handle those notifications properly. This is the
84 general philosophy of notifications in OpenStack. Note that a fault message
85 consumed by the Notifier is different from the fault message received by the
86 Inspector; the former message is related to virtual resources which are visible
87 to users with relevant ownership, whereas the latter is related to raw devices
88 or small entities which should be handled with an administrator privilege.
90 The northbound interface between the Notifier and the Consumer/Administrator is
91 specified in Section 5.5.
99 The detailed work flow for fault management is as follows (see also :num:`Figure
102 1. Request to subscribe to monitor specific virtual resources. A query filter
103 can be used to narrow down the alarms the Consumer wants to be informed
105 2. Each subscription request is acknowledged with a subscribe response message.
106 The response message contains information about the subscribed virtual
107 resources, in particular if a subscribed virtual resource is in "alarm"
109 3. The NFVI sends monitoring events for resources the VIM has been subscribed
110 to. Note: this subscription message exchange between the VIM and NFVI is not
111 shown in this message flow.
112 4. Event correlation, fault detection and aggregation in VIM.
113 5. Database lookup to find the virtual resources affected by the detected fault.
114 6. Fault notification to Consumer.
115 7. The Consumer switches to standby configuration (STBY)
116 8. Instructions to VIM requesting certain actions to be performed on the
117 affected resources, for example migrate/update/terminate specific
118 resource(s). After reception of such instructions, the VIM is executing the
119 requested action, e.g. it will migrate or terminate a virtual resource.
121 a. Query request from Consumer to VIM to get information about the current
122 status of a resource.
123 b. Response to the query request with information about the current status of
124 the queried resource. In case the resource is in "fault" state, information
125 about the related fault(s) is returned.
127 In order to allow for quick reaction to failures, the time interval between
128 fault detection in step 3 and the corresponding recovery actions in step 7 and 8
129 shall be less than 1 second.
133 .. figure:: images/figure7.png
136 Fault management work flow
141 .. figure:: images/figure8.png
144 Fault management scenario
146 :num:`Figure #figure8` shows a more detailed message flow (Steps 4 to 6) between
147 the 4 building blocks introduced in Section 5.1.
149 4. The Monitor observed a fault in the NFVI and reports the raw fault to the
151 The Inspector filters and aggregates the faults using pre-configured
155 a) The Inspector queries the Resource Map to find the virtual resources
156 affected by the raw fault in the NFVI.
157 b) The Inspector updates the state of the affected virtual resources in the
159 c) The Controller observes a change of the virtual resource state and informs
160 the Notifier about the state change and the related alarm(s).
161 Alternatively, the Inspector may directly inform the Notifier about it.
163 6. The Notifier is performing another filtering and aggregation of the changes
164 and alarms based on the pre-configured alarm configuration. Finally, a fault
165 notification is sent to northbound to the Consumer.
170 The detailed work flow for NFVI maintenance is shown in :num:`Figure #figure9`
171 and has the following steps. Note that steps 1, 2, and 5 to 8a in the NFVI
172 maintenance work flow are very similar to the steps in the fault management work
173 flow and share a similar implementation plan in Release 1.
175 1. Subscribe to fault/maintenance notifications.
176 2. Response to subscribe request.
177 3. Maintenance trigger received from administrator.
178 4. VIM switches NFVI resources to "maintenance" state. This, e.g., means they
179 should not be used for further allocation/migration requests
180 5. Database lookup to find the virtual resources affected by the detected
181 maintenance operation.
182 6. Maintenance notification to Consumer.
183 7. The Consumer switches to standby configuration (STBY)
184 8. Instructions from Consumer to VIM requesting certain recovery actions to be
185 performed (step 7a). After reception of such instructions, the VIM is
186 executing the requested action in order to empty the physical resources (step
188 9. Maintenance response from VIM to inform the Administrator that the physical
189 machines have been emptied (or the operation resulted in an error state).
190 10. Administrator is coordinating and executing the maintenance operation/work
193 A) Query request from Administrator to VIM to get information about the
194 current state of a resource.
195 B) Response to the query request with information about the current state of
196 the queried resource(s). In case the resource is in "maintenance" state,
197 information about the related maintenance operation is returned.
201 .. figure:: images/figure9.png
204 NFVI maintenance work flow
209 .. figure:: images/figure10.png
212 NFVI Maintenance implementation plan
214 :num:`Figure #figure10` shows a more detailed message flow (Steps 4 to 6)
215 between the 4 building blocks introduced in Section 5.1..
217 3. The Administrator is sending a StateChange request to the Controller residing
219 4. The Controller queries the Resource Map to find the virtual resources
220 affected by the planned maintenance operation.
223 a) The Controller updates the state of the affected virtual resources in the
224 Resource Map database.
226 b) The Controller informs the Notifier about the virtual resources that will
227 be affected by the maintenance operation.
229 6. A maintenance notification is sent to northbound to the Consumer.
233 9. The Controller informs the Administrator after the physical resources have
238 Implementation plan for OPNFV Release 1
239 ---------------------------------------
244 :num:`Figure #figure11` shows the implementation plan based on OpenStack and
245 related components as planned for Release 1. Hereby, the Monitor can be realized
246 by Zabbix. The Controller is realized by OpenStack Nova [NOVA]_, Neutron
247 [NEUT]_, and Cinder [CIND]_ for compute, network, and storage,
248 respectively. The Inspector can be realized by Monasca [MONA]_ or a simple
249 script querying Nova in order to map between physical and virtual resources. The
250 Notifier will be realized by Ceilometer [CEIL]_ receiving failure events
251 on its notification bus.
253 :num:`Figure #figure12` shows the inner-workings of Ceilometer. After receiving
254 an "event" on its notification bus, first a notification agent will grab the
255 event and send a "notification" to the Collector. The collector writes the
256 notifications received to the Ceilometer databases.
258 In the existing Ceilometer implementation, an alarm evaluator is periodically
259 polling those databases through the APIs provided. If it finds new alarms, it
260 will evaluate them based on the pre-defined alarm configuration, and depending
261 on the configuration, it will hand a message to the Alarm Notifier, which in
262 turn will send the alarm message northbound to the Consumer. :num:`Figure
263 #figure12` also shows an optimized work flow for Ceilometer with the goal to
264 reduce the delay for fault notifications to the Consumer. The approach is to
265 implement a new notification agent (called "publisher" in Ceilometer
266 terminology) which is directly sending the alarm through the "Notification Bus"
267 to a new "Notification-driven Alarm Evaluator (NAE)" (see Sections 5.6.2 and
268 5.6.3), thereby bypassing the Collector and avoiding the additional delay of the
269 existing polling-based alarm evaluator. The NAE is similar to the OpenStack
270 "Alarm Evaluator", but is triggered by incoming notifications instead of
271 periodically polling the OpenStack "Alarms" database for new alarms. The
272 Ceilometer "Alarms" database can hold three states: "normal", "insufficient
273 data", and "fired". It is representing a persistent alarm database. In order to
274 realize the Doctor requirements, we need to define new "meters" in the database
279 .. figure:: images/figure11.png
282 Implementation plan in OpenStack (OPNFV Release 1 ”Arno”)
287 .. figure:: images/figure12.png
290 Implementation plan in Ceilometer architecture
296 For NFVI Maintenance, a quite similar implementation plan exists. Instead of a
297 raw fault being observed by the Monitor, the Administrator is sending a
298 Maintenance Request through the northbound interface towards the Controller
299 residing in the VIM. Similar to the Fault Management use case, the Controller
300 (in our case OpenStack Nova) will send a maintenance event to the Notifier (i.e.
301 Ceilometer in our implementation). Within Ceilometer, the same workflow as
302 described in the previous section applies. In addition, the Controller(s) will
303 take appropriate actions to evacuate the physical machines in order to prepare
304 them for the planned maintenance operation. After the physical machines are
305 emptied, the Controller will inform the Administrator that it can initiate the
311 This section introduces all attributes and information elements used in the
312 messages exchange on the northbound interfaces between the VIM and the VNFO and
315 Note: The information elements will be aligned with current work in ETSI NFV IFA
319 Simple information elements:
321 * SubscriptionID: identifies a subscription to receive fault or maintenance
323 * NotificationID: identifies a fault or maintenance notification.
324 * VirtualResourceID (Identifier): identifies a virtual resource affected by a
325 fault or a maintenance action of the underlying physical resource.
326 * PhysicalResourceID (Identifier): identifies a physical resource affected by a
327 fault or maintenance action.
328 * VirtualResourceState (String): state of a virtual resource, e.g. "normal",
329 "maintenance", "down", "error".
330 * PhysicalResourceState (String): state of a physical resource, e.g. "normal",
331 "maintenance", "down", "error".
332 * VirtualResourceType (String): type of the virtual resource, e.g. "virtual
333 machine", "virtual memory", "virtual storage", "virtual CPU", or "virtual
335 * FaultID (Identifier): identifies the related fault in the underlying physical
336 resource. This can be used to correlate different fault notifications caused
337 by the same fault in the physical resource.
338 * FaultType (String): Type of the fault. The allowed values for this parameter
339 depend on the type of the related physical resource. For example, a resource
340 of type "compute hardware" may have faults of type "CPU failure", "memory
341 failure", "network card failure", etc.
342 * Severity (Integer): value expressing the severity of the fault. The higher the
343 value, the more severe the fault.
344 * MinSeverity (Integer): value used in filter information elements. Only faults
345 with a severity higher than the MinSeverity value will be notified to the
347 * EventTime (Datetime): Time when the fault was observed.
348 * EventStartTime and EventEndTime (Datetime): Datetime range that can be used in
349 a FaultQueryFilter to narrow down the faults to be queried.
350 * ProbableCause: information about the probable cause of the fault.
351 * CorrelatedFaultID (Integer): list of other faults correlated to this fault.
352 * isRootCause (Boolean): Parameter indicating if this fault is the root for
353 other correlated faults. If TRUE, then the faults listed in the parameter
354 CorrelatedFaultID are caused by this fault.
355 * FaultDetails (Key-value pair): provides additional information about the
356 fault, e.g. information about the threshold, monitored attributes, indication
357 of the trend of the monitored parameter.
358 * FirmwareVersion (String): current version of the firmware of a physical
360 * HypervisorVersion (String): current version of a hypervisor.
361 * ZoneID (Identifier): Identifier of the resource zone. A resource zone is the
362 logical separation of physical and software resources in an NFVI deployment
363 for physical isolation, redundancy, or administrative designation.
364 * Metadata (Key-Value-Pairs): provides additional information of a physical
365 resource in maintenance/error state.
367 Complex information elements (see also UML diagrams in :num:`Figure #figure13`
368 and :num:`Figure #figure14`):
370 * VirtualResourceInfoClass:
372 + VirtualResourceID [1] (Identifier)
373 + VirtualResourceState [1] (String)
374 + Faults [0..*] (FaultClass): For each resource, all faults
375 including detailed information about the faults are provided.
377 * FaultClass: The parameters of the FaultClass are partially based on ETSI TS
378 132 111-2 (V12.1.0) [*]_, which is specifying fault management in 3GPP, in
379 particular describing the information elements used for alarm notifications.
381 - FaultID [1] (Identifier)
383 - Severity [1] (Integer)
384 - EventTime [1] (Datetime)
386 - CorrelatedFaultID [0..*] (Identifier)
387 - FaultDetails [0..*] (Key-value pair)
389 .. [*] http://www.etsi.org/deliver/etsi_ts/132100_132199/13211102/12.01.00_60/ts_13211102v120100p.pdf
391 * SubscribeFilterClass
393 - VirtualResourceType [0..*] (String)
394 - VirtualResourceID [0..*] (Identifier)
395 - FaultType [0..*] (String)
396 - MinSeverity [0..1] (Integer)
398 * FaultQueryFilterClass: narrows down the FaultQueryRequest, for example it
399 limits the query to certain physical resources, a certain zone, a given fault
400 type/severity/cause, or a specific FaultID.
402 - VirtualResourceType [0..*] (String)
403 - VirtualResourceID [0..*] (Identifier)
404 - FaultType [0..*] (String)
405 - MinSeverity [0..1] (Integer)
406 - EventStartTime [0..1] (Datetime)
407 - EventEndTime [0..1] (Datetime)
409 * PhysicalResourceStateClass:
411 - PhysicalResourceID [1] (Identifier)
412 - PhysicalResourceState [1] (String): mandates the new state of the physical
415 * PhysicalResourceInfoClass:
417 - PhysicalResourceID [1] (Identifier)
418 - PhysicalResourceState [1] (String)
419 - FirmwareVersion [0..1] (String)
420 - HypervisorVersion [0..1] (String)
421 - ZoneID [0..1] (Identifier)
423 * StateQueryFilterClass: narrows down a StateQueryRequest, for example it limits
424 the query to certain physical resources, a certain zone, or a given resource
425 state (e.g., only resources in "maintenance" state).
427 - PhysicalResourceID [1] (Identifier)
428 - PhysicalResourceState [1] (String)
429 - ZoneID [0..1] (Identifier)
431 Detailed northbound interface specification
432 -------------------------------------------
434 This section is specifying the northbound interfaces for fault management and
435 NFVI maintenance between the VIM on the one end and the Consumer and the
436 Administrator on the other ends. For each interface all messages and related
437 information elements are provided.
439 Note: The interface definition will be aligned with current work in ETSI NFV IFA
442 All of the interfaces described below are produced by the VIM and consumed by
443 the Consumer or Administrator.
445 Fault management interface
446 ^^^^^^^^^^^^^^^^^^^^^^^^^^
448 This interface allows the VIM to notify the Consumer about a virtual resource
449 that is affected by a fault, either within the virtual resource itself or by the
450 underlying virtualization infrastructure. The messages on this interface are
451 shown in :num:`Figure #figure13` and explained in detail in the following
454 Note: The information elements used in this section are described in detail in
459 .. figure:: images/figure13.png
462 Fault management NB I/F messages
465 SubscribeRequest (Consumer -> VIM)
466 __________________________________
468 Subscription from Consumer to VIM to be notified about faults of specific
469 resources. The faults to be notified about can be narrowed down using a
474 - SubscribeFilter [1] (SubscribeFilterClass): Optional information to narrow
475 down the faults that shall be notified to the Consumer, for example limit to
476 specific VirtualResourceID(s), severity, or cause of the alarm.
478 SubscribeResponse (VIM -> Consumer)
479 ___________________________________
481 Response to a subscribe request message including information about the
482 subscribed resources, in particular if they are in "fault/error" state.
486 * SubscriptionID [1] (Identifier): Unique identifier for the subscription. It
487 can be used to delete or update the subscription.
488 * VirtualResourceInfo [0..*] (VirtualResourceInfoClass): Provides additional
489 information about the subscribed resources, i.e., a list of the related
490 resources, the current state of the resources, etc.
492 FaultNotification (VIM -> Consumer)
493 ___________________________________
495 Notification about a virtual resource that is affected by a fault, either within
496 the virtual resource itself or by the underlying virtualization infrastructure.
497 After reception of this request, the Consumer will decide on the optimal
498 action to resolve the fault. This includes actions like switching to a hot
499 standby virtual resource, migration of the fault virtual resource to another
500 physical machine, termination of the faulty virtual resource and instantiation
501 of a new virtual resource in order to provide a new hot standby resource.
502 Existing resource management interfaces and messages between the Consumer and
503 the VIM can be used for those actions, and there is no need to define additional
504 actions on the Fault Management Interface.
508 * NotificationID [1] (Identifier): Unique identifier for the notification.
509 * VirtualResourceInfo [1..*] (VirtualResourceInfoClass): List of faulty
510 resources with detailed information about the faults.
512 FaultQueryRequest (Consumer -> VIM)
513 ___________________________________
515 Request to find out about active alarms at the VIM. A FaultQueryFilter can be
516 used to narrow down the alarms returned in the response message.
520 * FaultQueryFilter [1] (FaultQueryFilterClass): narrows down the
521 FaultQueryRequest, for example it limits the query to certain physical
522 resources, a certain zone, a given fault type/severity/cause, or a specific
525 FaultQueryResponse (VIM -> Consumer)
526 ____________________________________
528 List of active alarms at the VIM matching the FaultQueryFilter specified in the
533 * VirtualResourceInfo [0..*] (VirtualResourceInfoClass): List of faulty
534 resources. For each resource all faults including detailed information about
535 the faults are provided.
540 The NFVI maintenance interfaces Consumer-VIM allows the Consumer to subscribe to
541 maintenance notifications provided by the VIM. The related maintenance interface
542 Administrator-VIM allows the Administrator to issue maintenance requests to the
543 VIM, i.e. requesting the VIM to take appropriate actions to empty physical
544 machine(s) in order to execute maintenance operations on them. The interface
545 also allows the Administrator to query the state of physical machines, e.g., in
546 order to get details in the current status of the maintenance operation like a
549 The messages defined in these northbound interfaces are shown in :num:`Figure
550 #figure14` and described in detail in the following subsections.
554 .. figure:: images/figure14.png
557 NFVI maintenance NB I/F messages
559 SubscribeRequest (Consumer -> VIM)
560 __________________________________
562 Subscription from Consumer to VIM to be notified about maintenance operations
563 for specific virtual resources. The resources to be informed about can be
564 narrowed down using a subscribe filter.
568 * SubscribeFilter [1] (SubscribeFilterClass): Information to narrow down the
569 faults that shall be notified to the Consumer, for example limit to specific
570 virtual resource type(s).
572 SubscribeResponse (VIM -> Consumer)
573 ___________________________________
575 Response to a subscribe request message, including information about the
576 subscribed virtual resources, in particular if they are in "maintenance" state.
580 * SubscriptionID [1] (Identifier): Unique identifier for the subscription. It
581 can be used to delete or update the subscription.
582 * VirtualResourceInfo [0..*] (VirtalResourceInfoClass): Provides additional
583 information about the subscribed virtual resource(s), e.g., the ID, type and
584 current state of the resource(s).
586 MaintenanceNotification (VIM -> Consumer)
587 _________________________________________
589 Notification about a physical resource switched to "maintenance" state. After
590 reception of this request, the Consumer will decide on the optimal action to
591 address this request, e.g., to switch to the standby (STBY) configuration.
595 * VirtualResourceInfo [1..*] (VirtualResourceInfoClass): List of virtual
596 resources where the state has been changed to maintenance.
598 StateChangeRequest (Administrator -> VIM)
599 _________________________________________
601 Request to change the state of a list of physical resources, e.g. to
602 "maintenance" state, in order to prepare them for a planned maintenance
607 * PhysicalResourceState [1..*] (PhysicalResourceStateClass)
609 StateChangeResponse (VIM -> Administrator)
610 __________________________________________
612 Response message to inform the Administrator that the requested resources are
613 now in maintenance state (or the operation resulted in an error) and the
614 maintenance operation(s) can be executed.
618 * PhysicalResourceInfo [1..*] (PhysicalResourceInfoClass)
620 StateQueryRequest (Administrator -> VIM)
621 ________________________________________
623 In this procedure, the Administrator would like to get the information about
624 physical machine(s), e.g. their state ("normal", "maintenance"), firmware
625 version, hypervisor version, update status of firmware and hypervisor, etc. It
626 can be used to check the progress during firmware update and the confirmation
627 after update. A filter can be used to narrow down the resources returned in the
632 * StateQueryFilter [1] (StateQueryFilterClass): narrows down the
633 StateQueryRequest, for example it limits the query to certain physical
634 resources, a certain zone, or a given resource state.
636 StateQueryResponse (VIM -> Administrator)
637 _________________________________________
639 List of physical resources matching the filter specified in the
644 * PhysicalResourceInfo [0..*] (PhysicalResourceInfoClass): List of physical
645 resources. For each resource, information about the current state, the
646 firmware version, etc. is provided.
651 This section is listing a first set of blueprints that have been proposed by the
652 Doctor project to the open source community. Further blueprints addressing other
653 gaps identified in Section 4 will be submitted at a later stage of the OPNFV. In
654 this section the following definitions are used:
656 * "Event" is a message emitted by other OpenStack services such as Nova and
657 Neutron and is consumed by the "Notification Agents" in Ceilometer.
658 * "Notification" is a message generated by a "Notification Agent" in Ceilometer
659 based on an "event" and is delivered to the "Collectors" in Ceilometer that
660 store those notifications (as "sample") to the Ceilometer "Databases".
662 Instance State Notification (Ceilometer) [*]_
663 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
665 The Doctor project is planning to handle "events" and "notifications" regarding
666 Resource Status; Instance State, Port State, Host State, etc. Currently,
667 Ceilometer already receives "events" to identify the state of those resources,
668 but it does not handle and store them yet. This is why we also need a new event
669 definition to capture those resource states from "events" created by other
672 This BP proposes to add a new compute notification state to handle events from
673 an instance (server) from nova. It also creates a new meter "instance.state" in
676 .. [*] https://etherpad.opnfv.org/p/doctor_bps
678 Event Publisher for Alarm (Ceilometer) [*]_
679 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
681 **Problem statement:**
683 The existing "Alarm Evaluator" in OpenStack Ceilometer is periodically
684 querying/polling the databases in order to check all alarms independently from
685 other processes. This is adding additional delay to the fault notification
686 send to the Consumer, whereas one requirement of Doctor is to react on faults
689 The existing message flow is shown in :num:`Figure #figure12`: after receiving
690 an "event", a "notification agent" (i.e. "event publisher") will send a
691 "notification" to a "Collector". The "collector" is collecting the
692 notifications and is updating the Ceilometer "Meter" database that is storing
693 information about the "sample" which is capured from original "event". The
694 "Alarm Evaluator" is periodically polling this databases then querying "Meter"
695 database based on each alarm configuration.
697 In the current Ceilometer implementation, there is no possibility to directly
698 trigger the "Alarm Evaluator" when a new "event" was received, but the "Alarm
699 Evaluator" will only find out that requires firing new notification to the
700 Consumer when polling the database.
702 **Change/feature request:**
704 This BP proposes to add a new "event publisher for alarm", which is bypassing
705 several steps in Ceilometer in order to avoid the polling-based approach of
706 the existing Alarm Evaluator that makes notification slow to users.
708 After receiving an "(alarm) event" by listening on the Ceilometer message
709 queue ("notification bus"), the new "event publisher for alarm" immediately
710 hands a "notification" about this event to a new Ceilometer component
711 "Notification-driven alarm evaluator" proposed in the other BP (see Section
714 Note, the term "publisher" refers to an entity in the Ceilometer architecture
715 (it is a "notification agent"). It offers the capability to provide
716 notifications to other services outside of Ceilometer, but it is also used to
717 deliver notifications to other Ceilometer components (e.g. the "Collectors")
718 via the Ceilometer "notification bus".
720 **Implementation detail**
722 * "Event publisher for alarm" is part of Ceilometer
723 * The standard AMQP message queue is used with a new topic string.
724 * No new interfaces have to be added to Ceilometer.
725 * "Event publisher for Alarm" can be configured by the Administrator of
726 Ceilometer to be used as "Notification Agent" in addition to the existing
728 * Existing alarm mechanisms of Ceilometer can be used allowing users to
729 configure how to distribute the "notifications" transformed from "events",
730 e.g. there is an option whether an ongoing alarm is re-issued or not
733 .. [*] https://etherpad.opnfv.org/p/doctor_bps
735 Notification-driven alarm evaluator (Ceilometer) [*]_
736 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
738 **Problem statement:**
740 The existing "Alarm Evaluator" in OpenStack Ceilometer is periodically
741 querying/polling the databases in order to check all alarms independently from
742 other processes. This is adding additional delay to the fault notification send
743 to the Consumer, whereas one requirement of Doctor is to react on faults as fast
746 **Change/feature request:**
748 This BP is proposing to add an alternative "Notification-driven Alarm Evaluator"
749 for Ceilometer that is receiving "notifications" sent by the "Event Publisher
750 for Alarm" described in the other BP. Once this new "Notification-driven Alarm
751 Evaluator" received "notification", it finds the "alarm" configurations which
752 may relate to the "notification" by querying the "alarm" database with some keys
753 i.e. resource ID, then it will evaluate each alarm with the information in that
756 After the alarm evaluation, it will perform the same way as the existing "alarm
757 evaluator" does for firing alarm notification to the Consumer. Similar to the
758 existing Alarm Evaluator, this new "Notification-driven Alarm Evaluator" is
759 aggregating and correlating different alarms which are then provided northbound
760 to the Consumer via the OpenStack "Alarm Notifier". The user/administrator can
761 register the alarm configuration via existing Ceilometer API [*]_. Thereby, he
762 can configure whether to set an alarm or not and where to send the alarms to.
764 **Implementation detail**
766 * The new "Notification-driven Alarm Evaluator" is part of Ceilometer.
767 * Most of the existing source code of the "Alarm Evaluator" can be re-used to
769 * No additional application logic is needed
770 * It will access the Ceilometer Databases just like the existing "Alarm
772 * Only the polling-based approach will be replaced by a listener for
773 "notifications" provided by the "Event Publisher for Alarm" on the Ceilometer
775 * No new interfaces have to be added to Ceilometer.
778 .. [*] https://etherpad.opnfv.org/p/doctor_bps
779 .. [*] https://wiki.openstack.org/wiki/Ceilometer/Alerting
781 Report host fault to update server state immediately (Nova) [*]_
782 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
784 **Problem statement:**
786 * Nova state change for failed or unreachable host is slow and does not reliably
787 state host is down or not. This might cause same server instance to run twice
788 if action taken to evacuate instance to another host.
789 * Nova state for server(s) on failed host will not change, but remains active
790 and running. This gives the user false information about server state.
791 * VIM northbound interface notification of host faults towards VNFM and NFVO
792 should be in line with OpenStack state. This fault notification is a Telco
793 requirement defined in ETSI and will be implemented by OPNFV Doctor project.
794 * Openstack user cannot make HA actions fast and reliably by trusting server
795 state and host state.
799 There needs to be a new API for Admin to state host is down. This API is used to
800 mark services running in host down to reflect the real situation.
802 Example on compute node is:
804 * When compute node is up and running:::
806 vm_state: activeand power_state: running
807 nova-compute state: up status: enabled
809 * When compute node goes down and new API is called to state host is down:::
811 vm_state: stopped power_state: shutdown
812 nova-compute state: down status: enabled
816 There is no attractive alternative to detect all different host faults than to
817 have an external tool to detect different host faults. For this kind of tool to
818 exist there needs to be new API in Nova to report fault. Currently there must be
819 some kind of workarounds implemented as cannot trust or get the states from
820 OpenStack fast enough.
822 .. [*] https://blueprints.launchpad.net/nova/+spec/update-server-state-immediately
827 This section lists some BPs related to Doctor, but proposed by drafters outside
830 pacemaker-servicegroup-driver [*]_
831 __________________________________
833 This BP will detect and report host down quite fast to OpenStack. This however
834 might not work properly for example when management network has some problem and
835 host reported faulty while VM still running there. This might lead to launching
836 same VM instance twice causing problems. Also NB IF message needs fault reason
837 and for that the source needs to be a tool that detects different kind of faults
838 as Doctor will be doing. Also this BP might need enhancement to change server
839 and service states correctly.
841 .. [*] https://blueprints.launchpad.net/nova/+spec/pacemaker-servicegroup-driver
844 vim: set tabstop=4 expandtab textwidth=80: