架构图
- Yagi 从rabbitmq拿到 notifications并传递给Yagi Handlers 链。
- Yagi Handlers 包括：
-
- Shoebox for long-term archiving.
- Atom-Hopper for pub-sub ATOM feeds.
- Stack-distiller/Winchester for StackTach.v3 stream processing.
Winchester 来蒸馏notification，变成event，再利用Streams 转换成新的event或者notification。
类似于Ceilometer event处理，由event handler把notification转换成Ceilometer event并存储起来。
最后Notabene 可以把Winchester出来的notification重新发布到message bus上做进一步的处理，这个过程可以反复。
这个对应到Ceilometer上就是把event transformer出来的event转换成notification在重新发布到message bus上，可以进一步，比如把它处理成meter，
Ceilometer正在实现这个功能,BP在这儿：event to sample publisher https://review.openstack.org/#/c/223926/
Yagi配置：
[global] verbose = False debug = True update_timer = 10 [event_worker] event_driver = yagi.broker.rabbit.Broker [rabbit_broker] host = localhost user = guest password = guest port = 5672 vhost = / poll_delay = 0.5 [consumers] queues = monitor.info [consumer:monitor.info] apps = winchester.yagi_handler.WinchesterHandler, yagi.handler.shoebox_handler.ShoeboxHandler exchange = monitor exchange_type = topic routing_key = monitor.info durable = True max_messages = 100
从Notifications 到 Events的提取，由winchester.yagi_handler:WinchesterHandler来做这部分工作。
把Notifications中有用的field提取为event的 Traits。在配置文件中定义。
例子：
- event_type: compute.instance.*
traits:
tenant_id:
fields:
- payload.tenant_id
- _context_project_id
user_id:
fields: payload.user_id
request_id:
fields: _context_request_id
instance_id:
fields:
- payload.instance_uuid
- payload.instance_id
- exception.kwargs.uuid
- instance.uuid
这点和ceilometer很类似，ceilometer的配置文件event_definitions中定义了每个event需要提前哪些traits。
接下来就是在streams中处理event，streams就是事件分发到的bucket（Streams are buckets that collect events）。
采用类似于ceilometer中的event pipeline机制。把event分发到pipeline，pipeline中最后部分是handler，
Winchester 配置文件：
---
database:
url: mysql://winchester:testpasswd@localhost/winchester
distiller_config: event_definitions.yaml
trigger_definitions: triggers.yaml
pipeline_config: pipelines.yaml
pipeline_handlers:
logger: winchester.pipeline_handler:LoggingHandler
usage: winchester.pipeline_handler:UsageHandler
notabene: winchester.pipeline_handler:NotabeneHandler
其中：
trigger_definitions: triggers.yaml 中决定
- into which streams to place incoming events
- which events to accept into a stream
- when a stream is ready for processing
- when a stream expires
- how to process ready streams
- how to process expired streams
- name: my_trigger distinguished_by: - request_id expiration: "$last + 1h" fire_pipeline: "my_fire_pipeline" expire_pipeline: "my_expire_pipeline" match_criteria: - event_type: - compute.instance.* fire_criteria: - event_type: compute.instance.*.end
上面 distinguished_by:request_id，也就是不同的request发到不同的stream bucket中。
match_criteria定义这个pipelinez需要处理的事件，类似ceilometer event_pipeline,yaml中的source - events配置。
fire_criteria: 定义了何时触发stream bucket中的event的处理。比如最后一次收到事件之后的2小时，或者基于特定事件如收到compute.instance.*.end。
最后如何处理stream bucket中的event定义在fire_pipeline: "my_fire_pipeline"中。可以有多个pipeline handlers， 类似于ceilometer event_pipeline,yaml中
的sinks中的publisher。
fire_pipeline的配置：
---
my_fire_pipeline:
- logger
- usage
- name: notabene
params:
host: localhost
user: guest
password: guest
exchange_type: topic
queue_name: monitor.info
env_keys:
- usage_notifications
my_expire_pipeline:
- logger
- usage
再来看看pipeline中的handler：
UsageHandler is a pipeline handler for determining the daily usage of every instance with an OpenStack Nova deployment
NotabeneHandler will take any new notifications (not events)publish them to the RabbitMQ exchange specified.
总结与Ceilometer的区别：
stacktach只用来处理notification
ceilometer可以处理notification和meter还有alarm
都可以把notification转换成event来保存和处理，stacktach的处理更丰富和到位
ceilometer主要还是生成数据和meter
stacktach中的pipeline不能单个或者2个event做transformer。
stacktach有处理触发机制，一次对bucket event处理，效率更高。
https://github.com/openstack/stacktach
http://stacktach.readthedocs.org/