This section will introduce apmec monitoring framework and describes the various actions that a user can take when a specific event occurs.
Apmec monitoring framework provides the MEM operators and MEA vendors to write a pluggable driver that monitors the various status conditions of the MEA entities it deploys and manages.
A monitor driver for apmec is a python module which contains a class that inherits from “apmec.mem.monitor_drivers.abstract_driver.MEAMonitorAbstractDriver”. If the driver depends/imports more than one module, then create a new python package under apmec/mem/monitor_drivers folder. After this we have to mention our driver path in setup.cfg file in root directory.
For example:
apmec.apmec.monitor_drivers =
ping = apmec.mem.monitor_drivers.ping.ping:MEAMonitorPing
Following methods need to be overridden in the new driver:
def get_type(self)
def get_name(self)
def get_description(self)
def monitor_get_config(self, plugin, context, mea)
def monitor_url(self, plugin, context, mea)
def monitor_call(self, mea, kwargs)
As mentioned in above section, if the return value of monitor_call method is other than boolean value ‘True’, then we have to map those event to the corresponding action as described below.
For example:
vdu1:
monitoring_policy:
ping:
actions:
failure: respawn
In this example, we have an event called ‘failure’. So whenever monitor_call returns ‘failure’ apmec will respawn the MEA.
The available actions that a monitor driver can call when a particular event occurs.
In the vdus section, under vdu you can specify the monitors details with corresponding actions and parameters.The syntax for writing monitor policy is as follows:
vduN:
monitoring_policy:
<monitoring-driver-name>:
monitoring_params:
<param-name>: <param-value>
...
actions:
<event>: <action-name>
...
...
vdu1:
monitoring_policy:
ping:
actions:
failure: respawn
vdu2:
monitoring_policy:
http-ping:
monitoring_params:
port: 8080
url: ping.cgi
actions:
failure: respawn
acme_scaling_driver:
monitoring_params:
resource: cpu
threshold: 10000
actions:
max_foo_reached: scale_up
min_foo_reached: scale_down
Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.