This documentation gives you some clues on how to write a new agent or plugin for Ceilometer if you wish to instrument a measurement which has not yet been covered by an existing plugin.
Polling agent might be run either on central cloud management nodes or on the compute nodes (where direct hypervisor polling is quite logical).
The agent running on each compute node polls for compute resources usage. Each meter collected is tagged with the resource ID (such as an instance) and the owner, including tenant and user IDs. The meters are then reported to the collector via the message bus. More detailed information follows.
The agent running on the cloud central management node polls other types of resources from a management server (usually using OpenStack services API to collect this data).
The polling agent is implemented in ceilometer/agent/manager.py. As you will see in the manager, the agent loads all plugins defined in the namespace ceilometer.poll.agent, then periodically calls their get_samples() method.
A polling agent can support multiple plugins to retrieve different information and send them to the collector. As stated above, an agent will automatically activate all possible plugins if no additional information about what to poll was passed. Previously we had separated compute and central agents with different namespaces with plugins (pollsters) defined within. Currently we keep separated namespaces - ceilometer.poll.compute and ceilometer.poll.central for quick separation of what to poll depending on where is polling agent running. This will load, among others, the ceilometer.compute.pollsters.cpu.CPUPollster, which is defined in the folder ceilometer/compute/pollsters.
Notifications mechanism uses plugins as well, for instance ceilometer.telemetry.notifications.TelemetryApiPost plugin which is defined in the ceilometer/telemetry/notifications folder, Though in most cases, this is not needed. A meter definition can be directly added to ceilometer/meter/data/meter.yaml to match the event type. For more information, see the New measurements page.
We are using these two existing plugins as examples as the first one provides an example of how to interact when you need to retrieve information from an external system (pollster) and the second one is an example of how to forward an existing event notification on the standard OpenStack queue to ceilometer.
Compute plugins are defined as subclasses of the ceilometer.compute.pollsters.BaseComputePollster class as defined in the ceilometer/compute/pollsters/__init__.py file. Pollsters must implement one method: get_samples(self, manager, cache, resources), which returns a sequence of Sample objects as defined in the ceilometer/sample.py file.
In the CPUPollster plugin, the get_samples method is implemented as a loop which, for each instances running on the local host, retrieves the cpu_time from the hypervisor and sends back two Sample objects. The first one, named cpu, is of type “cumulative”, meaning that between two polls, its value is not reset while the instance remains active, or in other words that the CPU value is always provided as a duration that continuously increases since the creation of the instance. The second one, named cpu_util, is of type “gauge”, meaning that its value is the percentage of cpu utilization.
Note that the LOG method is only used as a debugging tool and does not participate in the actual metering activity.
There is the way to specify either namespace(s) with pollsters or just list of concrete pollsters to use, or even both of these parameters on the polling agent start via CLI parameter:
ceilometer-polling --polling-namespaces central compute
This command will basically make polling agent to load all plugins from the central and compute namespaces and poll everything it can. If you need to load only some of the pollsters, you can use pollster-list option:
ceilometer-polling --pollster-list image image.size storage.*
If both of these options are passed, the polling agent will load only those pollsters specified in the pollster list, that can be loaded from the selected namespaces.
Agents coordination cannot be used in case of pollster-list option usage. This allows to avoid both samples duplication and their lost.
This should only be needed for cases where a complex arithmetic or non-primitive data types are used. In most cases, adding a meter definition to the ceilometer/meter/data/meter.yaml should suffice.
Notifications are defined as subclass of the ceilometer.agent.plugin_base.NotificationBase meta class. Notifications must implement:
- A sequence of strings defining the event types to be given to the plugin
- process_notification(self, message)
- Receives an event message from the list provided to event_types and returns a sequence of Sample objects as defined in the ceilometer/sample.py file.
In the InstanceNotifications plugin, it listens to three events:
Using the get_event_type method and subsequently the method process_notification will be invoked each time such events are happening which generates the appropriate sample objects to be sent to the collector.
Although we have described a list of the meters Ceilometer should collect, we cannot predict all of the ways deployers will want to measure the resources their customers use. This means that Ceilometer needs to be easy to extend and configure so it can be tuned for each installation. A plugin system based on setuptools entry points makes it easy to add new monitors in the agents. In particular, Ceilometer now uses Stevedore, and you should put your entry point definitions in the entry_points.txt file of your Ceilometer egg.
Installing a plugin automatically activates it the next time the ceilometer daemon starts. Rather than running and reporting errors or simply consuming cycles for no-ops, plugins may disable themselves at runtime based on configuration settings defined by other components (for example, the plugin for polling libvirt does not run if it sees that the system is configured using some other virtualization tool). Additionally, if no valid resources can be discovered the plugin will be disabled.
Any new plugin or agent contribution will only be accepted into the project if provided together with unit tests. Those are defined for the compute agent plugins in the directory tests/compute and for the agent itself in test/agent. Unit tests are run in a continuous integration process for each commit made to the project, thus ensuring as best as possible that a given patch has no side effect to the rest of the project.