Emitting Software Metrics¶
Beginning with the Newton (6.1.0) release, the ironic services support emitting internal performance data to statsd. This allows operators to graph and understand performance bottlenecks in their system.
This guide assumes you have a statsd server setup. For information on using and configuring statsd, please see the statsd README and documentation.
These performance measurements, herein referred to as “metrics”, can be emitted from the Bare Metal service, including ironic-api, ironic-conductor, and ironic-python-agent. By default, none of the services will emit metrics.
It is important to stress that not only statsd is supported for metrics collection and transmission. This is covered later on in our documentation.
Configuring the Bare Metal Service to Enable Metrics with Statsd¶
Enabling metrics in ironic-api and ironic-conductor¶
The ironic-api and ironic-conductor services can be configured to emit metrics
to statsd by adding the following to the ironic configuration file, usually
backend = statsd
If a statsd daemon is installed and configured on every host running an ironic service, listening on the default UDP port (8125), no further configuration is needed. If you are using a remote statsd server, you must also supply connection information in the ironic configuration file:
# Point this at your environments' statsd host
statsd_host = 192.0.2.1
statsd_port = 8125
Enabling metrics in ironic-python-agent¶
The ironic-python-agent process receives its configuration in the response from the initial lookup request to the ironic-api service. This means to configure ironic-python-agent to emit metrics, you must enable the agent metrics backend in your ironic configuration file on all ironic-conductor hosts:
agent_backend = statsd
In order to reliably emit metrics from the ironic-python-agent, you must provide a statsd server that is reachable from both the configured provisioning and cleaning networks. The agent statsd connection information is configured in the ironic configuration file as well:
# Point this at a statsd host reachable from the provisioning and cleaning nets
agent_statsd_host = 198.51.100.2
agent_statsd_port = 8125
Use of a different metrics backend with the agent is not presently supported.
Transmission to the Message Bus Notifier¶
Regardless if you’re using Ceilometer,
or some scripting you wrote to consume the message bus notifications,
metrics data can be sent to the message bus notifier from the timer methods
and additional gauge counters by utilizing the
configuration option and setting it to
collector. When this is the case,
Information is cached locally and periodically sent along with the general sensor
data update to the messaging notifier, which can consumed off of the message bus,
or via notifier plugin (such as is done with ironic-prometheus-exporter).
Transmission of timer data only works for the Conductor or
Ironic service model. A separate webserver process presently does not have
the capability of triggering the call to retrieve and transmit the data.
This functionality requires ironic-lib version 5.4.0 to be installed.
Types of Metrics Emitted¶
The Bare Metal service emits timing metrics for every API method, as well as for most driver methods. These metrics measure how long a given method takes to execute.
A deployer with metrics enabled should expect between 100 and 500 distinctly named data points to be emitted from the Bare Metal service. This will increase if the metrics.preserve_host option is set to true or if multiple drivers are used in the Bare Metal deployment. This estimate may be used to determine if a deployer needs to scale their metrics backend to handle the additional load before enabling metrics. To see which metrics have changed names or have been removed between releases, refer to the ironic release notes.
Additional conductor metrics in the form of counts will also be generated in limited locations where petinant to the activity of the conductor.
With the default statsd configuration, each timing metric may create additional metrics due to how statsd handles timing metrics. For more information, see statds documentation on metric types.
The ironic-python-agent ramdisk emits timing metrics for every API method.
Deployers who use custom HardwareManagers can emit custom metrics for their hardware. For more information on custom HardwareManagers, and emitting metrics from them, please see the ironic-python-agent documentation.
Adding New Metrics¶
If you’re a developer, and would like to add additional metrics to ironic, please see the ironic-lib developer documentation for details on how to use the metrics library. A release note should also be created each time a metric is changed or removed to alert deployers of the change.