Troubleshoot Telemetry ~~~~~~~~~~~~~~~~~~~~~~ Logging in Telemetry -------------------- The Telemetry service has similar log settings as the other OpenStack services. Multiple options are available to change the target of logging, the format of the log entries and the log levels. The log settings can be changed in :file:`ceilometer.conf`. The list of configuration options are listed in the logging configuration options table in the `Telemetry section `__ in the *OpenStack Configuration Reference*. By default ``stderr`` is used as standard output for the log messages. It can be changed to either a log file or syslog. The ``debug`` and ``verbose`` options are also set to false in the default settings, the default log levels of the corresponding modules can be found in the table referred above. | Recommended order of starting services -------------------------------------- As it can be seen in `Bug 1355809 `__, the wrong ordering of service startup can result in data loss. When the services are started for the first time or in line with the message queue service restart, it takes time while the **ceilometer-collector** service establishes the connection and joins or rejoins to the configured exchanges. Therefore, if the **ceilometer-agent-compute**, **ceilometer-agent-central**, and the **ceilometer-agent-notification** services are started before the **ceilometer-collector** service, the **ceilometer-collector** service may lose some messages while connecting to the message queue service. The possibility of this issue to happen is higher, when the polling interval is set to a relatively short period. In order to avoid this situation, the recommended order of service startup is to start or restart the **ceilometer-collector** service after the message queue. All the other Telemetry services should be started or restarted after and the **ceilometer-agent-compute** should be the last in the sequence, as this component emits metering messages in order to send the samples to the collector. | Notification agent ------------------ In the Icehouse release of OpenStack a new service was introduced to be responsible for consuming notifications that are coming from other OpenStack services. If the **ceilometer-agent-notification** service is not installed and started, samples originating from notifications will not be generated. In case of the lack of notification based samples, the state of this service and the log file of Telemetry should be checked first. For the list of meters that are originated from notifications, see the `Telemetry Measurements Reference `__. | Recommended ``auth_url`` to be used ----------------------------------- When using the Telemetry command line client, the credentials and the ``os_auth_url`` have to be set in order for the client to authenticate against OpenStack Identity. For further details about the credentials that have to be provided see the `Telemetry Python API `__. The service catalog provided by OpenStack Identity contains the URLs that are available for authentication. The URLs have different ``port``\s, based on whether the type of the given URL is ``public``, ``internal`` or ``admin``. OpenStack Identity is about to change API version from v2 to v3. The ``adminURL`` endpoint (which is available via the port: ``35357``) supports only the v3 version, while the other two supports both. The Telemetry command line client is not adapted to the v3 version of the OpenStack Identity API. If the ``adminURL`` is used as ``os_auth_url``, the :command:`ceilometer` command results in the following error message: :: $ ceilometer meter-list Unable to determine the Keystone version to authenticate with \ using the given auth_url: http://10.0.2.15:35357/v2.0 Therefore when specifying the ``os_auth_url`` parameter on the command line or by using environment variable, use the ``internalURL`` or ``publicURL``. For more details check the bug report `Bug 1351841 `__. .. TODO (karenb) The content in this file needs updating.