On 5/9/19 9:24 AM, Tim Bell wrote:
Is it time to rethink the approach to telemetry a bit?
Having each project provide its telemetry data (such as Swift with statsd - https://docs.openstack.org/swift/latest/admin/objectstorage-monitoring.html
or using a framework like Prometheus)?
In the end, the projects are the ones who have the best knowledge of how to get the metrics.
Tim
Tim, statsd for swift is for monitoring, it is *not* a usage metric. Likewise with Prometheus, who wont care if some data are missing. I very much would love to have each project handle metrics collection by themselves. Especially, I always though that the polling system implemented in Ceilometer is just wrong, and that every service must be able to report itself rather than being polled. I understand however that doing polling is easier than implementing such change in every service, so I get why it has been done this way. But then we need some kind of timeseries framework within OpenStack as a whole (through an Oslo library?), and also we must decide on a backend. Right now, the only serious thing we have is Gnocchi, since influxdb is gone through the open core model. Or do you have something else to suggest? Cheers, Thomas Goirand (zigo)