On 21/11/2020 02:38, Thomas Goirand wrote:
On 11/12/20 11:25 AM, Matthias Runge wrote:
Hi there,
one of the biggest challenges for the telemetry stack is currently the state of gnocchi, which is... undefined/unfortunate/under-contributed/...?
Telemetry started long time ago as a simple component ceilometer, which was split into several components ceilometer, aodh, panko, and gnocchi. Julien wrote a story about this some time ago[1].
There has also been an attempt to fork gnocchi back to OpenStack[2].
To my knowledge, the original contributors are not paid anymore to work on gnocchi, and at the same time, moving on to do something else is totally fine.
However, I am not sure if we (in OpenStack Telemetry) should or could maintain in addition to the rest of the telemetry stack a time-series database. I'd like to discuss this during a call.
Matthias,
I'm not sure I will have time to join the call. So hopefully, you understand that I prefer email (also because I wont be able to contribute to the project, so maybe joining the call would be overkill). Could you list the alternatives? If not using Gnocchi, what other backend would you use? As an operator I need a timeseries which is: - free software - HA - able to scale - packaged or packagable in my distro
Thomas, thank you for your email. Let me go into more details: since this requires some discussion, I am proposing to have a call instead of an email thread. The current stack uses gnocchi; you may be well aware of the back and forth around it. Also, I am not sure how much effort we can invest, or if that's feasible. I totally agree, we'd want a free software (not some open core or so) and also a scalable solution (you may have heard some load issues with old ceilometer or gnocchi before). Another solution may be to split use cases and to use gnocchi with an in-memory store for short lived telemetry data and a different backend for long time store, for example for billing data etc. If we stop providing the gnocchi API (or don't bring the old ceilometer API back), we'll also cut off applications currently using that API. All that requires some coordination. Does that explain my preference for a call?
I don't know any time series db (apart from Gnocchi) that check all the bullets above. Do you?
Personally, I wouldn't check gnocchi on the scalable bullet. (I'm eager to hear your success stories, that would make things much easier). Or let me rephrase this: it seems to be easier to achieve ingesting much more metrics per time interval e.g by using prometheus and at the same time using less hardware resources. Matthias
On 11/20/20 3:48 PM, Matthias Runge wrote:
If it works for you, let's use this bluejeans channel[4] for the discussion.
Gosh, reading that you'd be using bluejeans, then definitively, I will not be able to join the call... :/
Cheers,
Thomas Goirand (zigo)