On 23/11/2020 09:30, Thomas Goirand wrote:
On 11/23/20 8:54 AM, Matthias Runge wrote:
Personally, I wouldn't check gnocchi on the scalable bullet. (I'm eager to hear your success stories, that would make things much easier).
Maybe you should read this: https://julien.danjou.info/gnocchi-4-performance/
Julien Danjou pretends Gnocchi is able to eat 100k record per second. I wouldn't bet on this, but that's probably enough for OpenStack.
Though the biggest bottleneck is probably the notification bus, not the time series. I've heard that switching to kafka may help, but I'm really not a fan of this "solution" (which IMO isn't one, as Kafka brings its own set of problems).
all of this should be discussed in the community and also decided on. If we pick gnocchi, then we should also buy in and help out. I know that friendly and fellow OpenStackers already contribute there, which doesn't mean that there could or should be more helping hands.
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.
Prometheus is already in Debian, and it's been there for the last 2 releases, so I'd be ok switching to it. However, you'd have to provide a migration path for those already using Gnocchi.
If there has to be a migration path, that's also up for discussion. We are not a service provider here, we are simply providing the software. And as for all software, a new release may also bring backwards incompatible changes.... Matthias