[renlo]

I think usage statistics tend to require more retention time to discover user behavior and understand how to optimize revenue. In the general case people probably won't care much whether their widget was running at X% CPU on Dec 5th 2019 but they might care more about what percent of users did Y action on that date. When I worked on an observability team (not as an expert but as a general swe) we had two metrics pipelines; one was strictly usage statistics which came from the client, the other was purely server metrics but a subset of them were considered usage metrics which were aggregated and sent down with the client metrics for the folks upstairs.