You can do whatever that scales. I see some tools advocating for generating metrics and then advocating for downsampling. It is very confusing and technically wrong in some cases such as latency analysis. Generating things from traces didn’t scale for our case. Each trace adds another 10–15% overhead to a single RPC. This is why we think it is good to provide the both options. As an OC maintainer, I have to say, OC is not opinionated. You are free to write an exporter wrapper and generate these metrics on the trace export path if you have no downsampling.