This is an area I’m particularly interested in. Some tools advocate a sample everything and generate metrics at query time approach. Others like StackDriver and OpenCensus seem to be going this way; having separate solutions for 1. low cardinality pre-aggregated metrics, 2. Sampled end to end traces, and 3. Code profiling. I’d love to see StackDriver be more explicit about that approach (if this is correct) and explain the pros and cons.