Modern compute infrastructure exposes hundreds of thousands of metrics for platform engineers to understand the health and performance of the systems they are responsible for. For a variety of reasons, these metrics are collected and stored in a timeseries database, even if the metric itself is not actually timeseries data. This choice is generally the correct one if you are trying to monitor your infrastructure; however, if you are trying to _understand_ your infrastructure, timeseries data can sometimes hinder or obscure your results.

In this talk, we present ten types of graphs you can’t make in Grafana, and show how they were nevertheless helpful for identifying cost regressions, performance improvements, and reliability bottlenecks in real systems. We also discuss approaches for building a data pipeline off for automating this type of analysis based off traditional timeseries observability tooling.