I’ve found myself expressing from time to time that there are different types or “levels” of metrics we can use in and around engineering, and that each has different requirements.
An incomplete set of examples that I feel are most relevant to engineers follows. Note that as we move “up” levels of abstraction here, the value of the metrics is increasingly about the conversation that “pushing on the numbers” can stimulate.
See Will Larson’s notes for more discussion and a broader taxonomy, mostly focused on “investment-level” metrics below.
1. Operations-level metrics
Emitted by our software services: the stuff we put on service dashboards, representing workloads and how they’re doing. May have alerting attached, but probably shouldn’t.
Precision is king, here: we want each metric to have a clear meaning and implications. “Muddiness” or subtlety is the enemy. A good example of the latter is Linux load average. Caveat operator.
2. Service-level metrics
Emitted by our services and the infrastructure around them (for example, synthetic probes). Often used as service level indicators to understand customer experience and user journeys; may also include adoption and capacity metrics. Often have alerting attached.
While we want to define these precisely - see for example the Art of SLOs workshop - it’s OK for there to be uncertainties, subtleties, and warnings attached. Often we have to choose proxies for customer experience, or it’s prohibitively expensive to measure some important part of a user journey.
So we’re fine with service-level metrics needing to have an explanation attached: some bullet points, or a few paragraphs.
3. Investment-level metrics
Gathered in lots of different ways, often across a broad swath of systems and services. Used to drive or inform a specific large-scale investment by a group or organization: a difficult migration, the health of some org-level process. As likely to be delivered by a more traditional BI system as by Datadog dashboards. Rarely have alerting attached.
Again, we want a reasonably clear definition of a metric like this. However, it can be even less precise than a service-level metric, and have more caveats and edge-cases attached. It’s almost always a proxy - often a “trailing edge” indicator - for some important effort we’re trying to drive across the business.
It’s OK for a metric like this to have a page or a whole document explaining it, as long as we agree it provides a useful “needle we can move” over time.