Cost and capacity as signals

Which cost and capacity signals reveal a fault, when totals and utilization hide it?

Two quantities report themselves only as totals that hide the fault.

Cost. Who pays for what, and how to split a shared cluster between teams, is cost attribution and unit economics. For finding a fault, the signal that matters is a rate, not a total. Total spend grows with the business and tells you nothing. Unit costs hold still: $ per TB ingested, $ per pipeline run, $ per 1,000 queries. Their derivative (the change over time) is the only cost-regression detector that works. If $ per TB ingested rises 30% in a week while the pipeline is unchanged, something regressed.

Capacity. Utilization is a lagging, misleading signal: response time rises sharply well before 100%, so a system at 85% CPU with a latency objective has already lost its budget, and how much headroom a scaling delay requires is cost mechanics' arithmetic. Watch a leading signal, such as queue depth or the age of the oldest waiting request, which moves before the latency does.