In Prometheus and similar monitoring systems, recording rules precompute expensive or repeated queries. They are useful because raw queries over histograms, labels, or long time windows can be slow and hard to read. The rule turns that query into a new time series with a simpler name. The risk is that the new series starts to feel like truth even when it hides detail.
A concrete example is a rule that records service-level p95 latency by endpoint group. It makes dashboards fast, but a new route with low traffic and high latency may disappear inside the aggregate. Operators see green panels while a subset of users gets slow responses. The failure is not that recording rules are bad. The failure is losing the path back to raw metrics when the precomputed answer no longer matches the production question.
Good recording rules have owners, comments, tests or review, and dashboard escape hatches. They should reduce cost without removing the evidence needed during an incident.