Recording Feature Flag Decisions with Four Signals for Tenant Cost Allocation

The author argues that a feature flag's enabled boolean alone cannot explain how experiment costs were distributed across tenants. Their approach records four signals with each evaluation: the evaluation result, flag version, tenant cohort, and a stable operation ID. The application needs a fast decision, while SRE and finance teams need the tenant, rule revision, and downstream work tied to that decision, so the author keeps the former in the response and the latter in structured telemetry. The article advises against high-cardinality or sensitive identifiers such as raw emails or tokens as correlation keys, and recommends keeping tenant-level detail in event logs rather than metric labels. It also notes that queue retries should reuse the original decision via the job payload with idempotent consumers, so a single operation produces one accounting event.