Minimum-Exposure Observability: Task-Specific Diagnostic Utility Under Distributed-Trace Minimization

Abstract

Distributed traces support diagnosis but can expose sensitive identifiers, payload content, and business context, motivating telemetry minimization. Prior privacy-preserving telemetry work quantifies utility, but does not generally organize application-trace evaluation around whether the same minimized representation preserves different diagnostic tasks differently. We therefore ask how privacy-driven minimization affects task-specific diagnostic utility and whether different diagnostic tasks require different minimum telemetry representations. We evaluate five minimization treatments — suppression, pseudonymization, generalization, whole-trace sampling, and a composite treatment — using a deterministic five-service in-process simulator and six independently defined and independently validated diagnostic tasks. Under the composite treatment, which suppresses actor identity while generalizing payload and business context, cross-request incident-cohort attribution was indeterminate in all 500 confirmatory windows and five fault strata, while context-associated failure diagnosis remained fully preserved over the same windows and strata. Through a different minimization mechanism within the same confirmatory design, whole-trace sampling at p = 0.50 produced a statistically significant task-specific difference in operational affected-set-F1 loss between these tasks (D = −0.0396, 95% CI [−0.0632, −0.0161], p = 0.00110). These results support a task-dependent view of diagnostic sufficiency. We do not identify a globally minimal trace schema; rather, utility preservation must be evaluated with respect to the diagnostic capability being preserved.