Alice. Alice Is Impatient
6 hours ago
- #Inspection Paradox
- #Tail Latency
- #Customer Experience
- The inspection paradox explains why users perceive mean wait times or outage durations differently from service metrics, as users weight longer events more heavily in seconds/minutes.
- Users experience a t-weighted version of the latency or outage distribution, calculated as ℕ[X] + Var(X)/ℕ[X], which often results in significantly higher perceived mean times compared to the service's reported mean.
- A simulation with log-normal distributions (used for convenience, though not ideal) demonstrates that even with moderate median and high p99 values, customer-experienced mean times can be much longer than service metrics indicate.
- Tail latency and long recovery times are critical because they dominate user experience, and techniques like timeout-and-retry may hide some latency but not for recovery events.
- Trimmed measurements (e.g., trimmed means) are discouraged as they remove important context about the right tail, which affects customer experience and capacity planning via Little's Law.