Marc Brooker: why customers perceive worse latency and recovery than the metrics

Category: Engineering post
Status: discussion
Sources: brooker.co.za, discussion
Summary: Marc Brooker's post, dated 2026-06-19, applies the inspection paradox to service latency and outage recovery: a customer who is waiting is disproportionately likely to be inside a long request or a long outage, so the experience is a time-weighted distribution rather than the mean, expressed as E_a[X] = E[X] + Var(X)/E[X]. He works an example where a 30-minute median and 10-hour p99 recovery yield a one-hour mean time to recovery but a roughly six-hour customer-perceived recovery time, and notes that timeout-and-retry can hide latency some of the time but offers no such mitigation during recovery. He argues against trimming measurements and for understanding the shape of the tail distribution, with an interactive log-normal fitting tool in the post.
Why it matters: It gives a concrete reason that tail variance, not the mean, drives perceived latency and recovery, which informs how SLOs and MTTR are measured and communicated.