The Problem with MTTR: Learning from Incident Reports
summary
Argues that MTTR is a misleading single number: a curated database of roughly 10,000 reports from about 600 organizations across 15 years shows incident durations form skewed distributions where the mean is dominated by outliers, and Monte Carlo tests show mean-based changes often aren’t statistically meaningful. Recommends tracking human-centered signals instead — cost of coordination (who was paged, how many teams/tools involved), logging and counting near misses, using SLOs and richer post-incident reviews, and paying attention to language and mental models.
AI-written summary · about
chapters (9)
- 0:00Intro
- 1:29The end of MTTR?
- 4:24Library of incidents
- 12:48What is an incident?
- 19:15Cost of coordination
- 21:47Near misses
- 23:55Mental models
- 27:49Role of language in shaping public discourse
- 29:13Learnings from The Void
related talks
27mTales from the VOID: The Scary Truth about Incident Metrics
37mIncident Management Metrics That Matter
38mResilience in Modern Software Systems
45mHow To Measure MTTx Values in SRE
24mThe Metrics Question: Evolving Past MTTX
41mAre We Getting Better Yet? Progress Toward Safer Operations
29mShrinking the Impact of Production Incidents
40mFar from the Shallows: The Value of Deeper Incident Analysis