Culturing Resiliency with Data: A Taxonomy of Outages
summary
Uber analyzes roughly a thousand major tech incidents and finds 65–85% change-management, 15–25% capacity, ~2–5% internal-platform distribution; each incident is tagged with a golden signal (throughput/utilization/latency) and a time-series topology so anomaly detection uses time-series golden signals plus STL decomposition and change‑point detection. Incremental deploys, automated rollbacks, weekly failover game days and CI-driven runbook exercises feed back into process changes and fuller annotation rates from 2017→2018.
AI-written summary · about
chapters (8)
- 0:00Culturing resiliency with data A Taxonomy of incidents
- 0:3401 Background & context 02 Categories 03 Key learnings
- 1:22What is an Incident?
- 7:44Change Induced incidents
- 10:17Capacity Induced incidents
- 12:06Platform Induced incidents
- 16:23Restorative justice processes
- 18:22Resiliency through improved SDLC
related talks
38mResilience in Modern Software Systems
29mBuilding Resilient Systems
24mThe Metrics Question: Evolving Past MTTX
34mToward continuous resilience
19mHow to handle incidents at scale
20mData Incidents - When things go wrong
15mMaintaining reliable systems
44mProduction Incident Evaluation as Cost of Quality