The Quality
Flywheel
How failures become experiments, better decisions, and knowledge that survives the next release.
The tests pass. A change goes live. Websites stop responding.
A quality flywheel begins with the gap between those three sentences. What did the tests miss? Which change would prevent a repeat? How do we know the proposed fix will help somewhere beyond the original failure?
Preface
Our starting point is Cloudflare's outage on 2 July 2019. A new security rule consumed so much processing time that websites returned errors. The rule had passed its tests; those tests did not cover the way it could exhaust CPU.
The outage and Cloudflare's account are real. Atlas, the on-call AI assistant we use to explore the lessons, is hypothetical. Its examples ask a practical question: how would we teach an assistant to investigate a slow or unavailable website, then verify that it had learned something useful?
The chapters follow that question through measurement, investigation, hill climbing, release, and the knowledge left behind.