84% of CI failures are flaky, not bugs

I keep hearing "flaky tests are a problem" but I wanted actual numbers, not vibes.

Some study at Google found that around 84% of pass-to-fail transitions in their CI are flaky. Not real bugs. Just noise. Atlassian documented 150k+ dev hours per year wasted on flaky reruns in a single repo. Buildkite measured 800 million seconds of flaky re-runs per month across their platform. 25 years of compute time. Per month. Just re-runs...

We all know the pattern. You stop trusting CI, re-run the pipeline, merge anyway. Eventually we're paying CI costs for tests nobody believes lol

One team deleted half their flaky e2e tests and production bugs went down slightly. Features per sprint up 56%. Nubank deleted their entire e2e suite and deploys went from ~100/week to ~1000/week.

At what flake rate do you quarantine vs just delete? Has anyone actually recovered a flaky suite rather than nuking it?