Your CI pipeline is flaky and the team has started retrying red builds without reading them. How do you fix that?

Treat flakiness as a production incident, because a pipeline nobody trusts provides no safety at all. Measure it first: track pass rate and quarantine tests that fail intermittently so main goes green again. Then fix the quarantined tests by cause, usually shared state, timing assumptions, or a real race condition. Enforce a policy that quarantine is temporary with an owner and a deadline, so it does not become a graveyard.

The dangerous part is not the wasted time, it is that people have learned that red does not mean broken, so the day a real regression lands it gets retried through as well. I start by measuring, because opinions about which tests are flaky are usually wrong. Rerun the suite on an unchanged commit repeatedly, or track per test pass rates over a few hundred runs, and you get a ranked list. Then quarantine the worst offenders out of the blocking path so main is green and red means something again, but quarantine with an owner and an expiry date, otherwise it becomes a graveyard nobody looks at. Fixing them, the causes cluster: tests sharing state through a database or a singleton, sleeps standing in for proper waits, and reliance on ordering. And some of them are real bugs. I once chased a test that failed one time in fifty and it turned out to be a genuine race in a connection pool, which would have been an intermittent production issue nobody could reproduce.