How would you run chaos experiments on a live streaming pipeline without viewers ever noticing?

The 30-second answer

I treat chaos as hypothesis testing with a seatbelt. Each experiment states a viewer-facing steady state — stall ratio, manifest age, start time — and a falsifiable prediction, then climbs a ladder: staging with a looped real feed and synthetic viewers, then production on a test channel or cell that only synthetic viewers watch, using header-scoped faults so real requests are untouched, and only then small real-traffic experiments and game days. Abort conditions are written before the run and wired to automation that removes the fault in seconds. I never run on match days, during incidents, or with a low error budget, and every finding becomes a fix plus a regression experiment.