Question
Your news-ranking model updates online: it continuously retrains on click feedback every few minutes. Over the last 12 hours, content diversity has collapsed — the feed now shows almost exclusively one narrow topic cluster, engagement per session is falling, and user complaints about 'the same stuff over and over' are rising. No deploy, no error, latency normal. Dashboards: the online-learning loop trains on its OWN served impressions' clicks; a popular breaking story 12 hours ago got heavy clicks, the model up-weighted that cluster, surfaced more of it, which got more clicks (because that's mostly what was shown), reinforcing itself; exploration/randomization in the ranker was effectively zero. How do you triage and respond?
Stop the bleeding first (mitigate), then form hypotheses from real signals. Separate root cause from symptom, communicate status as you go, and close with what prevents a repeat.