A newly deployed model is underperforming. What do you do?

Roll back to the previous known-good version immediately if user impact is real — recovery first, diagnosis second. Then investigate with your monitoring: compare input distributions (drift?), check for a broken feature pipeline (training-serving skew), review the deploy diff, and inspect failing cases. Reproduce offline, fix, re-run evals, and redeploy progressively (shadow/canary). Capture the root cause so the same class of failure is caught earlier next time. Good rollback and observability are what make this a controlled event rather than a crisis.