Conversion Dropped 12% After the Deploy
Decide what you would do from the brief alone, including whether you would change anything at all. Everything below it is available, but the exercise stops working if you open it first.
Checkout conversion fell about 12% (illustrative) within an hour of a scheduled model deploy that changed both the pricing model and the backend release that serves it. The model's offline evaluation was clean. The incident commander needs to know: roll back which thing, and how do we find out what happened?
Rolling back only the model, since the offline evaluation was clean and "it must be the model". The old model loads on all hosts, the hosts with the cached old schema now match it and recover, and the hosts that correctly loaded the new schema now feed the old model a vector with an extra column — the failure moves to the other subset of hosts, conversion recovers only partly, and the incident commander concludes the drop had two causes. The clean offline evaluation was read as evidence about serving, which it never was.
Read this even if you are confident. It is here rather than behind a button because it is the answer most teams actually ship, it passes review, and its cost arrives weeks later when the labels do.