These items mix the machine learning unit on new products. A good answer keeps the two worlds straight: offline work (feature history, training, evaluation) can be slow and thorough, while online work (feature lookups, inference, monitoring) must be fast and safe.
For each decision, say what breaks if you pick the other option: leaked features inflate offline scores, recomputed features drift from training, unbounded funnels blow latency budgets and missing monitors let quality fall quietly. The products and numbers are new, so work each one out rather than recalling an earlier answer.
Apply feature, serving, ranking and monitoring decisions to new products.