The aim is to reduce training-serving inconsistency caused by separate online and offline computation paths.
At reported peak prediction traffic of 8 million QPS, logging every prediction would create hundreds of billions of records per day, with an estimated 1.7 PB daily footprint. Uber therefore logs allow-listed model features, aliases verbose names to integer IDs, and joins prediction logs with client impression events in Flink so that training records retain predictions actually seen by users. The company says the allow list cuts payload size by 4–5×.