The mechanism is selective capture: allow-listed features are serialized with compact integer aliases, then prediction records are joined to client impression events in Flink. Uber reports peak traffic of 8 million QPS, estimates that full logging could reach about 1.7 PB per day, and says its allow list reduces payload size by 4–5×. The immediate implication is that feature consistency becomes an operational data-contract problem: retain the serving values and the outcome together, while avoiding the cost of retaining every scored candidate.
Why now: transformer-backed recommendation traffic makes a canonical record desirable but economically infeasible without aggressive filtering and stream-state discipline. The evidence boundary is one company’s engineering description; it reports architecture and internal estimates, not an independently audited model-lift result. Watch whether Uber publishes measured changes in model quality, training freshness, or incident rates after deployment. This starts a distinct analytical record because the supplied material documents a specific inference-to-training logging architecture and its scaling trade-offs, while no prior New Runtime record is supplied for comparison.