Teaching the Old Dog a New Trick

Vinted's engineering team solved tail latency spikes (100ms+) in their Vespa search infrastructure by eliminating summary fetching during data redistribution. By encoding required data as match-features and leveraging a new optimization in Vespa 8.596.7, they reduced p99 latencies from ~9ms to ~3ms and eliminated latency spikes that occurred during high-volume feeding bursts (500M+ updates/hour). The solution removes a network round-trip and avoids document summary fetching failures when data moves between content nodes.