exploring documents, and supporting newer authors.
Previously, candidate selection, filtering, and ranking were shared, with documents separated only near the end of the process. VK says this made configuration conditional and difficult to inspect, made changes harder to isolate, and caused experiments for different objectives to compete for shared slots.
A request now splits into multiple streams after user-data collection. Each stream has its own task and configuration and can independently select candidates, filter, and rank them. Their results are still combined into one output through quota-based blending. Streams reuse common pipeline components rather than maintaining separate codebases, and the approach has been deployed in VK Video and made available to other products through VK’s platform.
The source does not provide latency, relevance, revenue, experiment-throughput, or reliability measurements. It also does not establish how widely other products have adopted the design.