AvitoTech reports that its collaborative model previously refreshed every six hours, while its new near-real-time path processes clickstream in 20-minute windows. It also reports 240 million active listings and says that recommendations on the home page drive roughly half of views and about 30% of seller contacts. The first incremental implementation updated only embeddings touched by fresh events; AvitoTech reports a 6% Recall@K gain and a 1% increase in buyer–seller contacts in an A/B test.
The mechanism is selective state mutation rather than repeated full retraining: a baseline model supplies initial embeddings, fresh event windows identify affected users and items, and workers update and publish only the relevant state. The production design adds the unglamorous but decisive pieces—windowing, queue scheduling, watermark handling, deduplication, retries, GPU allocation, and Redis publication. That explains why the system can pursue freshness without treating every 20-minute update as a new full model build.
Why now: in marketplaces, a listing’s highest engagement may occur in its first hours, so a six-hour lag can remove commercially valuable inventory from the recommendation opportunity. The implication is that teams should measure model freshness against the half-life of the underlying inventory, not only against offline ranking metrics. A second implication is that an incremental model can create hidden operational debt if it runs alongside the batch model; consolidation may matter as much as the initial quality lift.
The evidence boundary is important: the reported Recall@K, A/B result, memory reduction, and architecture details are AvitoTech’s own account, and the supplied source does not provide experiment design, confidence intervals, absolute contact volumes, or independent replication. What to watch is whether Avito reports how the 20-minute system performs across its category-by-region partitions, during queue backlogs or failures, and after changes in listing mix or user behavior.