Avito's homepage recommendations now update 18 times faster

Avito's recommendations team on Habr described the transition from batch updating the collaborative model avitofm every 6 hours to a near-time architecture with embedding fine-tuning every 20 minutes, which resulted in a +1% increase in buyer-seller contacts and a threefold reduction in memory usage.

Compiled by Sergey KostenchukPublished 2026-09-30Updated 2026-09-30

2026-09-30 Coding

💻 Avito's homepage recommendations are updated every 20 minutes

An analysis was published on the AvitoTech blog on Habr: the avitofm model was updated once every 6 hours, and new listings appeared in the search results with a delay. First, incremental fine-tuning was added (+6% Recall@K offline, +1% contacts in A/B), then a near-time architecture: one model per category instead of 4420 chunked ones, 20-minute clickstream windows from Kafka, A100/H100 GPUs fine-tune embeddings and output them to Redis.

🌍 Instead of a model for each category×region pair — one per category. Updates 18 times more often, p99 'event → recommendation' ≤ 25 minutes, memory usage is three times lower. The scheme is reproducible by recsys teams that are limited by the freshness of embeddings.

👤 Buyers see new listings almost immediately after publication — during peak buying hours. According to the team's measurements, this resulted in a +1% increase in contacts with sellers.

Source 1: https://habr.com/ru/companies/avito/articles/1086210/

Sources