Feed and Recommendation Architecture at VKontakte
summary
Architecture and engineering of VKontakte’s feed and recommendation systems, showing how the feed serves as the primary interaction point for 100 million users and must deliver relevant content quickly across devices and network conditions. Covers technologies and system design used to generate, rank and serve personalized content at scale (feed generation, recommendation pipelines, low-latency serving and infrastructure considerations).
AI-written summary · about
chapters (21)
- 0:00Introduction
- 1:15Scale and challenges
- 3:04Evolution of recommendations
- 5:12Building chronological feed
- 8:24Popular authors problem
- 10:34Purpose of ranking
- 12:23Filtering viewed posts
- 15:59Feed optimization
- 17:48Selecting 500 relevant posts
- 18:09Feature extraction for ranking
- 21:23Decaying counters in ML
- 23:11ML model code generation
- 24:59Ranking system optimization
- 26:27Optimization results
- 27:11Author recommendations
- 30:43Recommendation infrastructure
- 33:12Candidate search for recommendations
- 35:44Ranking recommendations
- 37:55Caching for speed
- 40:45Final feed assembly
- 41:32Conclusion and Q&A
related talks
41mVK Media Recommendations: How We Build the Personalized "For You" Feed
43mArchitecture of Yandex Market’s infinite personal feed
12mTechTalk #4: How search works on VKontakte
41mThe Infinite Feed — No Jokes
39mRecommendations at XING: From theory to production
34mThe Billion-Request Content Recommendation System Challenge
32mFeed generation at ShareChat
47mHow We Made VKontakte's Mobile Version Four Times Faster