This research paper details Monolith, a real-time recommendation system developed by Bytedance. Monolith addresses challenges in building scalable recommendation systems, such as sparse and dynamic data, and concept drift, by employing a collisionless embedding table and an online training architecture. Key innovations include a Cuckoo HashMap for efficient sparse parameter representation, incorporating features like expirable embeddings and frequency filtering, and a system for real-time parameter synchronization between training and serving. The authors present experimental results demonstrating Monolith's superior performance compared to systems using traditional hash tables and batch training, showcasing the benefits of its design choices in terms of model accuracy and efficiency. Finally, the paper compares Monolith to existing solutions, highlighting its unique advantages for industrial-scale applications.

https://arxiv.org/pdf/2209.07663