# Monolith: Real Time Recommendation System With Collisionless Embedding Table

**Authors:** Zhuoran Liu, Leqi Zou, Xuan Zou, Caihua Wang, Biao Zhang, Da Tang, Bolin Zhu, Yijie Zhu, Peng Wu, Ke Wang, Youlong Cheng  
**arXiv:** [2209.07663](https://arxiv.org/abs/2209.07663)  
**Date:** 2022-09-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2209.07663v2/

## Key points

- Having a scalable and real-time recommendation system is crucial in customer feedback-driven businesses
- Traditional deep learning frameworks like TensorFlow or PyTorch may fall short for recommendation scenarios due to being optimized for static parameters and dense computations
- Monolith is a novel approach tailored specifically for online training, addressing challenges of traditional frameworks by observing application workloads and production environments
- Key contributions of Monolith include collisionless embedding table creation, memory-saving optimizations, robust online training architecture with high fault-tolerance capabilities
- Monolith prioritizes real-time interactions over absolute reliability, successfully integrated into BytePlus Recommend product

## Summary

, , , , 

In the fast-paced world of customer feedback-driven businesses, having a scalable and real-time recommendation system is crucial. While widely used in production settings, traditional deep learning frameworks like TensorFlow or PyTorch often fall short when it comes to meeting the specific demands of recommendation scenarios. These frameworks are optimized for static parameters and dense computations, making them less effective for dynamic and sparse feature-based recommendations. Additionally, their separation of batch-training and serving stages hinders real-time interaction with customer feedback. To address these challenges, a novel approach has been developed - Monolith, a system tailored specifically for online training. This innovative design sets it apart from conventional recommendation systems by carefully observing application workloads and production environments. The key contributions of Monolith include the creation of a collisionless embedding table with memory-saving optimizations such as expirable embeddings and frequency filtering. Furthermore, the system offers a robust online training architecture with high fault-tolerance capabilities. A significant aspect highlighted by this research is the trade-off between system reliability and real-time learning efficiency. By prioritizing real-time interactions over absolute reliability, Monolith has proven its effectiveness in practice by successfully integrating into the BytePlus Recommend product. This groundbreaking approach marks a departure from traditional methodologies and opens up new possibilities for enhancing recommendation systems in time-critical business settings.