USDB: Scaling Storage for Next-Generation Recommender Systems at ByteDance

ByteDance researchers detailed USDB, a unified sequence storage system for streaming and batch workloads in next-generation recommender systems, in an arXiv paper (2611.02618) authored by Xiaoran Zhu, Jieming Cai, Weiping Yu and colleagues. USDB combines CascadeLSM time-based compaction, differentiated replicas sharing persistent files, partial dumps that reconstruct inference inputs from recent state and versioned history, and a hybrid columnar format for projected reads. Production migration records show 85.8% lower mean write latency and 42.1% lower infrastructure cost under normalized deployment accounting, while aggregation pushdown sustained 8.5x the throughput of ordinary reads under a constrained receive budget. Controlled comparisons against an Optimized RocksDB sequence baseline showed faster long-history reads across projections and client loads. The work matters because recommendation models increasingly consume long, attribute-rich user histories, creating a storage asymmetry between small continuous updates and repeated large reads that ByteDance says USDB reuses across ingestion, inference and training. Next to watch is whether these production results translate beyond ByteDance's internal deployment and how the approach compares with established storage engines at scale.