ByteDance USDB: Scaling Storage for Next-Generation Recommendation Systems

ByteDance proposes the unified sequence storage system USDB, which uses CascadeLSM to materialize user history in time-based tiers, shares persistent files across differentiated replicas, reconstructs inference inputs from partial dumps, and reuses maintenance results for writes, inference, and training. Production migration records show an 85.8% reduction in average write latency and a 42.1% reduction in normalized deployment cost; compared with an optimized RocksDB sequence baseline, long-history reads are faster, and aggregation pushdown achieves 8.5x the throughput of ordinary reads under a constrained receive budget.