USDB: Scaling Storage for Next-Generation Recommender Systems at ByteDance
AI summary
Editorial policyByteDance presents USDB, a unified sequence storage practice designed for next-generation recommender systems to handle streaming and batch workloads efficiently.
Key points
- •The increasing length and attribute richness of user histories consumed by recommendation models create a central storage asymmetry.
- •CascadeLSM time-based compaction stages are used to materialize application-specific histories.
- •The design employs differentiated replicas to share persistent files and partial dumps to reconstruct inference inputs.
#Semiconductors & Infrastructure#LLM / Model Trends#ByteDance#USDB