Technology1 min read

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

By · Published by Everything Blog

In short

The architecture relies on CascadeLSM time-based compaction, differentiated replicas sharing persistent files, and a hybrid columnar format to accelerate projected reads. To prepare inference inputs efficiently, the system utilizes partial dumps that merge versioned history with recently captured state. Furthermore, pushing aggregations down directly into storage allows the platform to achieve 8.5 times the throughput of ordinary reads under a constrained receive budget. Controlled evaluations against an optimized RocksDB baseline confirmed faster historical reads across different projection configurations and client loads.

Key points

  • Consolidating maintenance across ingestion, inference, and training workloads resolves op…: Consolidating maintenance across ingestion, inference, and training workloads resolves operational friction caused by asymmetric recommendation access patterns.
  • Tailored architectural mechanisms like CascadeLSM compaction and partial state dumps sign…: Tailored architectural mechanisms like CascadeLSM compaction and partial state dumps significantly accelerate historical data retrieval for demanding model pipelines.
  • Shared sequence maintenance maintains low write latencies while concurrently supporting h…: Shared sequence maintenance maintains low write latencies while concurrently supporting heavy batch access required for ongoing machine learning model training.

Sources checked by our analysis pipeline: 1

Computer Science > Databases

[Submitted on 2 Nov 2026]

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

View PDF HTML (experimental)Abstract:Recommendation models consume increasingly long, attribute-rich user histories, making small continuous updates and repeated large reads a central storage asymmetry. We present USDB, ByteDance's unified sequence storage practice for streaming and batch workloads. CascadeLSM time-based compaction stages materialize application-specific histories; differentiated replicas share persistent files; partial dumps reconstruct inference inputs from captured recent state and versioned history; and a hybrid columnar format accelerates projected reads. These designs reuse maintenance results across ingestion, inference, and training. Production migration records show 85.8% lower mean write latency, while normalized deployment accounting shows 42.1% lower infrastructure cost. Controlled comparisons with an Optimized RocksDB sequence baseline show faster long-history reads across projections and client loads. Aggregation pushdown sustains 8.5x the throughput of ordinary reads under a constrained receive budget. Production reconstruction and direct-file training results demonstrate how shared sequence maintenance supports fresh serving and repeated historical access.

Additional Features

References & Citations

Loading...

Bibliographic and Citation Tools

Bibliographic Explorer (What is the Explorer?)

Connected…

Technology

More in Technology