← Back to feed

ByteDance has published the design of USDB, the unified storage backend that serves user behaviour sequences to its recommendation models across Douyin, TikTok and live streaming. The paper, posted to arXiv on 2 November, reports 85.8% lower mean write latency, 72.6% lower mean CPU utilisation and 42.1% lower normalised infrastructure cost in production.

The problem

Large recommendation models, unlike older feature-engineered ones, scale with the length of the user history they read. ByteDance says the sequence length its models consume rose 2,000-fold between 2023 and 2025, from 50 to 100,000 actions. Representative retained-history configurations reached 800,000 actions in September 2026. Feature columns per record grew from dozens to hundreds.

That creates what the authors call a storage asymmetry: small continuous writes on one side, repeated large reads on the other. The paper lists four failure points in the previous design:

• Aggregating sequences in memory with a read-before-write pattern turns each small write into an SSD-bound bottleneck once streaming volume exceeds memory.
• LSM-tree compaction is CPU-heavy while batch storage needs disk, and running both on the same nodes inflates cost.
• Out-of-order arrivals and delayed feedback make the sequence seen at inference diverge from the one used in training, and large recommendation models are sensitive to that skew.
• Row-oriented SST files force training jobs to deserialise whole records when they only need some feature columns.

The four mechanisms

• CascadeLSM. A time-tiered LSM-tree that holds streaming and batch data in one structure, with differentiated compaction by time tier. It removes the read-before-write step.
• Differentiated replicas. Read-write, read-heavy and read-only replicas over disaggregated storage and compute. They share persistent files and synchronise without repeating compaction. This is where the 42% cost reduction comes from.
• Versioned partial dumps. A sequence-number scheme records what inference saw, so earlier inference inputs can be rebuilt from a captured recent fragment plus reusable history, instead of dumping full sequences repeatedly.
• Hybrid columnar format. Adaptive encoding that supports column scans for training and row-level indexing for inference. In the codec evaluation it decompresses 4.6 to 5.5 times faster than ZSTD, with encoded sizes between LZ4 and ZSTD.

Benchmarks against RocksDB

A 374-run study against an optimised RocksDB sequence baseline found:

• Repeated 10K and 100K reads reached 1.79x and 1.75x the throughput.
• 100K client-read jobs finished 1.62 to 2.52 times faster.
• Aggregation pushdown sustained 8.5 times the throughput of ordinary reads under a constrained receive budget.

Caveats

USDB has been deployed since 2024, so the production figures come from migration records, and the cost figure is “normalised deployment accounting” rather than a billing comparison. The paper says the growth charts are representative internal estimates, and the baseline comparison is against ByteDance’s own tuned RocksDB setup. No absolute latency or dollar figures are given in the abstract.

The broader point is that long-history recommendation, not chat, is where ByteDance is pushing sequence length into the hundreds of thousands, and the cost is landing on databases rather than GPUs.