StaleFlow: Staleness-Aware Data Management for Mitigating Data Skewness in Fully Disaggregated RL Post-Training

Summary: StaleFlow jointly manages trajectory staleness and length-induced skew in asynchronous, fully disaggregated RL post-training, combining lifecycle consistency tracking with trajectory/parameter servers and staleness-aware scheduling. It boosts throughput up to 2.68× without sacrificing convergence. (summarized by gpt-6-luna on Oct 22 2026)

Paper ID: h332d557fee00d4a7
Venue: SIGMOD
Year: 2027
Pagerank: 4.9605343e-05
Overall Rank: 10,410 | 30.23%
DOI: 10.1145/3837125
PDF: Download (CC BY 4.0)

Incoming Non-self Citations Over Time

No non-self incoming citations found for this paper in this database.

Authors

1. Haoyang Li (Peking University)
2. Sheng Lin (Peking University)
3. Fangcheng Fu (Shanghai Jiao Tong University)
4. Yuming Zhou (Peking University)
5. Xiaodong Ji (Peking University)
6. Yanfeng Zhao (Shanghai Jiao Tong University)
7. Lefeng Wang (Shanghai Jiao Tong University)
8. Jie Jiang (Tencent)
9. Bin Cui (Peking University)

BibTeX Citation

@inproceedings{li_sigmod27,
        title = {{StaleFlow: Staleness-Aware Data Management for Mitigating Data Skewness in Fully Disaggregated RL Post-Training}},
        author = {Li, Haoyang and Lin, Sheng and Fu, Fangcheng and Zhou, Yuming and Ji, Xiaodong and Zhao, Yanfeng and Wang, Lefeng and Jiang, Jie and Cui, Bin},
        series = {{SIGMOD} '27},
        booktitle = {Proceedings of the {ACM} {SIGMOD} International Conference on Management of Data},
        publisher = {Association for Computing Machinery},
        doi = {10.1145/3837125},
        url = {https://dl.acm.org/doi/10.1145/3837125},
        year = {2027}
}

Incoming Citations (Sorted by Pagerank)

Showing 0 of 0 citing papers.

Rank	Citing Paper	Year	Venue	Pagerank

Outgoing Citations (Sorted by Pagerank)

Showing 8 of 8 cited papers.

Citations counted here include only citations to other VLDB/SIGMOD/CIDR/PODS papers in this database.

Rank	Cited Paper	Year	Venue	Pagerank
1,612	PyTorch FSDP: Experiences on Scaling Fully Sharded Data Parallel	2023	VLDB	0.00010061487
3,844	HotPrefix: Hotness-Aware KV Cache Scheduling for Efficient Prefix Sharing in LLM Inference Systems	2026	SIGMOD	6.9683696e-05
3,972	PQCache: Product Quantization-based KVCache for Long Context LLM Inference	2025	SIGMOD	6.8744954e-05
4,910	Apt-Serve: Adaptive Request Scheduling on Hybrid Cache for Scalable LLM Inference Serving	2025	SIGMOD	6.3468554e-05
5,011	Galvatron: Efficient Transformer Training over Multiple GPUs Using Automatic Parallelism	2023	VLDB	6.3030066e-05
10,650	Hydraulis: Balancing Large Transformer Model Training via Co-designing Parallel Strategies and Data Assignment	2026	SIGMOD	4.9605343e-05
11,257	Malleus: Straggler-Resilient Hybrid Parallel Training of Large-scale Models via Malleable Data and Model Parallelization	2025	SIGMOD	4.9605343e-05
11,346	LobRA: Multi-tenant Fine-tuning over Heterogeneous Data	2025	VLDB	4.9605343e-05

Semantically Similar Papers

#	Overall Rank	Paper	Year	Venue
1	11,257	Malleus: Straggler-Resilient Hybrid Parallel Training of Large-scale Models via Malleable Data and Model Parallelization	2025	SIGMOD
2	9,232	Dalton: Learned Partitioning for Distributed Data Streams	2023	VLDB
3	10,800	Scarf: Self-Adaptive Tuning via Multi-Objective Reinforcement Learning for Apache Flink	2026	VLDB
4	10,573	ReStore: A Reinforcement Learning Approach for Data Migration in Multi-Tiered Storage	2026	SIGMOD
5	2,196	Heterogeneity-aware Distributed Parameter Servers	2017	SIGMOD
6	9,352	Stochastic Data Acquisition for Answering Queries as Time Goes by	2017	VLDB
7	4,927	Scalable Asynchronous Gradient Descent Optimization for Out-of-Core Models	2017	VLDB
8	1,472	Analyzing and Mitigating Data Stalls in DNN Training	2021	VLDB
9	7,333	Model-Free Control for Distributed Stream Data Processing using Deep Reinforcement Learning	2018	VLDB
10	4,961	Heterogeneity-Aware Distributed Machine Learning Training via Partial Reduce	2021	SIGMOD