DECK: Experiences on Delta Checkpointing for Industrial Recommendation Systems

Summary: DECK introduces a production-ready delta checkpointing system for multi‑terabyte industrial recommender training that extracts and streams model-state deltas with near-zero overhead and without halting training. Decoupled optimal merging of streamed deltas yields ~12× checkpoint frequency with negligible throughput loss. (summarized by gpt-5-mini on Feb 09 2026)

Paper ID: 14288
Venue: VLDB
Year: 2025
Pagerank: -
Overall Rank: 13,327 | 8.57%
DOI: 10.14778/3750601.3750621

Incoming Non-self Citations Over Time

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

Authors

1. Xin Gao (Meta Inc.)
2. Sibasish Acharya (Meta Inc.)
3. Sihui Han (Meta Inc.)
4. Yongxiong Ren (Meta Inc.)
5. Yanli Zhao (Meta Inc.)
6. Liang Luo (Meta Inc.)
7. Chucheng Wang (Meta Inc.)
8. Pradeep Fernando (Meta Inc.)
9. Saurabh Mishra (Meta Inc.)
10. Siqi Yan (Meta Inc.)
11. Yicong Du (Meta Inc.)
12. Elzbieta Krepska (Meta Inc.)
13. Intaik Park (Meta Inc.)
14. Min Ni (Meta Inc.)
15. Qunshu Zhang (Meta Inc.)
16. Shen Li (Meta Inc.)

BibTeX Citation

@article{gao_vldb25,
        title = {{DECK: Experiences on Delta Checkpointing for Industrial Recommendation Systems}},
        author = {Gao, Xin and Acharya, Sibasish and Han, Sihui and Ren, Yongxiong and Zhao, Yanli and Luo, Liang and Wang, Chucheng and Fernando, Pradeep and Mishra, Saurabh and Yan, Siqi and Du, Yicong and Krepska, Elzbieta and Park, Intaik and Ni, Min and Zhang, Qunshu and Li, Shen},
        journal = {PVLDB},
        series = {{VLDB} '25},
        volume = {18},
        number = {12},
        pages = {4978--4990},
        doi = {10.14778/3750601.3750621},
        url = {https://doi.org/10.14778/3750601.3750621},
        year = {2025}
}

Incoming Citations (Sorted by Pagerank)

Showing 0 of 0 citing papers.

Rank	Citing Paper	Year	Venue	Pagerank

Outgoing Citations (Sorted by Pagerank)

Showing 2 of 2 cited papers.

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

Rank	Cited Paper	Year	Venue	Pagerank
521	PyTorch Distributed: Experiences on Accelerating Data Parallel Training	2020	VLDB	0.0001713368
2,473	PyTorch FSDP: Experiences on Scaling Fully Sharded Data Parallel	2023	VLDB	8.5326287e-05

Semantically Similar Papers

#	Overall Rank	Paper	Year	Venue
1	9,520	Experimental Analysis of Large-scale Learnable Vector Storage Compression	2024	VLDB
2	2,233	FINEdex: A Fine-grained Learned Index Scheme for Scalable and Concurrent Memory Systems	2022	VLDB
3	6,708	Serving Deep Learning Models with Deduplication from Relational Databases	2022	VLDB
4	8,908	Privacy and Accuracy-Aware AI/ML Model Deduplication	2025	SIGMOD
5	6,954	DLRover-RM: Resource Optimization for Deep Recommendation Models Training in the Cloud	2024	VLDB
6	8,907	Scheduling Data Processing Pipelines for Incremental Training on MLP-based Recommendation Models	2025	SIGMOD
7	9,164	FEC: Efficient Deep Recommendation Model Training with Flexible Embedding Communication	2023	SIGMOD
8	7,300	DEX: Query Execution in a Delta-based Storage System	2017	SIGMOD
9	6,958	Efficient Fault Tolerance for Recommendation Model Training via Erasure Coding	2023	VLDB
10	10,804	IncrCP: Decomposing and Orchestrating Incremental Checkpoints for Effective Recommendation Model Training	2025	VLDB