Biathlon: Harnessing Model Resilience for Accelerating ML Inference Pipelines

Summary: Biathlon, an ML-serving system, exploits model resilience to input perturbations by selecting per-aggregation-feature approximation levels to maximize latency reduction while guaranteeing bounded end-to-end accuracy loss. Evaluated on real pipelines, it achieves 5.3x–16.6x speedups with negligible accuracy drop. (summarized by gpt-5-mini on Feb 09 2026)

Paper ID: h545ab1ee6c8e6be8
Venue: VLDB
Year: 2024
Pagerank: 5.4111208e-05
Overall Rank: 7,983 | 46.35%
DOI: 10.14778/3675034.3675052
PDF: Download (CC BY-NC-ND 4.0)

Incoming Non-self Citations Over Time

Authors

1. Chaokun Chang (Chinese University of Hong Kong)
2. Eric Lo (Chinese University of Hong Kong)
3. Chunxiao Ye (Chinese University of Hong Kong)

BibTeX Citation

@article{chang_vldb24,
        title = {{Biathlon: Harnessing Model Resilience for Accelerating ML Inference Pipelines}},
        author = {Chang, Chaokun and Lo, Eric and Ye, Chunxiao},
        journal = {PVLDB},
        series = {{VLDB} '24},
        volume = {17},
        number = {10},
        pages = {2631--2640},
        doi = {10.14778/3675034.3675052},
        url = {https://doi.org/10.14778/3675034.3675052},
        year = {2024}
}

Incoming Citations (Sorted by Pagerank)

Showing 3 of 3 citing papers.

Rank	Citing Paper	Year	Venue	Pagerank
6,168	Apt-Serve: Adaptive Request Scheduling on Hybrid Cache for Scalable LLM Inference Serving	2025	SIGMOD	5.8603375e-05
11,078	KEN: An Execution Engine for Unstructured Database Systems	2026	VLDB	4.9769913e-05
11,198	Malleus: Straggler-Resilient Hybrid Parallel Training of Large-scale Models via Malleable Data and Model Parallelization	2025	SIGMOD	4.9769913e-05

Outgoing Citations (Sorted by Pagerank)

Showing 27 of 27 cited papers.

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

Rank	Cited Paper	Year	Venue	Pagerank
9	Online Aggregation	1997	SIGMOD	0.00076265429
271	NoScope: Optimizing Neural Network Queries over Video at Scale	2017	VLDB	0.0002256866
318	DeepDB: Learn from Data, not from Queries!	2020	VLDB	0.00021166957
596	Wander Join: Online Aggregation via Random Walks	2016	SIGMOD	0.00015782051
772	VerdictDB: Universalizing Approximate Query Processing	2018	SIGMOD	0.0001409096
841	Quickr: Lazily Approximating Complex AdHoc Queries in BigData Clusters	2016	SIGMOD	0.00013543
1,428	Knowing When You’re Wrong: Building Fast and Reliable Approximate Query Processing Systems	2014	SIGMOD	0.0001069161
1,828	DBEst: Revisiting Approximate Query Processing Engines with Machine Learning Models	2019	SIGMOD	9.547768e-05
1,869	G-OLA: Generalized On-Line Aggregation for Interactive Analysis on Big Data	2015	SIGMOD	9.4749419e-05
2,003	Sample + Seek: Approximating Aggregates with Distribution Precision Guarantee	2016	SIGMOD	9.2071735e-05
2,028	Database Learning: Toward a Database that Becomes Smarter Every Time	2017	SIGMOD	9.1584244e-05
2,184	Extending Relational Query Processing with ML Inference	2020	CIDR	8.8953085e-05
2,251	Evaluating End-to-End Optimization for Data Analytics Applications in Weld	2018	VLDB	8.750953e-05
2,460	Query Processing on Tensor Computation Runtimes	2022	VLDB	8.4308406e-05
2,661	End-to-end Optimization of Machine Learning Prediction Queries	2022	SIGMOD	8.1568473e-05
2,672	Relational Confidence Bounds Are Easy With The Bootstrap*	2005	SIGMOD	8.1456115e-05
3,213	Turbo-Charging Estimate Convergence in DBO	2009	VLDB	7.5304969e-05
3,401	A Demonstration of Willump: A Statistically-Aware End-to-end Optimizer for Machine Learning Inference	2020	VLDB	7.3294462e-05
3,508	Serving and Optimizing Machine Learning Workflows on Heterogeneous Infrastructures	2023	VLDB	7.2455881e-05
4,182	Rafiki: Machine Learning as an Analytics Service System	2019	VLDB	6.7519862e-05
6,038	Incremental Computation of Common Windowed Holistic Aggregates	2016	VLDB	5.9059794e-05
6,239	Optimizing In-memory Database Engine for AI-powered On-line Decision Augmentation Using Persistent Memory	2021	VLDB	5.8381685e-05
6,418	Containerized Execution of UDFs: An Experimental Evaluation	2022	VLDB	5.7899052e-05
6,552	JoinBoost: Grow Trees Over Normalized Data Using Only SQL	2023	VLDB	5.7475822e-05
9,189	FEBench: A Benchmark for Real-Time Relational Data Feature Extraction	2023	VLDB	5.2097224e-05
10,113	Everest: A Top-K Deep Video Analytics System	2022	SIGMOD	5.0765311e-05
10,127	RALF: Accuracy-Aware Scheduling for Feature Store Maintenance	2024	VLDB	5.0736184e-05

Semantically Similar Papers

#	Overall Rank	Paper	Year	Venue
1	10,005	cedar: Optimized and Unified Machine Learning Input Data Pipelines	2025	VLDB
2	6,168	Apt-Serve: Adaptive Request Scheduling on Hybrid Cache for Scalable LLM Inference Serving	2025	SIGMOD
3	976	Democratizing Data Science through Interactive Curation of ML Pipelines	2019	SIGMOD
4	9,171	Optimizing Inference Serving on Serverless Platforms	2022	VLDB
5	6,747	Mitigating the Impedance Mismatch between Prediction Query Execution and Database Engine	2025	SIGMOD
6	9,426	Towards Observability for Production Machine Learning Pipelines	2022	VLDB
7	5,774	Optimizing Data Pipelines for Machine Learning in Feature Stores	2023	VLDB
8	6,222	Materialization and Reuse Optimizations for Production Data Science Pipelines	2022	SIGMOD
9	5,271	InferDB: In-Database Machine Learning Inference Using Indexes	2024	VLDB
10	7,544	Automating and Optimizing Data-Centric What-If Analyses on Native Machine Learning Pipelines	2023	SIGMOD