A Hybrid Online-Offline Training Architecture for Maintaining Model Inference Freshness in Production Systems
DOI:
https://doi.org/10.63808/ihf.v2i3.472Keywords:
Inference freshness, Concept drift adaptation, Online-offline hybrid training, Streaming machine learning, MLOpsAbstract
Production drift erodes predictive performance between scheduled rebuilds. Periodic retraining incurs cadence-bound freshness lag, whereas continuous online learning is exposed to transient noise. We formalise inference freshness as observable sample lag plus a predictive-loss proxy and present a hybrid controller that coordinates online updates, offline rebuilds, and shadow-validated swaps. On four non-stationary INSECTS streams, the hybrid improves prequential accuracy by 1.1–5.3 percentage points over the stronger of periodic-offline and online-only. ARF remains 0.7–16.3 points more accurate but processes 455–610 predictions/s versus 6,231–9,820 for the hybrid in this CPU replay. A corrupted-rebuild test rejects all poor candidates and avoids a 20.63-point next-1,000-prediction regression. On synthetic low-, medium-, and high-transition streams, periodic retraining leads by 1.6, 2.9, and 0.5 points. The contribution is therefore a transparent freshness-aware scheduling and update-safety design, not state-of-the-art predictive performance.
References
[1] Arora, S., Rani, R., & Saxena, N. (2024). A systematic review on detection and adaptation of concept drift in streaming data using machine learning techniques. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 14(4), e1536. https://doi.org/10.1002/widm.1536
[2] Bifet, A., & Gavaldà, R. (2007). Learning from time-changing data with adaptive windowing. Proceedings of the 2007 SIAM International Conference on Data Mining, 443-448. https://doi.org/10.1137/1.9781611972771.42
[3] Chen, Y., Yang, X., & Dai, H. L. (2024). Cost-sensitive continuous ensemble kernel learning for imbalanced data streams with concept drift. Knowledge-Based Systems, 284, 111272. https://doi.org/10.1016/j.knosys.2023.111272
[4] Eken, B., Pallewatta, S., Tran, N., Tosun, A., & Babar, M. A. (2025). A multivocal review of MLOps practices, challenges and open issues. ACM Computing Surveys, 58(2), 1-35. https://doi.org/10.1145/3747346
[5] Gai, Y., Meng, K., & Wang, X. (2024). Online learning for streaming data classification in nonstationary environments. Statistical Analysis and Data Mining: The ASA Data Science Journal, 17(2), e11669. https://doi.org/10.1002/sam.11669
[6] Gomes, H. M., Bifet, A., Read, J., Barddal, J. P., Enembreck, F., Pfahringer, B., Holmes, G., & Abdessalem, T. (2017). Adaptive random forests for evolving data stream classification. Machine Learning, 106(9-10), 1469-1495. https://doi.org/10.1007/s10994-017-5642-8
[7] Palli, A. S., Jaafar, J., Md Saad, M. H., Mokhtar, A. A., Gomes, H. M., Soomro, A. A., & Gilal, A. R. (2025). Smart adaptive ensemble model for multiclass imbalanced nonstationary data streams. Scientific Reports, 15(1), 21140. https://doi.org/10.1038/s41598-025-05122-w
[8] Souza, V. M. A., Reis, D. M. D., Maletzke, A. G., & Batista, G. E. A. P. A. (2020). Challenges in benchmarking stream learning algorithms with real-world data. Data Mining and Knowledge Discovery, 34(6), 1805-1858. https://doi.org/10.1007/s10618-020-00698-5
[9] Steidl, M., Felderer, M., & Ramler, R. (2023). The pipeline for the continuous development of artificial intelligence models: Current state of research and practice. Journal of Systems and Software, 199, 111615. https://doi.org/10.1016/j.jss.2023.111615
[10] Suárez-Cetrulo, A. L., Quintana, D., & Cervantes, A. (2023). A survey on machine learning for recurring concept drifting data streams. Expert Systems with Applications, 213, 118934. https://doi.org/10.1016/j.eswa.2022.118934
[11] Wang, L., Zhang, X., Su, H., & Zhu, J. (2024). A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8), 5362-5383. https://doi.org/10.1109/TPAMI.2024.3367329
[12] Wooders, S., Mo, X., Narang, A., Lin, K., Stoica, I., Hellerstein, J. M., Crooks, N., & Gonzalez, J. E. (2023). RALF: Accuracy-aware scheduling for feature store maintenance. Proceedings of the VLDB Endowment, 17(3), 563-576. https://doi.org/10.14778/3632093.3632116
[13] Yang, S., Han, M., Zhu, S., Yang, W., Dai, Z., Li, J., & Ding, J. (2026). A survey of processing methods for different types of concept drift. Data & Knowledge Engineering, 161, 102525. https://doi.org/10.1016/j.datak.2025.102525
Downloads
Published
Issue
Section
License
Copyright (c) 2026 Zhengyuan Wang
This work is licensed under a Creative Commons Attribution 4.0 International License.