Self-Adaptive Service Selection for Machine Learning Continuous Delivery

In the dynamic landscape of machine learning applications on streaming data, the constant evolution of models and input data complicates optimal model deployment. The static selection of a model risks suboptimal performance as data patterns evolve, while frequent redeployments increase operational costs. This paper proposes a self-adaptive system that autonomously selects interchangeable models for processing streaming data while balancing the tradeoff of performance and redeployment frequency. Inspired by the MAPE-K reference model, our approach utilizes an adaptive model selection control loop to continuously monitor model performance on production and experimental data. "what-if" environments are introduced to collect additional experimental data, simulating production-like scenarios. A selection algorithm that employs two distinct adaptation policies is introduced that strategically plans the selection of the most suitable module for upcoming data. Leveraging a learning-based method, we improve the efficiency of our system by recognizing the patterns of selection eliminating the need for further experimental data collection. Empirical evaluation on an energy forecasting use case spans over 16 years of data demonstrates a substantial reduction in errors up to 34% compared to the best static selection, affirming the proposed framework’s effectiveness. Our findings reveal the potential to discontinue experimental "what-if" analyses with just 12% of historical data, which underlines the practicality of our adaptive strategy on a long-lasting task.