In this paper, we propose DHDA, an online configuration performance learning framework for modern composable software systems operating in dynamic environments. DHDA uses a dual-layered adaptation approach to adapt to both global drift (changes in performance across the entire configuration space) and local drift (changes that only affect specific sub-regions of the configuration space). At a high level, the data is repartitioned and local models are retrained within each partition to handle global drift only when necessary. At a low level, local models in each partition detect local drift and adapt asynchronously. To balance responsiveness and efficiency, DHDA combines incremental updates with periodic global retraining to minimize unnecessary computation when drift is not detected. Evaluations on eight software systems show that DHDA achieves significantly better accuracy than state-of-the-art techniques, effectively adapts to drift with up to a factor of 2 performance gains, and incurs reasonable overhead.
