Proactive Elastic Scaling for Online Service in Cloud Environments with Minimal Training Data

Elasticity is a cornerstone of cloud computing. However, conventional learning-based scaling methods typically demand extensive historical metrics for training, which may not be feasible for newly-deployed online service or those with highly dynamic patterns. Enpowered by recent advances in time series foundation models, we introduce a novel proactive scaling method that works well even when none or few training data is available, leveraging foundation model’s zero-shot and few-shot learning capabilities. Experiments results demonstrates superior predicting accuracy with the practical inference time.