A Reliable Extrapolative Capacity Estimation Method for Complex Distributed Web Services

Reliable estimation of system capacity in large-scale distributed web services is crucial for cost-performance optimization and auto-scaling. Existing methods, such as simulation and load testing, face challenges due to system complexity, incomplete observability, and high costs for modern distributed systems. We propose a novel extrapolative capacity estimation method, incorporating a queueing model-inspired parameterization, and an informative Bayesian prior for robust extrapolation. Furthermore, we devised a novel self-consistency filter that can also act as a grey-box model reliability measure using the principle of single value. This enables function-oriented belief reliability design for extrapolative learning algorithms without ground truth. Our method demonstrates reliability and robustness across noisy system behaviors and load variations, validated on a large-scale distributed system.