Predictive Bayesian Arbitration: A Scalable Noisy-OR Model with Service Criticality Awareness

TL;DR

This paper introduces a scalable, predictive arbitration framework for Geo-HA clusters using a Bayesian Noisy-OR model that learns failure dependencies and enables proactive switchovers, reducing downtime and detection time.

Contribution

It presents a novel Bayesian Noisy-OR based online learning system with expert priors for predictive failure detection in distributed cloud environments.

Findings

Achieves 60% reduction in failure detection time

Improves switchover efficiency by up to 77.8%

Enables proactive switchovers before failures occur

Abstract

Geographically High-Available (Geo-HA) cluster systems are essential for service continuity in distributed cloud-native environments. However, traditional arbitration mechanisms, which are often predicated on deterministic node-level heartbeats, are resource-intensive and inherently reactive. This necessitates a dedicated arbiter per deployment and leads to reactive switchovers that incur unavoidable downtime, occurring only after a failure has already compromised the system. This paper presents a novel predictive arbitration framework that utilizes a shared, microservice-based architecture to consolidate arbitration logic across multiple Geo-HA domains, significantly reducing the aggregate infrastructure footprint. Central to our approach is an adaptive online learning mechanism grounded in a Bayesian Noisy-OR model that autonomously discovers and learns temporal cascade dependencies from emergent failure patterns. To overcome the "cold start" challenge, the system utilizes expert-informed priors that are dynamically refined at runtime without manual configuration. Experimental results demonstrate that this framework achieves a 60\% reduction in Mean Time to Failure Detection (MTTFD) and improves total switchover efficiency by up to 77.8\% compared to traditional reactive standards. By enabling a significant predictive lead time, the system allows switchovers to initiate proactively before hard failures occur, while maintaining a linear $O(n)$ computational complexity. This approach provides a scalable, context-aware alternative that bridges the performance-durability gap in modern microservice architectures.