ViperProbe: Rethinking Microservice Observability with eBPF

Abstract

Recent shifts to microservice-based architectures and the supporting servicemesh radically disrupt the landscape of performance-oriented management tasks. While the adoption of frameworks like Istio and Kubernetes ease the management and organization of such systems, they do not themselves provide strong observability. Microservice observability requires diverse, highly specialized, and often adaptive, metrics and algorithms to monitor both the health of individual services and the larger application. However, modern metrics collection frameworks are relatively static and rigid. We introduce ViperProbe, an eBPF-based microservices collection framework that provides (1) dynamic sampling and (2) collection of deep, diverse, and precise system metrics. Viper-Probe builds on the observation that the adoption of a common set of design patterns, e.g., servicemesh, enables offline analysis. By examining the performance profile of these patterns before deploying on production, ViperProbe can effectively reduce the set of collected metrics, thereby improving the efficiency and effectiveness of those metrics. To the best of our knowledge, ViperProbe is the first scalable eBPF-based dynamic and adaptive microservices metrics collection framework. Our results show ViperProbe has limited overhead, while significantly more effective for traditional management tasks, e.g., horizontal autoscaling.