Abstract

Many applications span multiple cloud providers. The faults in one cloud provider can travel to other providers, which are not visible in the monitoring consoles of those providers. Threshold alerts and anomaly detection systems can only monitor each metric individually and therefore are often alerted too late to diagnose the problem quickly. This article reviews the application of graph machine learning techniques, specifically graph neural networks (GNNs), to incident prediction and root cause analysis and extends these techniques to multi-cloud environments. The metrics from each cloud provider are combined into a graph structure and GNNs learn from that graph structure. The results of the analysis can be explained by returning the subgraph of the graph that is most responsible for the results of the analysis. These results can be applied throughout the incident lifecycle. Also, metrics for evaluating the success of these systems are specified. This work is a proposal for such a system and has not been validated through new experiments.

Keywords

How to Cite This Article

Rights