Liberty Global's digital TV platform is a complex graph-based IT system that needs to be constantly monitored to prevent downtime. Anomaly detection and graph analytics support engineers with timely problem identification. We show how we evolve graph deep learning methods such as node embedding and message passing from unsupervised to semi-supervised as root causes are gradually being labeled.
With its digital TV platform, Liberty Global serves video content to over 10 million customers in six countries on a daily basis. In order to provide a seamless viewing experience, the video streaming platform is constantly monitored to detect and resolve issues before they result in downtime.
A large volume and variety of telemetry data streams is collected, describing each of the densely interconnected microservices in detail. As monitoring them all manually is impossible, Liberty Global developed its own AIOps solution. Using anomaly detection, real-time correlations and graph-based root cause analysis, IT operations engineers are provided with a prioritization of incidents along with suggested root causes.
The biggest challenge in identifying graph-based root causes is the lack of proper labels. This talk will take you through our approach to jump-starting a learning system using unsupervised techniques such as node embeddings and Google PageRank. Active root cause labeling from operations engineers subsequently allows us to transform the problem into a semi-supervised one. Techniques such as graph convolutions, node embeddings, supervised PageRank propagation can then really unlock the power of deep learning for graph analytics. This is an exciting field that is relatively new and unstandardized but growing quickly.
Besides taking you through the interesting challenges in monitoring large scale IT systems with AI, it serves as a primer to practical graph (deep) learning approaches. With example code on public data made available on github, you'll immediately be able to get started with graph deep learning yourself.