Scalable network overlay to support direct communications within Kubernetes multi-cluster topologies

The growth of the Internet knows no rest, and brings with it the necessity to deliver services at unprecedented orders of magnitude. The cloud has become the foundation upon which services can be delivered and managed at scale, along with engineering applications to be cloud-native. In this context, Kubernetes has been the driving force behind the successful management, deployment and delivery of large-scale services. With Kubernetes, computing nodes are brought together into clusters, where organizations have full visibility over the overall resources at their disposal, for an effective utilization of the deployed infrastructure. Moreover, companies may need to control multiple clusters to ensure their presence in multiple regions, increase availability, scale better in terms of maintenance, and may rely on different cloud providers for cost management or vendor lock-in avoidance. To take full advantage of a multi-cluster infrastructure, clusters may be connected together to create a single yet flexible environment where the deployment of heavy workloads can leverage the aggregate resources that clusters provide together. Liqo, an open-source project started at Politecnico di Torino, embraces this idea and makes it a reality. Liqo makes it possible to deploy an application consisting of several micro-services to different clusters. To preserve the overall functioning of the application, the different application’s components have to communicate with one another, from one cluster to the other. This was achieved by creating a hub-and-spoke topology, where traffic passed through a central hub cluster. This thesis presents the design and implementation of a solution whereby workloads deployed in a multi-cluster scenario can directly exchange traffic. With this approach, clusters connected to a central cluster are able to discover one another and exchange the network parameters required for direct communications, using the central cluster as a relay. Once this initial setup completes, traffic starts flowing directly, without traversing the central cluster, which is freed from networking overhead and overloads the previous topology implied, reducing the latency in communications and making them immune to temporary network outages on the central cluster. This results in a mesh, peer-to-peer topology and was made possible by evolving the networking and service reflection logic at the core of Liqo.