Container Orchestration in Limited Availability Domains

Mission-critical systems are of vital importance and their failure may result in catastrophic outcomes. Therefore, applying cloud-native technologies as container orchestration to support them is no trivial task. Even more so when the environment greatly differs: while usually cloud-native focuses on a geographically disperse scenario, servers hosted in distant data-centres, and thousands of clients, the mission-critical systems at hand are restricted to only 2 Availability Zones (AZ), exemplified by mission-critical systems in the naval domain. When AZs are limited, commonly used consensus protocols of cluster state store of container orchestrators fail to address the requirements of such vital systems: if one node fails, the cluster can never achieve consensus and no further changes can be made on its state. This thesis considers a solution to the problem, the use of a publish-subscribe database. A benchmark comprised of four metrics (latency, throughput, consistency, and partition tolerance) is described. In a prototype system, results are gathered, analysed and compared with proven cluster state stores (etcd and PostgreSQL). Final results indicate that a pub-sub data store should be capable of addressing the needs of cluster state stores for mission-critical systems, thereby opening a new avenue of research not thoroughly explored before.