On the feasibility of multi-leader replication in the early tiers

In traditional service architectures that follow the service statelessness principle, the state is primarily held in the data tier. Here, service operators utilize tailored storage solutions to guarantee the required availability; even though failures can occur at any time. This centralized approach to store and process an application’s state in the data tier implies that outages of the entire tier cannot be tolerated. An alternative approach, which is in focus of this thesis, is to decentralize the processing of state information and to use more stateful components in the early tiers. The possibility to tolerate a temporary outage of an entire tier implies that the application’s state can be manipulated by the remaining tiers without waiting for approval from the unavailable tier. This setup requires multi-leader replication, where every replica can accept writes and forwards the resulting changes to the other replicas. This thesis explores the feasibility of using multi-leader replication to store and process state in a decentralized manner across multiple tiers. To this end, two replication mechanisms, namely Conflict-free Replicated Data Types and Operational Transformation, are under particular investigation. We use and extend both mechanism to demonstrate that the aforementioned decentralization is worth considering when designing a service architecture. The challenges that arise when following our approach go back to fundamental impossibility results in distributed systems research, i. e. the impossibility to achieve a fault-tolerant consensus mechanism in asynchronous systems and the inevitable trade-off between availability and consistency in the presence of failures. With this thesis, we contribute to close the exposed gaps of both results by providing usable alternatives for standard IT services. We exemplify the feasibility of our alternatives with a fully distributed IMAP service and a programming library that provides the necessary extension to utilize our approach in a variety of web-based applications. All contributions of this thesis are based on both theory and practice. In particular, all extensions to the existing multi-leader replication mechanisms were proven to satisfy the necessary properties. Moreover, those extensions were also implemented as prototypical applications and evaluated against the corresponding de facto standard software from the industry.