Evaluating Failover and Recovery of Replicated SQL Databases

Abstract: Kubernetes has emerged as a popular container orchestration platform, providing elasticity for applications by enabling them to be rapidly provisioned, scaled and terminated. This works well for stateless applications, however stateful applications and in particular replicated databases which many applications rely on require careful handling to keep a consistent state in a dynamic environment. The purpose of this thesis is to investigate the current state of replication in MySQL-compatible databases and give insight to how reliable they are to run on Kubernetes. In thesis an exploration is done to find different database alternatives and select a few based on their capabilities and longevity. The selected alternatives are then subject to experimentation with different failover and recovery scenarios following chaos engineering principles to evaluate their reliability. MySQL, Vitess, and TiDB were all selected and of the three MySQL InnoDB Cluster shows the best availability properties, TiDB second with improved scalability, and Vitess third with possibilities for global scales.