As soon as several nodes in a Group Replication accept write access at the same time, certification-based conflict detection decides which transaction wins and which one gets rolled back. Anyone who does not understand this mechanism ends up building applications that regularly fail under multi-primary operation with seemingly random rollbacks.