A method and a device for processing a failure in at least one distributed cluster, and a system, where the at least one distributed cluster includes a first distributed cluster. The first distributed cluster includes a first Master node, a first Slave node, a first reference node, and a first secondary node that serves as a backup of the first Master node. The first secondary node receives a heartbeat message that includes first indication information. The first secondary node determines, according to the first indication information, that the first reference node is disconnected from the first Master node. The first secondary node determines that the first secondary node is also disconnected from the first Master node when it is detected that a heartbeat message from the first Master node to the first secondary node is interrupted. The first secondary node determines the first Master node is faulty.