HDFS HA architecture and principles

- JN station with 2N + 1 storage EditLog, each write data operation has a majority (> = N + 1) times to return it that the write is successful, the data will not be lost if successful. Of course, this algorithm can be tolerated that up to N machines hang, to hang up if more than N stations, the algorithm becomes ineffective. This principle is based on Paxos algorithm. HA architecture in which SecondaryNameNode this cold standby role does not exist, in order to maintain standby NN always consistent with the metadata master Active NN, interactive JournalNode lightweight processes through a series of guard between them. When modifying any operation is performed on Active NN, JN process will also modify the log to record more than half of JN least in this time Standby NN monitored inside JN change the synchronization log reads JN inside the modified log, and then sync to your directory tree of images which, when a fault occurs, Active hang after the NN, Standby NN will it become before Active NN, JN read all of which change log so that we can guarantee high reliability and NN hang a mirror directory tree is consistent, then seamlessly take over its duties, maintenance requests from clients, so as to achieve a high-availability purposes