HDFS HA Solution

HDFS HA Background

In versions earlier than Hadoop 2.0.0, SPOF occurs in the HDFS cluster. Each cluster has only one NameNode. If the host where the NameNode is located is faulty, the HDFS cluster cannot be used unless the NameNode is restarted or started on another host. This affects the overall availability of HDFS in the following aspects:

1. In the case of an unplanned event such as host breakdown, the cluster would be unavailable until the NameNode is restarted.
2. Planned maintenance tasks, such as software and hardware upgrade, will cause the cluster stop working.

To solve the preceding problems, the HDFS HA solution enables a hot-swap NameNode backup for NameNodes in a cluster in automatic or manual (configurable) mode. When a machine fails (due to hardware failure), the active/standby NameNode switches over automatically in a short time. When the active NameNode needs to be maintained, the administrator can manually perform an active/standby NameNode switchover to ensure cluster availability during maintenance.

For details about HDFS automatic failover, see

http://hadoop.apache.org/docs/r3.1.1/hadoop-project-dist/hadoop-hdfs/HDFSHighAvailabilityWithQJM.html#Automatic_Failover

HDFS HA Implementation

Figure 1 Typical HA deployment

In a typical HA cluster (as shown in Figure 1), two NameNodes need to be configured on two independent servers, respectively. At any time point, one NameNode is in the active state, and the other NameNode is in the standby state. The active NameNode is responsible for all client operations in the cluster, while the standby NameNode maintains synchronization with the active node to provide fast switchover if necessary.

To keep the data synchronized with each other, both nodes communicate with a group of JournalNodes. When the active node modifies any file system's metadata, it will store the modification log to a majority of these JournalNodes. For example, if there are three JournalNodes, then the log will be saved on two of them at least. The standby node monitors changes of JournalNodes and synchronizes changes from the active node. Based on the modification log, the standby node applies the changes to the metadata of the local file system. Once a switchover occurs, the standby node can ensure its status is the same as that of the active node. This ensures that the metadata of the file system is synchronized between the active and standby nodes if the switchover is incurred by the failure of the active node.

To ensure fast switchover, the standby node needs to have the latest block information. Therefore, DataNodes send block information and heartbeat messages to two NameNodes at the same time.

It is vital for an HA cluster that only one of the NameNodes be active at any time. Otherwise, the namespace state would split into two parts, risking data loss or other incorrect results. To prevent the so-called "split-brain scenario", the JournalNodes will only ever allow a single NameNode to write data to it at a time. During switchover, the NameNode which is to become active will take over the role of writing data to JournalNodes. This effectively prevents the other NameNodes from being in the active state, allowing the new active node to safely proceed with switchover.

For more information about the HDFS HA solution, visit the following website:

http://hadoop.apache.org/docs/r3.1.1/hadoop-project-dist/hadoop-hdfs/HDFSHighAvailabilityWithQJM.html