Enable HDFS NameNode High Availability with HDP 2.0.6 plugin
Extend HDP 2.0.6 plugin to include the setup and configuration of the HDFS NameNode High Availability after creating, configuring and starting the cluster.
Hadoop clusters are created with a single NameNode which represents a SPOF (Single Point Of Failure). If the NameNode becomes unavailable, the whole cluster becomes unavailable and we have to wait for the NameNode to come back up before using the cluster again. NameNode High Availability was introduced in Hadoop 2.0.0 and integrated into HDP as of the 2.0.0 release. The High Availability is achieved by having 2 NameNodes, an active NameNode and a stand by one. When the active fails the standby steps in and act as the cluster's NameNode. NameNode's High Availability can be configured manually on clusters deployed with HDP 2.0.6 plugin in Sahara through Ambari, but the process is long, tedious and error prone.
End users might not have the necessary skills required to setup the High Availability and deployers might prefer to deploy highly available clusters without manually configuring each one.
HDFS NameNode High Availability (Using Quorum Journal Manager (QJM)) uses Journal nodes to share HDFS edits between the active and the standby namenodes. The journal nodes ensure that the two namenodes have the same set of HDFS edits, but do not ensure the automatic failover. The automatic failover requires Zookeeper servers and Zookeeper Failover Controllers (ZKFC). A typical cluster with HDFS NameNode High Availability uses at least three (or a an odd number greater than three) journal nodes and a zookeeper cluster with three or five zookeeper servers. The Zookeeper Failover Controllers are installed on the servers acting as active and standby namenodes. The setup removes the secondary namenode (which is usually installed on a different server than the one hosting the namenode) and installs a second namenode process. The journal nodes and zookeeper servers can be installed on the same servers running the active and standby (old secondary namenode) namenodes. This leaves us with 2 journal nodes and 2 zookeeper servers. An additional server is all what's needed with journal node and zookeeper server to setup a minimally viable Hadoop cluster with HDFS NameNode High Availability. (For more info : http://hadoop.apache.org/docs/r2.3.0/hadoop-yarn/hadoop-yarn-site/HDFSHighAvailabilityWithQJM.html)
The 'Create Node Group Template' wizard will introduce a new process 'JOURNALNODE' to the list of available processes for HDP 2.0.6 plugin. Other processes necessary for HDFS HA are either already included in the list (NAMENODE, ZOOKEEPER_SERVER and SECONDARY_NAMENODE) or will be automatically setup (Zookeeper Failover Controllers).
The 'Launch Cluster' wizard for HDP 2.0.6 plugin will include a checkbox 'Enable HDFS HA'. This option will default to False and will be added to the cluster object.
The verification code will verify the necessary requirements for a Hadoop cluster and a set of additional requirements in the case where 'Enable HDFS HA' is set to True. The requirements include : NAMENODE and SECONDARY_NAMENODE on different servers At least three journal nodes on different servers At least three zookeeper servers on different servers
Upon successful validation the cluster will be created and once it's in the 'Active' state Availability' and if 'Enable HDFS HA' is True the service will instruct the plugin to start configuring the NameNode High Availability. The cluster will be set in 'Configuring HDFS HA' state and the plugin will start the configuration procedure. The procedure starts by the plugin stopping all the services and executing some preparation commands on the server with the namenode process (through the hadoopserver objects). Then the plugin installs and starts the journal nodes using Ambari REST API (POST, PUT, WAIT ASYNC). Next the configuration is updated using Ambari REST API (PUT), other services including Hive, Oozie and Hue might require configuration update if they are installed. Finally some more remote commands are executed on the namenodes and the Zookeeper Failover Controllers are installed and started and the SECONDARY_NAMENODE process is deleted. The plugin will return and the cluster will be set back in the 'Active' state.
Manual setup through Ambari web interface
Data model impact
REST API impact
Other end user impact
Developers of subsequent versions of the HDP plugin should take into account this option and the added functionality. The procedure is not likely to change in newer versions of HDP as it uses Ambari's API which stays intact with newer versions.
Sahara-dashboard / Horizon impact
A checkbox 'Enable HDFS HA' in the 'Launch Cluster' wizard when the user chooses HDP 2.0.6 plugin.
- Primary assignee:
- Add a new attribute to cluster-level configs to indicate whether HA is enabled or not.
- Add new service classes to HDP 2.0.6 for Journal nodes and Zookeeper Failover Controllers
- Add new remote methods to hdp hadoopserver.py for remote commands
- Add new methods to generate new configurations according to cluster configuration
Unit Test service classes Unit Test new cluster specs Integration Test cluster creation with HA Integration Test cluster creation without HA
Update documentation to reflect new changes and to explain new options.