Enabling High Availability for Spark Thrift Server

With EEPs 5.0.4 or 6.3.0 and later, you can enable high availability for the Spark Thrift Server. Note the following characteristics of high availability for the Spark Thrift Server:
  • Unlike a HiveServer2 high-availability (HA) configuration, all Spark Thrift Servers are in an active state. ZooKeeper keeps track of the Thrift Servers. ZooKeeper chooses one of them to work and makes a record of the choice. If one of the Thrift Servers goes down, ZooKeeper looks for another Thrift Server, makes a record, and works with it.
  • After configuration, you can use Beeline to connect to the Spark Thrift Server on each node. The Control System displays one thrift server as active with the others on standby, but you can connect to any of them.
  • If a Spark Thrift Server stops or fails, ZooKeeper removes the record for the failed Spark Thrift Server, and the client connects to the next one in the ZooKeeper list.
  • At its core, the running Spark Thrift Server is a job that you can start in YARN mode. This makes it possible to configure queues for the Spark Thrift Server in a multi-tenant cluster if high availability is enabled. You can do this by using the ./sbin/start-thriftserver script and applying the special properties that YARN provides for managing queues.
  • You don't need to configure load balancing. Spark handles load-balancing automatically through the use of parallelized requests and efficient resource management.
To enable high availability, use the following steps:
  1. Install Spark Thrift Server on all the cluster nodes where it is needed:
    On Ubuntu
    apt-get install mapr-spark-thriftserver
    On Red Hat / CentOS
    yum install mapr-spark-thriftserver
    On SLES
    zypper install mapr-spark-thriftserver
  2. Add the following properties to the /opt/mapr/spark/spark-<spark_version>/conf/hive-site.xml file on all the nodes where the Spark Thrift Server is installed
    <property>
    <name>hive.zookeeper.quorum</name>
    <value><zk_host1_>,<zk_host_2>,…,<zk_host_n></value>
    </property>
    
    <property>
    <name>hive.zookeeper.client.port</name>
    <value><zk_port></value>
    </property>
    
    <property>
    <name>hive.server2.support.dynamic.service.discovery</name>
    <value>true</value>
    </property>
    
    <property>
    <name>hive.server2.zookeeper.namespace</name>
    <value><zk_namespace></value>
    </property>
    For example:
    <property>
    <name>hive.zookeeper.quorum</name>
    <value>node1.cluster.com,node2.cluster.com,node3.cluster.com</value>
    </property>
    
    <property>
    <name>hive.zookeeper.client.port</name>
    <value>5181</value>
    </property>
    
    <property>
    <name>hive.server2.support.dynamic.service.discovery</name>
    <value>true</value>
    </property>
    
    <property>
    <name>hive.server2.zookeeper.namespace</name>
    <value>ts2-ts2</value>
    </property>
    NOTE
    The values that you provide for the hive.server2.zookeeper.namespace property should be different for the hive-site.xml in the Spark and Hive directories.
  3. Restart the Spark Thrift Server to apply the changes following the script in the .sbin directory at /opt/mapr/spark/spark-<spark_version>/ or by running a maprcli command on all configured nodes:
    ./sbin/stop-thriftserver.sh
    ./sbin/start-thriftserver.sh 
    or
    maprcli node services -nodes <host_1>,<host_2>,<host_n> -name spark-thriftserver -action restart
  4. Launch the Zookeeper command line interface, and check the Spark Thriftserver znode by running the following commands:
    /opt/mapr/zookeeper/zookeeper-<version>/bin/zkCli.sh -server <ip:port of zookeeper instance>
    ls /<hive.server2.zookeeper.namespace>
    For example:
    /opt/mapr/zookeeper/zookeeper-3.4.11/bin/zkCli.sh -server node1.cluster.com:5181
    ls /ts2-ts2
    [serverUri=node1.cluster.com:2304;version=;sequence=0000000000]
  5. Using Beeline, you can connect to the Spark Thrift Server by using the following string:
    beeline> !connect jdbc:hive2://<hostname -f>:5181/default;serviceDiscoveryMode=zooKeeper;zooKeeperNamespace=<hive.server2.zookeeper.namespace>;
    For example:
    ./bin/beeline
    Warning: Unable to determine $DRILL_HOME
    Beeline version 1.2.0-mapr-spark-MEP-6.0.0-1912 by Apache Hive
    beeline> !connect jdbc:hive2://node1.cluster.com:5181/default;ssl=true;serviceDiscoveryMode=zooKeeper;zooKeeperNamespace=ts2-ts2;auth=maprsasl;
    Connecting to jdbc:hive2://node1.cluster.com:5181/default;ssl=true;serviceDiscoveryMode=zooKeeper;zooKeeperNamespace=ts2-ts2;auth=maprsasl;
    20/03/29 21:38:19 WARN MaprSaslClient: SASL Server qopProperty: auth-confis different from Client: auth-conf,auth-int,auth.Using Server one
    Connected to: Spark SQL (version 2.4.4.0-mapr-630)
    Driver: Hive JDBC (version 1.2.0-mapr-spark-MEP-6.0.0-1912)
    Transaction isolation: TRANSACTION_REPEATABLE_READ
    1: jdbc:hive2://node1.cluster.com:5181/defaul> show databases;
    +-----------------+
    | databaseName |
    +-----------------+
    | default             |
    +-----------------+
    1 row selected (0.11 seconds)
NOTE
High availability for the Spark Thrift Server can be used in conjunction with HiveServer2 high availability. For more information about HiveServer2 high availability, see Enabling High Availability for Hive.