Making spark use /etc/hosts file for binding in YARN cluster mode

问题

Have a spark cluster setup on a machine with two inets, one public another private. The /etc/hosts file in the cluster has the internal ip of all the other machines in the cluster, like so.

internal_ip FQDN

However when I request a SparkContext via pyspark in YARN client mode(pyspark --master yarn --deploy-mode client), akka binds onto the public ip and thus a time out occurs.

15/11/07 23:29:23 INFO Remoting: Starting remoting
15/11/07 23:29:23 INFO Remoting: Remoting started; listening on addresses :[akka.tcp://sparkYarnAM@public_ip:44015]
15/11/07 23:29:23 INFO util.Utils: Successfully started service 'sparkYarnAM' on port 44015.
15/11/07 23:29:23 INFO yarn.ApplicationMaster: Waiting for Spark driver to be reachable.
15/11/07 23:31:30 ERROR yarn.ApplicationMaster: Failed to connect to driver at yarn_driver_public_ip:48875, retrying ...
15/11/07 23:31:30 ERROR yarn.ApplicationMaster: Uncaught exception: 
org.apache.spark.SparkException: Failed to connect to driver!
    at org.apache.spark.deploy.yarn.ApplicationMaster.waitForSparkDriver(ApplicationMaster.scala:427)
    at org.apache.spark.deploy.yarn.ApplicationMaster.runExecutorLauncher(ApplicationMaster.scala:293)
    at org.apache.spark.deploy.yarn.ApplicationMaster.run(ApplicationMaster.scala:149)
    at org.apache.spark.deploy.yarn.ApplicationMaster$$anonfun$main$1.apply$mcV$sp(ApplicationMaster.scala:574)
    at org.apache.spark.deploy.SparkHadoopUtil$$anon$1.run(SparkHadoopUtil.scala:66)
    at org.apache.spark.deploy.SparkHadoopUtil$$anon$1.run(SparkHadoopUtil.scala:65)
    at java.security.AccessController.doPrivileged(Native Method)
    at javax.security.auth.Subject.doAs(Subject.java:422)
    at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1657)
    at org.apache.spark.deploy.SparkHadoopUtil.runAsSparkUser(SparkHadoopUtil.scala:65)
    at org.apache.spark.deploy.yarn.ApplicationMaster$.main(ApplicationMaster.scala:572)
    at org.apache.spark.deploy.yarn.ExecutorLauncher$.main(ApplicationMaster.scala:599)
    at org.apache.spark.deploy.yarn.ExecutorLauncher.main(ApplicationMaster.scala)
15/11/07 23:31:30 INFO yarn.ApplicationMaster: Final app status: FAILED, exitCode: 10, (reason: Uncaught exception: org.apache.spark.SparkException: Failed to connect to driver!)
15/11/07 23:31:30 INFO yarn.ApplicationMaster: Unregistering ApplicationMaster with FAILED (diag message: Uncaught exception: org.apache.spark.SparkException: Failed to connect to driver!)
15/11/07 23:31:30 INFO yarn.ApplicationMaster: Deleting staging directory .sparkStaging/application_1446960366742_0002

As seen from the log, private IP is completely ignored, how can I make YARN and spark use the private IP address as specified in the hosts file ?

Cluster was provisioned using Ambari(HDP 2.4)

回答1:

+1 for the question.

Spark uses Akka for communication.

So it's more of an Akka question than Spark.

If you need to bind your network interface to a different address - use akka.remote.netty.tcp.bind-hostname and akka.remote.netty.tcp.bind-port settings.

http://doc.akka.io/docs/akka/snapshot/additional/faq.html#Why_are_replies_not_received_from_a_remote_actor_

回答2:

This is currently an issue in spark, the only way to get spark to bind to the proper interface is to use custom nameservers.

Spark essentially does a hostname lookup and uses the IP address that it finds to bind with Akka. Workaround is to create a custom bind zone and run a nameserver.

https://issues.apache.org/jira/browse/SPARK-5113

来源：https://stackoverflow.com/questions/33591393/making-spark-use-etc-hosts-file-for-binding-in-yarn-cluster-mode