[SPARK-15941] Netty RPC implementation ignores the executor bind address

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Incomplete
• Affects Version/s: 1.6.1
• Labels: bulk-closed
• Assignee: Unassigned
• Reporter: Marco Capuccini
• Created: 14/Jun/16 09:55
• Updated: 21/May/19 04:35
• Resolved: 21/May/19 04:35

Description

When using Netty RPC implementation, which is the default one in Spark 1.6.x, the executor addresses that I see in the Spark application UI (the one on port 4040) are the IP addresses of the machines, even if I start the slaves with the -H option, in order to bind each slave to the hostname of the machine.

This is a big deal when using Spark with HDFS, as the executor addresses need to match the hostnames of the DataNodes, to achieve data locality.

When setting spark.rpc=akka everything works as expected, and the executor addresses in the Spark UI match the hostname, which the slaves are bound to.