[SPARK-14437] Spark using Netty RPC gets wrong address in some setups

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.6.0, 1.6.1
• Fix Version/s: 2.0.0
• Component/s: Block Manager, Spark Core

Environment

AWS, Docker, Flannel

• Assignee: Shixiong Zhu
• Reporter: Kevin Hogeland
• Created: 06/Apr/16 19:33
• Updated: 15/Jun/17 11:29
• Resolved: 09/Apr/16 00:18

Description

Netty can't get the correct origin address in certain network setups. Spark should handle this, as relying on Netty correctly reporting all addresses leads to incompatible and unpredictable network states. We're currently using Docker with Flannel on AWS. Container communication looks something like: Container 1 (1.2.3.1) -> Docker host A (1.2.3.0) -> Docker host B (4.5.6.0) -> Container 2 (4.5.6.1)

If the client in that setup is Container 1 (1.2.3.4), Netty channels from there to Container 2 will have a client address of 1.2.3.0.

The RequestMessage object that is sent over the wire already contains a senderAddress field that the sender can use to specify their address. In NettyRpcEnv#internalReceive, this is replaced with the Netty client socket address when null. senderAddress in the messages sent from the executors is currently always null, meaning all messages will have these incorrect addresses (we've switched back to Akka as a temporary workaround for this). The executor should send its address explicitly so that the driver doesn't attempt to infer addresses based on possibly incorrect information from Netty.

Issue Links

• relates to: SPARK-14849 shuffle broken when accessing standalone cluster through NAT
• links to: [Github] Pull Request #12240 (zsxwing)