[YARN-4021] RuntimeException/YarnRuntimeException sent over to the client can cause client to assume a local fatal failure

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Hadoop Flags: Incompatible change
• Assignee: Unassigned
• Reporter: Anubhav Dhoot
• Created: 06/Aug/15 00:01
• Updated: 02/Jun/17 16:35

Description

Currently RuntimeException and its derived types such as YarnRuntimeExceptions are serialized over to the client and thrown at the client after YARN-731. This can cause issues like MAPREDUCE-6439 where we assume a local fatal exception has happened.
Instead we should have a way to distinguish local RuntimeException versus remote RuntimeException to avoid these issues. We need to go over all the current client side code that is expecting a remote RuntimeException inorder to make it work with this change.

Issue Links

• is related to: MAPREDUCE-6439 AM may fail instead of retrying if RM shuts down during the allocate call