[REEF-1796] All REEF YARN jobs end with FORCE_CLOSED status on the client side

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Invalid
• Labels: bug yarn
• Assignee: Sergiy Matusevych
• Reporter: Sergiy Matusevych
• Created: 10/May/17 21:47
• Updated: 17/May/17 19:52
• Resolved: 17/May/17 19:48

Description

It looks like the connection between REEF driver and REEF client is being closed prematurely either on the client or on the driver side when using YARN runtime. As a result, REEF driver fails to communicate its final status to the client, and the client times out with FORCE_CLOSED status. That causes all unit tests to fail on YARN. From the logs it can be seen that REEF jobs complete normally on YARN; it's just the final status message that does not make it to the client.

To reproduce: run e.g. HelloREEF application on YARN,

.\bin\runreef.ps1 -VerboseLog -Jars .\lang\java\reef-examples\target\reef-examples-0.16.0-SNAPSHOT-shaded.jar -Class org.apache.reef.examples.hello.HelloREEFYarn

or run unt tests

.\bin\runtests.ps1 -Yarn -Jars ".\lang\java\reef-examples\target\reef-examples-0.16.0-SNAPSHOT-shaded.jar;.\lang\java\reef-tests\target\reef-tests-0.16.0-SNAPSHOT-test-jar-with-dependencies.jar"

Issue Links

• blocks: REEF-1444 Release 0.16
• relates to: REEF-1801 Improve logging when sending remote events in Wake