[YARN-8346] Upgrading to 3.1 kills running containers with error "Opportunistic container queue is full"

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: 3.1.0, 3.0.2
• Fix Version/s: 3.1.0, 2.10.0, 3.2.0, 2.9.2, 3.0.3
• Target Version/s: 3.1.1, 3.0.3
• Hadoop Flags: Reviewed
• Assignee: Jason Darrell Lowe
• Reporter: Rohith Sharma K S
• Created: 23/May/18 12:30
• Updated: 07/Nov/18 01:34
• Resolved: 24/May/18 07:08

Description

It is observed while rolling upgrade from 2.8.4 to 3.1 release, all the running containers are killed and second attempt is launched for that application. The diagnostics message is "Opportunistic container queue is full" which is the reason for container killed.

In NM log, I see below logs for after container is recovered.

2018-05-23 17:18:50,655 INFO org.apache.hadoop.yarn.server.nodemanager.containermanager.scheduler.ContainerScheduler: Opportunistic container [container_e06_1527075664705_0001_01_000001] will not be queued at the NMsince max queue length [0] has been reached

Following steps are executed for rolling upgrade

1. Install 2.8.4 cluster and launch a MR job with distributed cache enabled.
2. Stop 2.8.4 RM. Start 3.1.0 RM with same configuration.
3. Stop 2.8.4 NM batch by batch. Start 3.1.0 NM batch by batch.

Issue Links

• blocks: HADOOP-15501 [Umbrella] Upgrade efforts to Hadoop 3.x