[SPARK-30191] AM should update pending resource request faster when driver lost executor

• Type: Improvement
• Status: In Progress
• Priority: Minor
• Resolution: Unresolved
• Affects Version/s: 3.1.0
• Component/s: Spark Core, YARN
• Assignee: Unassigned
• Reporter: Max Xie
• Created: 09/Dec/19 14:06
• Updated: 17/May/20 18:14

Description

I run spark on yarn.  I found that when driver lost its executors because of machine hardware problem and all of service includes nodemanager, executor on the node has been killed,  it means that Resourcemanager can't update the containers info on the node until Resourcemanager try to remove the node,   but it always takes 10 mins or longger, and in the meantime, AM don't add the new resource request and driver missing the executors.

So maybe AM should add the factor `numExecutorsExiting` in YarnAllocator's method `

updateResourceRequests`  to optimize it.

 

 

 

Issue Links

• links to: GitHub Pull Request #26816