[SPARK-8366] maxNumExecutorsNeeded should properly handle failed tasks

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.4.0
• Fix Version/s: 1.5.0
• Component/s: Spark Core
• Target Version/s: 1.5.0
• Assignee: meiyoula
• Reporter: meiyoula
• Created: 15/Jun/15 02:00
• Updated: 12/Aug/15 06:19
• Resolved: 12/Aug/15 06:19

Description

I use the dynamic executor allocation function.
When an executor is killed, all running tasks on it will be failed. Until reach the maxTaskFailures, this failed task will re-run with a new task id.
But the ExecutorAllocationManager won't concern this new tasks to total and pending tasks, because the total stage task number only set when stage submitted.

Issue Links

• links to: [Github] Pull Request #6817 (XuTingjun)