[SPARK-18761] Uncancellable / unkillable tasks may starve jobs of resoures

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 2.0.3, 2.1.1, 2.2.0
• Component/s: Spark Core
• Assignee: Josh Rosen
• Reporter: Josh Rosen
• Created: 07/Dec/16 06:26
• Updated: 20/Dec/16 23:57
• Resolved: 20/Dec/16 02:44

Description

Spark's current task cancellation / task killing mechanism is "best effort" in the sense that some tasks may not be interruptible and may not respond to their "killed" flags being set. If a significant fraction of a cluster's task slots are occupied by tasks that have been marked as killed but remain running then this can lead to a situation where new jobs and tasks are starved of resources because zombie tasks are holding resources.

I propose to address this problem by introducing a "task reaper" mechanism in executors to monitor tasks after they are marked for killing in order to periodically re-attempt the task kill, capture and log stacktraces / warnings if tasks do not exit in a timely manner, and, optionally, kill the entire executor JVM if cancelled tasks cannot be killed within some timeout.

Issue Links

• relates to: SPARK-17064 Reconsider spark.job.interruptOnCancel
• links to: [Github] Pull Request #16189 (JoshRosen)
• links to: [Github] Pull Request #16358 (JoshRosen)