[ https://issues.apache.org/jira/browse/SPARK-30325?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
haiyangyu updated SPARK-30325: ------------------------------ Description: Kill tasks which succeeded in origin stage when new retry stage has started the same task and hasn't finished. This can decreate stage run time, resouce cost and most important of all, it can avoid a bug which can cause app hung. The bugs occurs in the coren case as follows: 1. The stage occurs for fetchFailed and some task hasn't finished, scheduler will resubmit a new stage as retry with those unfinished tasks. 2. The unfinished task in origin stage finished and the same task on the new retry stage hasn't finished, it will mark the task partition on the new retry stage as succesuful. 3. The executor running those 'successful task' crashed, it cause taskSetManager run executorLost to rescheduler the task on the executor, here will cause copiesRunning decreate 1 twice, beause those 'successful task' are not finished, the variable copiesRunning will decreate to -1 as result. 4. 'dequeueTaskFromList' will use copiesRunning equal 0 as reschedule basis when rescheduler tasks, and now it is -1, can't to reschedule, and the app will hung forever. > Stage retry and executor crashed cause app hung up forever > ---------------------------------------------------------- > > Key: SPARK-30325 > URL: https://issues.apache.org/jira/browse/SPARK-30325 > Project: Spark > Issue Type: Bug > Components: Spark Core > Affects Versions: 2.4.0, 2.4.4 > Reporter: haiyangyu > Priority: Major > > Kill tasks which succeeded in origin stage when new retry stage has started > the same task and hasn't finished. > This can decreate stage run time, resouce cost and most important of all, it > can avoid a bug which can cause app hung. > The bugs occurs in the coren case as follows: > 1. The stage occurs for fetchFailed and some task hasn't finished, scheduler > will resubmit a new stage as retry with those unfinished tasks. > 2. The unfinished task in origin stage finished and the same task on the new > retry stage hasn't finished, it will mark the task partition on the new retry > stage as succesuful. > 3. The executor running those 'successful task' crashed, it cause > taskSetManager run executorLost to rescheduler the task on the executor, here > will cause copiesRunning decreate 1 twice, beause those 'successful task' are > not finished, the variable copiesRunning will decreate to -1 as result. > 4. 'dequeueTaskFromList' will use copiesRunning equal 0 as reschedule basis > when rescheduler tasks, and now it is -1, can't to reschedule, and the app > will hung forever. -- This message was sent by Atlassian Jira (v8.3.4#803005) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org