[
https://issues.apache.org/jira/browse/HADOOP-4068?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12628552#action_12628552
]
Amar Kamat commented on HADOOP-4068:
------------------------------------
I suggest we do something like
{code:title=JobInProgress#failedTask()|borderStyle=solid}
public void failedTask(TaskInProgress tip, TaskAttemptID taskid, String reason,
TaskStatus.Phase phase, TaskStatus.State state,
String trackerName, JobTrackerInstrumentation metrics)
{
TaskStatus status = ....
TaskStatus.State oldState = tip.getTaskStatus(taskid).getRunState();
updateTaskStatus(tip, status, metrics);
TaskStatus.State newState = tip.getTaskStatus(taskid).getRunState();
// Make sure that the tip fails only if the state changes for the attempt
// that fails it
if (oldState == newState) {
return;
}
JobHistory.Task.logFailed(...);
{code}
> JobTracker might wrongly log a tip as failed
> --------------------------------------------
>
> Key: HADOOP-4068
> URL: https://issues.apache.org/jira/browse/HADOOP-4068
> Project: Hadoop Core
> Issue Type: Bug
> Components: mapred
> Reporter: Amar Kamat
>
> Consider the following case
> 1) attempt _attempt_1_0_ from tip _tip_1_ that ran on tracker _tracker_1_
> failed
> 2) jobtracker will mark _attempt_1_0_ for removal under _tracker_1_. Marking
> basically means removal of the mapping _tracker_1_->_attempt_1_0_
> 3) Marked attempts are removed only on next heartbeat from _tracker__1 or
> when _tracker_1_ is lost.
> 4) Consider a case where _tracker_1_ goes down.
> 5) In the meanwhile attempt _attempt_1_1_ succeeds on _tracker_2_ and the
> jobtracker marks the tip _tip_1_ as complete
> 6) Now the expiry-tracker thread detect that _tracker_1_ is lost and fails
> all the attempt under _tracker_1_.
> 7) Here the jobtracker will kill _attempt_1_0_ *again* and log tip _tip_1_ as
> failed in the history although tip _tip_1_ is really complete/succeeded.
> The events in the history file would be something like
> {noformat}
> tip_1 start
> ---------
> attempt_1_0 start
> attempt_1_0 failed
> ---------
> attempt_1_1 start
> attempt_1_1 finished
> tip_1 finished
> ---------
> tip_1 failed
> {noformat}
> Note that this true even for tasks that expire. Tasks that are scheduled and
> never come back are killed by the {{ExpireLaunchingTasks}} thread. It will
> also call {{JobInProgress.failedTask()}} which will fail the attempt and log
> the TIP as failed.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
