[
https://issues.apache.org/jira/browse/HADOOP-3553?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Aaron Greenhouse updated HADOOP-3553:
-------------------------------------
Status: Open (was: Patch Available)
> Nested class TaskTracker.TaskInProgress needs additional synchronization
> ------------------------------------------------------------------------
>
> Key: HADOOP-3553
> URL: https://issues.apache.org/jira/browse/HADOOP-3553
> Project: Hadoop Core
> Issue Type: Bug
> Components: mapred
> Affects Versions: 0.17.0
> Environment: All Java platforms
> Reporter: Aaron Greenhouse
> Attachments: TaskTracker.patch
>
> Original Estimate: 1h
> Remaining Estimate: 1h
>
> The nested class TaskTracker.TaskInProgress needs additional synchronization
> to work properly with the Java Memory Model. Presumably this class is
> accessed by more than one thread, because it already contains
> synchronization. However, it needs additional synchronization, especially to
> protect access to the long fields lastProgressReport and taskTimeOut. Long
> fields are not guaranteed to be read/written atomically, so not only do you
> risk reading stale values, but you risk reading corrupted values.
> The field wasKilled also needs synchronization, as it is polled from within
> the TaskTracker class.
> I suggest the following improvements to the class
> - Make the fields task and taskStatus final. They are used this way already.
> Making them final clarifies there behavior in a current environment.
> - Add the synchronized modifier to the methods getLastProgressReport() and
> getTaskTimeout().
> - Make the field wasKilled private and add a new public synchronized
> wasKilled() getter method. Replace the use of the field with this method in
> TaskTracker.
> - Add a comment to localizeTask() indicating that the caller must be
> synchronized on this.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
