[ https://issues.apache.org/jira/browse/YARN-1036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13732637#comment-13732637 ]
Ravi Prakash commented on YARN-1036: ------------------------------------ Omkar! We don't want to delete a file. We want it to be localized again if it is detected to no longer exist. Which is what this patch is doing. This is what Koji is also saying. I already tested the patch does the right thing on my single node cluster. My test methodology was this. 1. Configure 2 directories for the NM. One on the node hard disk, and another on a pen drive. 2. Ran a sleep job with -files option specifying a file. 3. Make sure the file is localized on the pen drive. (If it isn't run another sleep job with a different file to be put in distcache) 4. unplug the pendrive (to simulate a bad disk). Before the patch, running a sleep job requesting the same old file in distcache didn't localize the file again. So if the job had required that file, it would have failed. After the patch, it detects that the file which was already localized is missing, and so it localizes it again. This is the behavior we want. Do you agree? > Distributed Cache gives inconsistent result if cache files get deleted from > task tracker > ----------------------------------------------------------------------------------------- > > Key: YARN-1036 > URL: https://issues.apache.org/jira/browse/YARN-1036 > Project: Hadoop YARN > Issue Type: Bug > Components: nodemanager > Affects Versions: 0.23.9 > Reporter: Ravi Prakash > Assignee: Ravi Prakash > Attachments: YARN-1036.branch-0.23.patch > > > This is a JIRA to backport MAPREDUCE-4342. I had to open a new JIRA because > that one had been closed. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira