[
https://issues.apache.org/jira/browse/UIMA-5428?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16045025#comment-16045025
]
Burn Lewis commented on UIMA-5428:
----------------------------------
Megs produced some initialization failures that may be due to non-atomic
creation of the AE descriptor by multiple JPs:
org.apache.uima.util.InvalidXMLException: Invalid descriptor at
file:/home/burn/ducc/logs/7/7-uima-ae-descriptor.xml.
Creating a unique temporary file and renaming with the atomic Files.move method
appears to be safe.
An alternative approach would be to:
- continue to create the descriptor in the JD
- if the JP cannot read it (presumably because of a non-shared FS) then create
it with a unique name
Pro: - avoids problems with multiple processes writing the same file
Con - multiple descriptors (small files) would be created in the non-shared
local FS if a job had multiple JPs on the same node
> DUCC minimally support user log directory not in shared file system
> -------------------------------------------------------------------
>
> Key: UIMA-5428
> URL: https://issues.apache.org/jira/browse/UIMA-5428
> Project: UIMA
> Issue Type: New Feature
> Components: DUCC
> Reporter: Lou DeGenaro
> Assignee: Burn Lewis
> Fix For: 2.2.1-Ducc
>
>
> Support submitted work where user specified log directory is not in shared
> file system.
> For example, if user specifies /tmp/foobar as the log directory for a
> submitted job, the job will run and the /tmp on each system where the JD and
> JPs run will contain written files, such as JD and JP logs files. But no
> ducc-mon support. ducc-mon will only be able to see the /tmp of its local
> file system. Any files located in /tmp on remote machines (e.g. machines
> other than the DUCC head node) will not be visible via ducc-mon.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
