JobState can be persisted to HDFS or Zookeeper. At various points in the lifecycle it's accessed with different user credentials thus may cause errors depending on how permissions are set.
When submitting a MR job, templeton.JarDelegator is used.
It calls LauncherDelegator#queueAsUser() which runs TempletonControllerJob with UserGroupInformation.doAs().
TempletonControllerJob will in turn create JobState and persist it.
LauncherDelegator.registerJob() also modifies JobState but w/o doing a doAs()
So in the later case it's possible that the persisted state of JobState by a different user than one that created/owns the file.
templeton.tool.HDFSCleanup tries to delete these files w/o doAs.
'childid' file, for example, is created with rw-r-
and it's parent directory (job_201308051224_0001) has rwxr-xr-x.
HDFSStorage doesn't set file permissions explicitly so it must be using default permissions.
So there is a potential issue here (depending on UMASK) especially once
HIVE-4601 is addressed.
Actually, even w/o
HIVE-4601 the user that owns the WebHCat process is likely different than the one submitting a request.
The default for templeton.storage.class is org.apache.hcatalog.templeton.toolHDFSStorage, but it's likely that most production environments change it to Zookeeper, which may explain why this issue is not commonly seen.