While running the simplest of jobs (Pi) on MR2 in a fully secure configuration I have noticed that the job was failing on the reduce side with the following messages littering the nodemanager logs:
2012-01-19 08:35:32,544 ERROR org.apache.hadoop.mapred.ShuffleHandler: Shuffle error
org.apache.hadoop.util.DiskChecker$DiskErrorException: Could not find usercache/rvs/appcache/application_1326928483038_0001/output/attempt_1326928483038_0001_m_000003_0/file.out.index in any of the configured local directories
While digging further I found out that the permissions on the files/dirs were prohibiting nodemanager (running under the user yarn) to access these files:
$ ls -l /data/3/yarn/usercache/testuser/appcache/application_1327102703969_0001/output/attempt_1327102703969_0001_m_000001_0
-rw-r----- 1 testuser testuser 28 Jan 20 15:41 file.out
-rw-r----- 1 testuser testuser 32 Jan 20 15:41 file.out.index
Digging even further revealed that the group-sticky bit that was faithfully put on all the subdirectories between testuser and application_1327102703969_0001 was gone from output and attempt_1327102703969_0001_m_000001_0.
Looking into how these subdirectories are created (org.apache.hadoop.yarn.server.nodemanager.containermanager.localizer.ContainerLocalizer.initDirs())
Path appFileCacheDir = new Path(appBase, FILECACHE);
appsFileCacheDirs[i] = appFileCacheDir.toString();
lfs.mkdir(appFileCacheDir, null, false);
lfs.mkdir(new Path(appBase, OUTPUTDIR), null, false);
Reveals that lfs.mkdir ends up manipulating permissions and thus clears sticky bit from output and filecache.
At this point I'm at a loss about how this is supposed to work. My understanding was
that the whole sequence of events here was predicated on a sticky bit set so
that daemons running under the user yarn (default group yarn) can have access
to the resulting files and subdirectories down at output and below. Please let
me know if I'm missing something or whether this is just a bug that needs to be fixed.
On a related note, when the shuffle side of the Pi job failed the job itself didn't.
It went into the endless loop and only exited when it exhausted all the local storage
for the log files (at which point the nodemanager died and thus the job ended). Perhaps
this is even more serious side effect of this issue that needs to be investigated