NameNode detects "Inconsistent diskspace" for directories with quota-enabled
subdirectories (introduced by HDFS-1377)
---------------------------------------------------------------------------------------------------------------------
Key: HDFS-2053
URL: https://issues.apache.org/jira/browse/HDFS-2053
Project: Hadoop HDFS
Issue Type: Bug
Components: name-node
Affects Versions: 0.20.3, 0.20.203.0
Environment: Hadoop release 0.20.203.0 with the HDFS-1377 patch
applied.
My impression is that the same issue exists also in the other branches where
the HDFS-1377 patch has been applied to (see description).
Reporter: Michael Noll
Priority: Minor
Fix For: 0.20.3, 0.20.203.0
*How to reproduce*
{code}
# create test directories
$ hadoop fs -mkdir /hdfs-1377/A
$ hadoop fs -mkdir /hdfs-1377/B
$ hadoop fs -mkdir /hdfs-1377/C
# ...add some test data (few kB or MB) to all three dirs...
# set space quota for subdir C only
$ hadoop dfsadmin -setSpaceQuota 1g /hdfs-1377/C
# the following two commands _on the parent dir_ trigger the warning
$ hadoop fs -dus /hdfs-1377
$ hadoop fs -count -q /hdfs-1377
{code}
Warning message in the namenode logs:
{code}
2011-06-09 09:42:39,817 WARN org.apache.hadoop.hdfs.server.namenode.NameNode:
Inconsistent diskspace for directory C. Cached: 433872320 Computed: 438465355
{code}
Note that the commands are run on the _parent directory_ but the warning is
shown for the _subdirectory_ with space quota.
*Background*
The bug was introduced by the HDFS-1377 patch, which is currently committed to
at least branch-0.20, branch-0.20-security, branch-0.20-security-204,
branch-0.20-security-205 and release-0.20.3-rc2. In the patch,
{{src/hdfs/org/apache/hadoop/hdfs/server/namenode/INodeDirectory.java}} was
updated to trigger the warning above if the cached and computed diskspace
values are not the same for a directory with quota.
The warning is written by {{computecontentSummary(long[] summary)}} in
{{INodeDirectory}}. In the method an inode's children are recursively walked
through while the {{summary}} parameter is passed and updated along the way.
{code}
/** {@inheritDoc} */
long[] computeContentSummary(long[] summary) {
if (children != null) {
for (INode child : children) {
child.computeContentSummary(summary);
}
}
{code}
The condition that triggers the warning message compares the current node's
cached diskspace (via {{node.diskspaceConsumed()}}) with the corresponding
field in {{summary}}.
{code}
if (-1 != node.getDsQuota() && space != summary[3]) {
NameNode.LOG.warn("Inconsistent diskspace for directory "
+getLocalName()+". Cached: "+space+" Computed: "+summary[3]);
{code}
However {{summary}} may already include diskspace information from other inodes
at this point (i.e. from different subtrees than the subtree of the node for
which the warning message is shown; in our example for the tree at
{{/hdfs-1377}}, {{summary}} can already contain information from
{{/hdfs-1377/A}} and {{/hdfs-1377/B}} when it is passed to inode
{{/hdfs-1377/C}}). Hence the cached value for {{C}} can incorrectly be
different from the computed value.
*How to fix*
The supplied patch creates a fresh summary array for the subtree of the current
node. The walk through the children passes and updates this {{subtreeSummary}}
array, and the condition is checked against {{subtreeSummary}} instead of the
original {{summary}}. The original {{summary}} is updated with the values of
{{subtreeSummary}} before it returns.
*Unit Tests*
I have run "ant test" on my patched build without any errors*. However the
existing unit tests did not catch this issue for the original HDFS-1377 patch,
so this might not mean anything. ;-)
That said I am unsure what the most appropriate way to unit test this issue
would be. A straight-forward approach would be to automate the steps in the
_How to reproduce section_ above and check whether the NN logs an incorrect
warning message. But I'm not sure how this check could be implemented. Feel
free to provide some pointers if you have some ideas.
*Note about Fix Version/s*
The patch _should_ apply to all branches where the HDFS-1377 patch has
committed to. In my environment, the build was Hadoop 0.20.203.0 release with
a (trivial) backport of HDFS-1377 (0.20.203.0 release does not ship with the
HDFS-1377 fix). I could apply the patch successfully to
{{branch-0.20-security}}, {{branch-0.20-security-204}} and
{{release-0.20.3-rc2}}, for instance. Since I'm a bit confused regarding the
upcoming 0.20.x release versions (0.20.x vs. 0.20.20x.y) I have been so bold
and added 0.20.203.0 to the list of affected versions even though it is
actually only affected when HDFS-1377 is applied to it...
Best,
Michael
*Well, I get one error for {{TestRumenJobTraces}} but first this seems to be
completely unrelated and second I get the same test error when running the
tests on the stock 0.20.203.0 release build.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
