[jira] [Commented] (HDFS-3061) Cached directory size in INodeDirectory can get permantently out of sync with computed size, causing quota issues

["Kihwal Lee (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HDFS-3061?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13268446#comment-13268446 ] 

Kihwal Lee commented on HDFS-3061:
----------------------------------

We (mostly Koji) have tracked it down to abandonBlock().  Branch-1 is missing HDFS-1487. 
                
> Cached directory size in INodeDirectory can get permantently out of sync with computed size, causing quota issues
> -----------------------------------------------------------------------------------------------------------------
>
>                 Key: HDFS-3061
>                 URL: https://issues.apache.org/jira/browse/HDFS-3061
>             Project: Hadoop HDFS
>          Issue Type: Bug
>          Components: name-node
>    Affects Versions: 0.20.203.0
>         Environment: 0.20.203 with HDFS-1377 and HDFS-2053 patches applied
>            Reporter: Alex Holmes
>         Attachments: QuotaTestSimple.java
>
>
> It appears that there's a condition under which a HDFS directory with a space quota set can get to a point where the cached size for the directory can permanently differ from the computed value.  When this happens the following command:
> {code}
> hadoop fs -count -q /tmp/quota-test
> {code}
> results in the following output in the NameNode logs:
> {code}
> WARN org.apache.hadoop.hdfs.server.namenode.NameNode: Inconsistent diskspace for directory quota-test. Cached: 6000 Computed: 6072
> {code}
> I've observed both transient and persistent instances of this happening.  In the transient instances this warning goes away, but in the persistent instances every invocation of the {{fs -count -q}} command yields the above warning.
> I've seen instances where the actual disk usage of a directory is 25% of the cached value in INodeDirectory, which creates problems since the quota code uses this cached value to determine whether block write requests are permitted.
> This isn't easy to reproduce - I am able to (inconsistently) get HDFS into this state with a simple program which:
> #  Writes files into HDFS
> #  When a DSQuotaExceededException is encountered removes all files created in step 1
> #  Repeat step 1
> I'm going to try and come up with a more repeatable test case to reproduce this issue.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira