[
https://issues.apache.org/jira/browse/HDFS-1282?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12885955#action_12885955
]
Andrew Ryan commented on HDFS-1282:
-----------------------------------
If 'df' parsing is broken on some operating systems, we could either fix them,
or declare their df output to be 'untrusted'. But on platforms with reliable df
output, I think we should take advantage of it as a means by which to catch a
rare but unpleasant class of failures.
> namenode should reject datanodes which send impossible block reports
> --------------------------------------------------------------------
>
> Key: HDFS-1282
> URL: https://issues.apache.org/jira/browse/HDFS-1282
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: data-node, name-node
> Affects Versions: 0.20.1
> Reporter: Andrew Ryan
>
> Over the past few weeks we've had several datanodes with bad disks that
> suffer ext3 corruption, and consequently start reporting impossible values
> for how full they are. This particular node, for example, has a configured
> capacity of 10.86TB but reports 1733.95TB used, for a total of 15973.57%
> utilization.
> Node Last Contact Admin State Configured Capacity (TB) Used
> (TB) Non DFS Used (TB) Remaining (TB) Used (%)
> Used (%) Remaining (%) Blocks
> hadoop2254 44 In Service 10.86 1733.95 0 5.24
> 15973.57 48.25 65602
> If we can avoid generating such bogus data on the datanode that would be
> great. But if the namenode receives such an impossible block report, it
> should definitely consider that datanode to be not trustworthy, and in my
> opinion, make it dead.
> The "fix" in our case was either to fsck or replace the bad disk.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
