[
https://issues.apache.org/jira/browse/HADOOP-4483?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12641916#action_12641916
]
Tsz Wo (Nicholas), SZE commented on HADOOP-4483:
------------------------------------------------
+1 HADOOP-4483-v2.patch looks good to me.
>Are you folks saying that the approach adopted by this patch is not sufficient
>and it needs more changes to make it efficient?
Current fix is good enough for this issue. If there is anything we could do
for better performance, we could do it in a separated issue.
> getBlockArray in DatanodeDescriptor does not honor passed in maxblocks value
> ----------------------------------------------------------------------------
>
> Key: HADOOP-4483
> URL: https://issues.apache.org/jira/browse/HADOOP-4483
> Project: Hadoop Core
> Issue Type: Bug
> Components: dfs
> Affects Versions: 0.18.1
> Environment: hadoop-0.18.1 running on a cluster of 16 nodes.
> Reporter: Ahad Rana
> Priority: Critical
> Fix For: 0.18.2
>
> Attachments: HADOOP-4483-v2.patch, patch.HADOOP-4483
>
> Original Estimate: 1h
> Remaining Estimate: 1h
>
> The getBlockArray method in DatanodeDescriptor.java should honor the passed
> in maxblocks parameter. In its current form it passed in an array sized to
> min(maxblocks,blocks.size()) into the Collections.toArray method. As the
> javadoc for Collections.toArray indicates, the toArray method may discard the
> passed in array (and allocate a new array) if the number of elements returned
> by the iterator exceeds the size of the passed in array. As a result, the
> flawed implementation of this method would return all the invalid blocks for
> a data node in one go, and thus trigger the NameNode to send a DNA_INVALIDATE
> command to the DataNode with an excessively large number of blocks. This
> INVALIDATE command, in turn, could potentially take a very long time to
> process at the DataNode, and since DatanodeCommand(s) are processed in
> between heartbeats at the DataNode, this would trigger the NameNode to
> consider the DataNode to be offline / unresponsive (due to a lack of
> heartbeats).
> In our use-case at CommonCrawl.org, we regularly do large scale hdfs file
> deletions after certain stages of our map-reduce pipeline. These deletes
> would make certain DataNode(s) unresponsive, and thus impact the cluster's
> capability to properly balance file-system reads / writes across the whole
> available cluster. This problem only surfaced once we migrated from our 16.2
> deployment to the current 18.1 release.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
