[
https://issues.apache.org/jira/browse/HADOOP-5724?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12708466#action_12708466
]
Raghu Angadi commented on HADOOP-5724:
--------------------------------------
One reason could be because if there is a block report that is sent by DN
before seeing the deletion request, and seen by NN after sending the deletion
request, that block will be added again to the datanode. There might be other
related issues. Unless these are all sorted out, it is better for block list
for a DN at NN to reflect what DN says it is.
> But in case of block deletion due to file deletion etc., blocks are removed
> from blocksMap immediately.
I don't think so... the blocks are removed from the datanode list only after
the blockreport. from what I can see, could be mistaken. Otherwise, when a
file is deleted, a block could be safely deleted from the datanode map. Does
getstamp handle all the case where NN might reuse the same block id before
datanodes delete the prev incarnations?
Also the extra overhead for reporting the deleted blocks is more than
compensated by less frequent block reports.
> Datanode should report deletion of blocks to Namenode explicitly
> ----------------------------------------------------------------
>
> Key: HADOOP-5724
> URL: https://issues.apache.org/jira/browse/HADOOP-5724
> Project: Hadoop Core
> Issue Type: Bug
> Reporter: Suresh Srinivas
> Assignee: Suresh Srinivas
> Fix For: 0.21.0
>
> Attachments: blockdel.patch, blockdel.patch
>
>
> Currently datanode notifies namenode newly added blocks and the blocks that
> are corrupt. There is no explicit message from the datanode to the namenode
> to indicate the deletion of blocks. Block reports from the datanode is the
> only way for the namenode to learn about the deletion of blocks at a
> datanode. With the addition of explicit request to indicate to block
> deletion, block report interval (which is currently 1 hour) can be increased
> to a longer duration. This reduces load on both namenode and datanodes.
>
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
