[
https://issues.apache.org/jira/browse/HDFS-3875?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13662214#comment-13662214
]
Kihwal Lee commented on HDFS-3875:
----------------------------------
bq. Can you explain this sleep here a little further? The assumption is that
the responder will come back and interrupt the streamer? Why do we need to wait
instead of just bailing out immediately with the IOE? Will this cause a
3-second delay in re-establishing the pipeline again?
This gives the responder time to send the checksum error back upstream, so that
the upstream node can blow up and exclude itself from the pipeline. This may
not be always ideal since there can be many different failure modes, but if
anything needs to be eliminated without knowing the cause, the source seems to
be a better candidate than the sink who actually verifies checksum.
Unless there is network issue in sending ACKs, responder will immediately
terminate and interrupt the main writer thread, so the thread won't stay up.
Even if the thread stays up for some reason, recoverRbw() during pipeline
recovery will interrupt the thread, so there won't be 3 second delay.
If the last node in a pipeline has a faulty NIC, two upstream nodes will be
eliminated (in 3 replica case) and after adding a new dn to the end of
pipeline, the faulty node will be removed. Issues on intermediate nodes will be
handled in less number of iterations and the worst case will be when data is
corrupt in DFSOutputStream, which will be detected after hitting the maximum
number of retries. There will be no recovery in this case.
> Issue handling checksum errors in write pipeline
> ------------------------------------------------
>
> Key: HDFS-3875
> URL: https://issues.apache.org/jira/browse/HDFS-3875
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: datanode, hdfs-client
> Affects Versions: 2.0.2-alpha
> Reporter: Todd Lipcon
> Assignee: Kihwal Lee
> Priority: Critical
> Attachments: hdfs-3875.branch-0.23.no.test.patch.txt,
> hdfs-3875.branch-0.23.patch.txt, hdfs-3875.branch-0.23.patch.txt,
> hdfs-3875.branch-0.23.with.test.patch.txt, hdfs-3875.patch.txt,
> hdfs-3875.patch.txt, hdfs-3875.trunk.no.test.patch.txt,
> hdfs-3875.trunk.no.test.patch.txt, hdfs-3875.trunk.patch.txt,
> hdfs-3875.trunk.patch.txt, hdfs-3875.trunk.with.test.patch.txt,
> hdfs-3875.trunk.with.test.patch.txt, hdfs-3875-wip.patch
>
>
> We saw this issue with one block in a large test cluster. The client is
> storing the data with replication level 2, and we saw the following:
> - the second node in the pipeline detects a checksum error on the data it
> received from the first node. We don't know if the client sent a bad
> checksum, or if it got corrupted between node 1 and node 2 in the pipeline.
> - this caused the second node to get kicked out of the pipeline, since it
> threw an exception. The pipeline started up again with only one replica (the
> first node in the pipeline)
> - this replica was later determined to be corrupt by the block scanner, and
> unrecoverable since it is the only replica
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
