[
https://issues.apache.org/jira/browse/HDFS-1467?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Todd Lipcon updated HDFS-1467:
------------------------------
Attachment: hdfs-1467.txt
It turns out there's a really big bug in pipeline construction for append...
I'm not sure why more tests aren't failing more often!
Currently in BlockReceiver's constructor, we copy the reference to the provided
block, and increment its generation stamp to the {{newGs}}. But, this is just a
reference to the same Block object used in DataXceiver.opWriteBlock. So, when
the DN tries to set up the append pipeline to its downstream mirror, it ends up
passing the new generation stamp version of the block, and the downstream
mirror will fail to construct a pipeline.
The end result is that we can never successfully construct a pipeline for
append with more than one datanode in it! This test would fail about 2/3 of the
time since it would see the old replica on one of the nodes that didn't make it
into the append pipeline.
This patch fixees the BlockReceiver constructor to take a copy of the block,
and TestPipelines seems to pass now. I also added some extra debug logging just
to illustrate the problem.
This probably should not be committed until we can add another test which shows
the problem 100% of the time.
> TestPipelines failing on trunk
> ------------------------------
>
> Key: HDFS-1467
> URL: https://issues.apache.org/jira/browse/HDFS-1467
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: data-node
> Affects Versions: 0.22.0
> Reporter: Todd Lipcon
> Priority: Critical
> Attachments: failed-TestPipelines.txt, hdfs-1467.txt
>
>
> TestPipelines appears to be failing on trunk:
> Should be RBW replica after sequence of calls append()/write()/hflush()
> expected:<RBW> but was:<FINALIZED>
> junit.framework.AssertionFailedError: Should be RBW replica after sequence of
> calls append()/write()/hflush() expected:<RBW> but was:<FINALIZED>
> at
> org.apache.hadoop.hdfs.TestPipelines.pipeline_01(TestPipelines.java:109)
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
