[
https://issues.apache.org/jira/browse/HDFS-8145?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jing Zhao updated HDFS-8145:
----------------------------
Attachment: HDFS-8145.002.patch
Thanks for the review, Zhe! Update the patch to address your comments:
# I just removed isInECZone method from ECZoneManager since it is only called
once there. Also in this way in FSDirectory#isInECZone we do not need to
acquire the read lock again.
# Actually for editlog transaction we do not need to persist these two numbers
even if later we no longer use the hard-coded numbers. For both {{addBlock}}
and {{updateBlocks}}, the schema information can be retrieved from the
corresponding EC zone. We need to write schema information for INode only if we
decide to drop the ECZone design, and that change will have a much bigger scope
and a lot of pending issues in my opinion.
> Fix the editlog corruption exposed by failed TestAddStripedBlocks
> -----------------------------------------------------------------
>
> Key: HDFS-8145
> URL: https://issues.apache.org/jira/browse/HDFS-8145
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Reporter: Jing Zhao
> Assignee: Jing Zhao
> Attachments: HDFS-8145.000.patch, HDFS-8145.001.patch,
> HDFS-8145.002.patch
>
>
> {{TestAddStripedBlocks}} failed with some editlog corruption. Did some
> debugging, I can see at least two issues:
> # DFSStripedOutputStream tries to send out an empty packet to close the block
> even if writing 0 bytes
> # Because of the above, NN tries to close the file. This exposes another bug
> in {{BlockInfoStriped}}, which writes its data/parity block numbers into the
> close editlog but do not read them while loading.
> This jira aims to only fix #2 and HDFS-8166 is used for fixing #1.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
