[
https://issues.apache.org/jira/browse/HDFS-8033?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14508177#comment-14508177
]
Jing Zhao commented on HDFS-8033:
---------------------------------
Thanks for updating the patch, Zhe! Some comments on the 002 patch:
# I think we do not need to override the blockSeekTo method. The return type of
blockSeekTo is a DataNodeInfo, which does not suit striped stream very well.
Thus I guess we only need to define a new {{seekToNextBlock}} method which have
the same functionality but different signature.
# To me {{currentNode}} and {{blockReader}} should be internal states only used
by {{DFSInputStream}}. {{DFSStripedInputStream}} should only use its own
reader/DN list, instead of setting {{currentNode}} and {{blockReader}} and
calling {{DFSInputStream#readBuffer}}. We can define a different version of
{{readBuffer}} method which utilizes the reader list to read. And this new
{{readBuffer}} method does not need to override the one in DFSInputStream.
# Another reason that we do not reuse the {{DFSInutStream#readBuffer}} is that
the retry logic is different. From striped inputstream we do not need to try
another source when hitting any failure.
# We need more tests. Now instead of injecting ec blocks, we can directly use
DFSStripedOutputstream for writing. And then we can compare the reading results
(of both stateful read and pread) with the original written content.
> Erasure coding: stateful (non-positional) read from files in striped layout
> ---------------------------------------------------------------------------
>
> Key: HDFS-8033
> URL: https://issues.apache.org/jira/browse/HDFS-8033
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Reporter: Zhe Zhang
> Assignee: Zhe Zhang
> Attachments: HDFS-8033.000.patch, HDFS-8033.001.patch,
> HDFS-8033.002.patch
>
>
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
