[
https://issues.apache.org/jira/browse/HDFS-4273?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14549996#comment-14549996
]
Masatake Iwasaki commented on HDFS-4273:
----------------------------------------
I'm looking into this and writing my understanding for other reviewers here:
All of HDFS-4273, HDFS-5917 and HDFS-6022 addresses improvement of refreshing
{{deadNodes}}. I think HDFS-6022 is the most promising. (Actually, newest v8
patch omitted the deadNodes part.)
Other issues addressed here are
# There is a case it should retry but don't
# There is race condition around {{failures}}
HDFS-5776 changed a lot of relevant code around {{chooseDataNode}} method and
the v8 patch is difficult to rebase.
Fixes around {{seekToNewSource}} alone may work and I'll try to update the
patch.
> Fix some issue in DFSInputstream
> --------------------------------
>
> Key: HDFS-4273
> URL: https://issues.apache.org/jira/browse/HDFS-4273
> Project: Hadoop HDFS
> Issue Type: Bug
> Affects Versions: 2.0.2-alpha
> Reporter: Binglin Chang
> Assignee: Binglin Chang
> Priority: Minor
> Attachments: HDFS-4273-v2.patch, HDFS-4273.patch, HDFS-4273.v3.patch,
> HDFS-4273.v4.patch, HDFS-4273.v5.patch, HDFS-4273.v6.patch,
> HDFS-4273.v7.patch, HDFS-4273.v8.patch, TestDFSInputStream.java
>
>
> Following issues in DFSInputStream are addressed in this jira:
> 1. read may not retry enough in some cases cause early failure
> Assume the following call logic
> {noformat}
> readWithStrategy()
> -> blockSeekTo()
> -> readBuffer()
> -> reader.doRead()
> -> seekToNewSource() add currentNode to deadnode, wish to get a
> different datanode
> -> blockSeekTo()
> -> chooseDataNode()
> -> block missing, clear deadNodes and pick the currentNode again
> seekToNewSource() return false
> readBuffer() re-throw the exception quit loop
> readWithStrategy() got the exception, and may fail the read call before
> tried MaxBlockAcquireFailures.
> {noformat}
> 2. In multi-threaded scenario(like hbase), DFSInputStream.failures has race
> condition, it is cleared to 0 when it is still used by other thread. So it is
> possible that some read thread may never quit. Change failures to local
> variable solve this issue.
> 3. If local datanode is added to deadNodes, it will not be removed from
> deadNodes if DN is back alive. We need a way to remove local datanode from
> deadNodes when the local datanode is become live.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
