[
https://issues.apache.org/jira/browse/HDFS-814?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12794446#action_12794446
]
dhruba borthakur commented on HDFS-814:
---------------------------------------
> but it has the following issue: Applications that don't care about very
> accurate file lengths will pay the cost for files
This will happen only if the file is being written to when somebody else does a
getFileStatus on the file. This should never happen for the most typical app
that runs on HDFS... a map-reduce job.
>Cost of ls -r of a dir (say MR output dir) can go up when some of the files in
>the subtree are open for writing.
I suspect that this is not a typical use-case. The MR-job output directory will
typically be empty until the job is committed and all files get renamed into
the out directory (from the tmp directory).
I am good for this patch because this does not introduce a
FileSystem/FileContext API.
> Add an api to get the visible length of a DFSDataInputStream.
> -------------------------------------------------------------
>
> Key: HDFS-814
> URL: https://issues.apache.org/jira/browse/HDFS-814
> Project: Hadoop HDFS
> Issue Type: New Feature
> Components: hdfs client
> Reporter: Tsz Wo (Nicholas), SZE
> Assignee: Tsz Wo (Nicholas), SZE
> Fix For: 0.21.0, 0.22.0
>
> Attachments: h814_20091221.patch, h814_20091221_0.21.patch
>
>
> Hflush guarantees that the bytes written before are visible to the new
> readers. However, there is no way to get the length of the visible bytes.
> The visible length is useful in some applications like SequenceFile.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
