[
https://issues.apache.org/jira/browse/HADOOP-2078?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12536341
]
Konstantin Shvachko commented on HADOOP-2078:
---------------------------------------------
testZeroSizeFile() does not fail because DFSClient writes 0 bytes into the file
before closing.
When the output stream buffer is empty DFSClient .endBlock() allocates one
block and sends 0 bytes to 2 data-nodes.
The data-nodes create one empty data-file, and one not empty meta-file each and
then report to the name-node that the block have been received.
So empty file is represented by one block of size 0.
In the case you describe fileBlocks == null and pendingFile != null. So
condition
{code}
fileBlocks == null || pendingFile == null
{code}
is equivalent to
{code}
fileBlocks == null
{code}
They are not equivalent only if fileBlocks != null and pendingFile == null,
which never happens. Because no file means no blocks.
> Name-node should be able to close empty files.
> ----------------------------------------------
>
> Key: HADOOP-2078
> URL: https://issues.apache.org/jira/browse/HADOOP-2078
> Project: Hadoop
> Issue Type: Bug
> Affects Versions: 0.15.0
> Reporter: Konstantin Shvachko
> Assignee: Konstantin Shvachko
> Fix For: 0.16.0
>
> Attachments: emptyClose.patch
>
>
> When I try to close an empty file, the name-node throws an exception "Could
> not complete write to file"
> and issues a warning "NameSystem.completeFile: failed to complete".
> I don't see any reason why empty files should not be allowed.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
