[
https://issues.apache.org/jira/browse/HDFS-3701?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13461744#comment-13461744
]
Tsz Wo (Nicholas), SZE commented on HDFS-3701:
----------------------------------------------
- fetchLocatedBlocksAndGetLastBlockLength() returns the last block length but
the length is only used for checking whether it equals to -1. So how about
changing the return type to boolean (false means location unavailable) and
renaming it to fetchLocatedBlocks()? Then, openInfo could be simplified as
below
{code}
synchronized void openInfo() throws IOException {
for(int retries = 3; retries > 0; retries--) {
if (fetchLocatedBlocks()) {
//fetch block success
return;
} else {
// Last block location unavailable. When a cluster restarts,
// DNs may not report immediately. At this time partial block
// locations will not be available with NN for getting the length.
// Lets retry a few times to get the length.
DFSClient.LOG.warn("Last block locations unavailable. "
+ "Datanodes might not have reported blocks completely."
+ " Will retry for " + retries + " times");
waitFor(4000);
}
}
throw new IOException("Could not obtain the last block locations.");
}
{code}
- We may also change the return type of updateBlockInfo(..) to boolean since
the length is not used. I am fine if you want to keep the length.
- Throw InterruptedIOException, a subclass of IOException, instead of
IOException in waitFor(..).
> HDFS may miss the final block when reading a file opened for writing if one
> of the datanode is dead
> ---------------------------------------------------------------------------------------------------
>
> Key: HDFS-3701
> URL: https://issues.apache.org/jira/browse/HDFS-3701
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: hdfs client
> Affects Versions: 1.0.3
> Reporter: nkeywal
> Priority: Critical
> Attachments: HDFS-3701.branch-1.v2.merged.patch,
> HDFS-3701.ontopof.v1.patch, HDFS-3701.patch
>
>
> When the file is opened for writing, the DFSClient calls one of the datanode
> owning the last block to get its size. If this datanode is dead, the socket
> exception is shallowed and the size of this last block is equals to zero.
> This seems to be fixed on trunk, but I didn't find a related Jira. On 1.0.3,
> it's not fixed. It's on the same area as HDFS-1950 or HDFS-3222.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
