[
https://issues.apache.org/jira/browse/HDFS-9092?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14804689#comment-14804689
]
Yongjun Zhang commented on HDFS-9092:
-------------------------------------
Thanks [~brandonli].
Good idea, below is some comment I made in the code, hope it helps:
{quote}
/**
* Overlapping Write Request Handling
* A write request can be in three states:
* s0. just created, with data != null
* s1. dumped as length "count", and data set to null
* s2. read back from dumped area as length "count"
*
* Write requests may have overlapping range, we detect this by comparing
* the data offset range of the request against the current offset of data
* already written to HDFS. There are two categories:
*
* 1. If the beginning part of a new write request data is already written
* due to an earlier request, we alter the new request by trimming this
* portion before the new request enters state s0, and the originalCount is
* remembered.
*
* 2. If the lower end of the write request range is beyond the current
* offset of data already written, we put the request into cache, and detect
* the overlapping when taking the request out from cache.
*
* For category 2, if we find out that a write request overlap with another,
* this write request is already in state s0, s1, or s3. We trim the
* beginning part of this request, by remembering the size of this portion
* as trimDelta. So the resulted offset of the write request is
* "offset + trimDelta" and the resulted size of the write request is
* "count - trimDelta".
*
* What important to notice is, if the request is in s1 when we do the
* trimming, the data dumped is of size "count", so when we load
* the data back from dumped area, we should set the position of the data
* buffer to trimDelta.
*/
{quote}
> Nfs silently drops overlapping write requests, thus data copying can't
> complete
> -------------------------------------------------------------------------------
>
> Key: HDFS-9092
> URL: https://issues.apache.org/jira/browse/HDFS-9092
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: nfs
> Affects Versions: 2.7.1
> Reporter: Yongjun Zhang
> Assignee: Yongjun Zhang
> Attachments: HDFS-9092.001.patch
>
>
> When NOT using 'sync' option, the NFS writes may issue the following warning:
> org.apache.hadoop.hdfs.nfs.nfs3.OpenFileCtx: Got an overlapping write
> (1248751616, 1249677312), nextOffset=1248752400. Silently drop it now
> and the size of data copied via NFS will stay at 1248752400.
> Found what happened is:
> 1. The write requests from client are sent asynchronously.
> 2. The NFS gateway has handler to handle the incoming requests by creating an
> internal write request structuire and put it into cache;
> 3. In parallel, a separate thread in NFS gateway takes requests out from the
> cache and writes the data to HDFS.
> The current offset is how much data has been written by the write thread in
> 3. The detection of overlapping write request happens in 2, but it only
> checks the write request against the curent offset, and trim the request if
> necessary. Because the write requests are sent asynchronously, if two
> requests are beyond the current offset, and they overlap, it's not detected
> and both are put into the cache. This cause the symptom reported in this case
> at step 3.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
