[
https://issues.apache.org/jira/browse/MAPREDUCE-1752?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12927178#action_12927178
]
Patrick Kling commented on MAPREDUCE-1752:
------------------------------------------
There is something really strange about the semantics of the offsets and
lengths returned by this. Consider the following part file consisting of 3
blocks containing a file f starting at offset 896 with length 512:
{code}
+---------------+
| ... |
+---------------+
0
+-----------+---+
| ... | f |
+-----------+---+
512 896
+-----------+---+
| f |...|
+-----------+---+
1024 1408
{code}
Calling getFileBlockLocations on this file will return 2 LocatedBlocks:
b1=<offset=0, length=512>, b2=<offset=512, length=512>. This indicates that b1
contains the first 512 bytes of the block, even though in fact it only contains
the first 128 bytes. This is a problem when the client uses these LocatedBlocks
to detect whether a portion of f has been corrupted.
I can think of 2 possible ways of fixing this:
1) Fix the offset of the returned blocks by subtracting hstatus.getStartIndex()
(i.e., the offset of f in the part file) from the block offset. This would
return b1=<offset=-384, length=512> and b2=<offset=128, length=512>, indicating
to the client that the first 384 bytes of b1 are not part of 1 and correctly
indicating the length of each block. In a way, this is similar to how
FSNamesystem.getBlockLocations returns entire blocks even if the caller asks
for a range that covers only part of these blocks.
2) Fix the length on the first block returned to reflect the portion of f that
is contained in this block, i.e., return b1=<offset=128, length=128>,
b2=<offset=128, length=512>. This seems somewhat less clean to me but avoids
negative offsets. Also, it would break the convention that all blocks of a file
with the exception of the last block are the same length.
> Implement getFileBlockLocations in HarFilesystem
> ------------------------------------------------
>
> Key: MAPREDUCE-1752
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-1752
> Project: Hadoop Map/Reduce
> Issue Type: Improvement
> Components: harchive
> Reporter: Dmytro Molkov
> Assignee: Dmytro Molkov
> Fix For: 0.22.0
>
> Attachments: MAPREDUCE-1752.2.patch, MR-1752.patch
>
>
> To efficiently run map reduce on the data that has been HAR'ed it will be
> great to actually implement getFileBlockLocations for a given filename.
> This way the JobTracker will have information about data locality and will
> schedule tasks appropriately.
> I believe the overhead introduced by doing lookups in the index files can be
> smaller than that of copying data over the wire.
> Will upload the patch shortly, but would love to get some feedback on this.
> And any ideas on how to test it are very welcome.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
