[
https://issues.apache.org/jira/browse/MAPREDUCE-5958?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jason Lowe reassigned MAPREDUCE-5958:
-------------------------------------
Assignee: Emilio Coppa
Thanks for the report and patch, Emilio. We just ran across this as well, and
my apologies for missing this earlier. Looks like this was caused by
MAPREDUCE-2264 when the getRawDataLength stuff was added.
Patch looks good overall, and I manually verified it fixes the issue with
compressed map outputs. However it would be nice to have a unit test to verify
this doesn't break again in the future. I see there's already unit tests in
TestMerger that were intended to test this, but the tests are ineffective
because they create no segments. The {{for (int i = 1; i < 1; i++)}} code in
the test ends up creating no segments so nothing is merged.
[~sandyr] [~devaraj.k], could you also take a look since you were familiar with
the changes in MAPREDUCE-2264?
> Wrong reduce task progress if map output is compressed
> ------------------------------------------------------
>
> Key: MAPREDUCE-5958
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-5958
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Affects Versions: 2.2.0, 2.3.0, 2.2.1, 2.4.0, 2.4.1
> Reporter: Emilio Coppa
> Assignee: Emilio Coppa
> Priority: Minor
> Labels: progress, reduce
> Fix For: 2.4.1
>
> Attachments: HADOOP-5958-v2.patch
>
>
> If the map output is compressed (_mapreduce.map.output.compress_ set to
> _true_) then the reduce task progress may be highly underestimated.
> In the reduce phase (but also in the merge phase), the progress of a reduce
> task is computed as the ratio between the number of processed bytes and the
> number of total bytes. But:
> - the number of total bytes is computed by summing up the uncompressed
> segment sizes (_Merger.Segment.getRawDataLength()_)
> - the number of processed bytes is computed by exploiting the position of the
> current _IFile.Reader_ (using _IFile.Reader.getPosition()_) but this may
> refer to the position in the underlying on disk file (which may be compressed)
> Thus, if the map outputs are compressed then the progress may be
> underestimated (e.g., only 1 map output ondisk file, the compressed file is
> 25% of its original size, then the reduce task progress during the reduce
> phase will range between 0 and 0.25 and then artificially jump to 1.0).
> Attached there is a patch: the number of processed bytes is now computed by
> exploiting _IFile.Reader.bytesRead_ (if the the reader is in memory, then
> _getPosition()_ already returns exactly this field).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
