[
https://issues.apache.org/jira/browse/HADOOP-5539?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12688544#action_12688544
]
Billy Pearson edited comment on HADOOP-5539 at 3/23/09 8:14 PM:
----------------------------------------------------------------
Someone can use my patch as a starting point.
The ReduceTask.java call that is the problem is line 2145
The maptask.java call that is the problem is line 1269
I use streaming without a combiner so that should be looked at also to see if
it uses o.a.h.mapred.Merger
the basic problem is the codec is not passed from these function to the merger
so its always null the call to
o.a.h.mapred.Merger should include codec somehow if compression is not used
then codec is null
in both ReduceTask and Maptask.
I thank this is a major bug that effects all MR jobs with disk bandwidth that
uses compression.
was (Author: viper799):
someone can use my patch as a starting point.
the ReduceTask.java call that is the problem is line 2145
the maptask.java call that is the problem is line 1269
I use streaming without a combiner so that should be looked at also to see if
it uses o.a.h.mapred.Merger
> o.a.h.mapred.Merger not maintaining map out compression on intermediate files
> -----------------------------------------------------------------------------
>
> Key: HADOOP-5539
> URL: https://issues.apache.org/jira/browse/HADOOP-5539
> Project: Hadoop Core
> Issue Type: Bug
> Components: mapred
> Affects Versions: 0.19.1
> Environment: 0.19.2-dev, r753365
> Reporter: Billy Pearson
> Assignee: Billy Pearson
> Priority: Blocker
> Fix For: 0.19.2, 0.20.0
>
> Attachments: 5539.patch
>
>
> hadoop-site.xml :
> mapred.compress.map.output = true
> map output files are compressed but when the in memory merger closes
> on the reduce the on disk merger runs to reduce input files to <=
> io.sort.factor if needed.
> when this happens it outputs files called intermediate.x files these
> do not maintain compression setting the writer (o.a.h.mapred.Merger.class
> line 432)
> passes the codec but I added some logging and its always null map output
> compression set true or false.
> This causes task to fail if they can not hold the uncompressed size of the
> data of the reduce its holding
> I thank this is just and oversight of the codec not getting set correctly for
> the on disk merges.
> {code}
> 2009-03-20 01:30:30,005 INFO org.apache.hadoop.mapred.Merger: Merging 30
> intermediate segments out of a total of 3000
> 2009-03-20 01:30:30,005 INFO org.apache.hadoop.mapred.Merger: intermediate.1
> used codec: null
> {code}
> I added
> {code}
> // added my me
> if (codec != null){
> LOG.info("intermediate." + passNo + " used codec: " +
> codec.toString());
> } else {
> LOG.info("intermediate." + passNo + " used codec: Null");
> }
> // end added by me
> {code}
> Just before the creation of the writer o.a.h.mapred.Merger.class line 432
> and it outputs the second line above.
> I have confirmed this with the logging and I have looked at the files on the
> disk of the tasktracker. I can read the data in
> the intermediate files clearly telling me that there not compressed but I can
> not read the map.out files direct from the map output
> telling me the compression is working on the map end but not on the on disk
> merge that produces the intermediate.
> I can see no benefit for these not maintaining the compression setting and as
> it looks they where intended to maintain it.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
