hadoop-common-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Owen O'Malley (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HADOOP-3227) Implement a binary input/output format for Streaming
Date Fri, 04 Jul 2008 06:25:36 GMT

    [ https://issues.apache.org/jira/browse/HADOOP-3227?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12610462#action_12610462

Owen O'Malley commented on HADOOP-3227:

I agree with Runping. For input, you *need* to know the record breaks, or you can't split
file between maps. This requires that \n be special and therefore quoted when in the non-record
boundary case. I think we should just do that quoting so that the streaming apps can just
echo their input to output and get an Identity map/reduce. Of course this all got a little
trickier now that you can change the delimiter from tab to an arbitrary string...

> Implement a binary input/output format for Streaming
> ----------------------------------------------------
>                 Key: HADOOP-3227
>                 URL: https://issues.apache.org/jira/browse/HADOOP-3227
>             Project: Hadoop Core
>          Issue Type: Improvement
>          Components: contrib/streaming
>            Reporter: Arun C Murthy
>            Assignee: Arun C Murthy
> Lots of streaming applications process textual data with 1 record per line and fields
separated by a delimiter. It turns out that there is no point in using any of Hadoop's input/output
formats since the streaming script/binary itself will parse the input and break into records
and fields. In such cases we should provide users with a binary input/output format which
just sends 64k (or so) blocks of data directly from HDFS to the streaming application.
> I did something very similar for Pig-Streaming (PIG-94 - BinaryStorage) which resulted
in 300%+ speedup for scanning (identity mapper & map-only jobs) data... the parsing done
by input/output formats in these cases were pure-overhead.

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message