Home | About | Sematext search-lucene.com search-hadoop.com
NEW: Monitor These Apps!
elasticsearch, apache solr, apache hbase, hadoop, redis, casssandra, amazon cloudwatch, mysql, memcached, apache kafka, apache zookeeper, apache storm, ubuntu, centOS, red hat, debian, puppet labs, java, senseiDB
 Search Hadoop and all its subprojects:

Switch to Plain View
MapReduce >> mail # user >> Re: produce a large sequencefile (1TB)


+
Jerry Lam 2013-08-20, 02:25
+
Bing Jiang 2013-08-20, 02:55
+
Harsh J 2013-08-20, 05:38
Copy link to this message
-
Re: produce a large sequencefile (1TB)
Hi Harsh,

Thank you for the reply. It really answers my question and provides
practical advices.

Best Regards,

Jerry
On Tue, Aug 20, 2013 at 1:38 AM, Harsh J <[EMAIL PROTECTED]> wrote:

> Unfortunately given the way Reducers work today you wouldn't be able
> to do this. They are designed to fetch all data before the merge, sort
> and process it through the reducer implementation. For that to work,
> as you've yourself deduced, you will need as much space locally
> available.
>
> What you could do however, is perhaps just run a Map-only job, let it
> produce smaller files, then run a non-MR java app that reads them all
> one by one, and appends to a single HDFS SequenceFile. This is like a
> reducer, but minus a local sort phase. If the sort is important to you
> as well, then your tweaking will have to go further into using
> multiple reducers with Total Order Partitioning, and then running this
> external java app.
>
> On Tue, Aug 20, 2013 at 8:25 AM, Bing Jiang <[EMAIL PROTECTED]>
> wrote:
> > Hi Jerry,
> >
> > I think whether it is acceptable to set multiple reducers to generate
> more
> > MapFile(IndexFile, DataFile)s.
> >
> > I want to know the real difficulties of multiply reducer to
> post-processing.
> > Maybe there are some questions about app?
> >
> >
> >
> > 2013/8/20 Jerry Lam <[EMAIL PROTECTED]>
> >>
> >> Hi Bing,
> >>
> >> you are correct. The local storage does not have enough capacity to hold
> >> the temporary files generated by the mappers. Since we want a single
> >> sequence file at the end, we are forced to use 1 reducer.
> >>
> >> The use case is that we want to generate an index for the 1TB sequence
> >> file that we can randomly access each row in the sequence file. In
> practice,
> >> this is simply a MapFile.
> >>
> >> Any idea how to resolve this dilemma is greatly appreciated.
> >>
> >> Jerry
> >>
> >>
> >>
> >> On Mon, Aug 19, 2013 at 8:14 PM, Bing Jiang <[EMAIL PROTECTED]>
> >> wrote:
> >>>
> >>> hi,Jerry.
> >>> I think you are worrying about the volumn of mapreduce local file, but
> >>> would  you give us more details about your apps.
> >>>
> >>> On Aug 20, 2013 6:09 AM, "Jerry Lam" <[EMAIL PROTECTED]> wrote:
> >>>>
> >>>> Hi Hadoop users and developers,
> >>>>
> >>>> I have a use case that I need produce a large sequence file of 1 TB in
> >>>> size when each datanode has  200GB of storage but I have 30 datanodes.
> >>>>
> >>>> The problem is that no single reducer can hold 1TB of data during the
> >>>> reduce phase to generate a single sequence file even I use aggressive
> >>>> compression. Any datanode will run out of space since this is a single
> >>>> reducer job.
> >>>>
> >>>> Any comment and help is appreciated.
> >>>>
> >>>> Jerry
> >>
> >>
> >
> >
> >
> > --
> > Bing Jiang
> > Tel:(86)134-2619-1361
> > weibo: http://weibo.com/jiangbinglover
> > BLOG: www.binospace.com
> > BLOG: http://blog.sina.com.cn/jiangbinglover
> > Focus on distributed computing, HDFS/HBase
>
>
>
> --
> Harsh J
>
NEW: Monitor These Apps!
elasticsearch, apache solr, apache hbase, hadoop, redis, casssandra, amazon cloudwatch, mysql, memcached, apache kafka, apache zookeeper, apache storm, ubuntu, centOS, red hat, debian, puppet labs, java, senseiDB