hadoop-hdfs-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From java8964 <java8...@hotmail.com>
Subject RE: Hadoop InputFormat - Processing large number of small files
Date Thu, 21 Aug 2014 18:16:27 GMT
If you want to use NLineInputFormat, and also want the individual file to be processed in the
map task which prefer to be on the same task node as data node,  you need to implement and
control that kind of logic by yourself.
Extend the NLineInputFormat, Override the getSplits() method, read the location of the HDFS
file, and assign the server of the same file/block.
Keep in mind that NLineInputFormat will send one file to one mapper. So you need to make sure
that the file size is less than one block, Split is on block level instead of file.
About detail how to control the Mapper execution location, you can refer this book:
Professional Hadoop Solution
Chapter 4 Customizing MapReduce Execution describes some examples how to do that.

Date: Thu, 21 Aug 2014 22:26:12 +0530
Subject: Re: Hadoop InputFormat - Processing large number of small files
From: rabmdu@gmail.com
To: user@hadoop.apache.org

This means that a file with names of all the files that need to be processed and is fed to
hadoop with NLineInputFormat? 
If this is the case, then, how can we ensure that map processes are scheduled in the node
where blocks containing the files are stored already?


On Thu, Aug 21, 2014 at 9:07 PM, Felix Chern <idryman@gmail.com> wrote:

If I were you, I’ll first generate a file with those file name:

hadoop fs -ls > term_file
Then run the normal map reduce job

On Aug 21, 2014, at 1:38 AM, rab ra <rabmdu@gmail.com> wrote:
Thanks for the link. If it is not required for CFinputformat to have contents of the files
in the map process but only the filename, what changes need to be done in the code?


On 20 Aug 2014 22:59, "Felix Chern" <idryman@gmail.com> wrote:

I wrote a post on how to use CombineInputFormat:http://www.idryman.org/blog/2013/09/22/process-small-files-on-hadoop-using-combinefileinputformat-1/

In the RecordReader constructor, you can get the context of which file you are reading in.In
my example, I created FileLineWritable to include the filename in the mapper input key.Then
you can use the input key as:


  public static class TestMapper extends Mapper<FileLineWritable, Text, Text, IntWritable>{

    private Text txt = new Text();

    private IntWritable count = new IntWritable(1);

    public void map (FileLineWritable key, Text val, Context context) throws IOException,

      StringTokenizer st = new StringTokenizer(val.toString());

        while (st.hasMoreTokens()){

          txt.set(key.fileName + st.nextToken());          

          context.write(txt, count);





On Aug 20, 2014, at 8:19 AM, rab ra <rabmdu@gmail.com> wrote:

Thanks for the response.
Yes, I know wholeFileInputFormat. But i am not sure filename comes to map process either as
key or value. But, I think this file format reads the contents of the file. I wish to have
a inputformat that just gives filename or list of filenames.

Also, files are very small. The wholeFileInputFormat spans one map process per file and thus
results huge number of map processes. I wish to span a single map process per group of files.

I think I need to tweak CombineFileInputFormat's recordreader() so that it does not read the
entire file but just filename.



On Wed, Aug 20, 2014 at 6:48 PM, Shahab Yunus <shahab.yunus@gmail.com> wrote:

Have you looked at the WholeFileInputFormat implementations? There are quite a few if search
for them...




On Wed, Aug 20, 2014 at 1:46 AM, rab ra <rabmdu@gmail.com> wrote:


I have a use case wherein i need to process huge set of files stored in HDFS. Those files
are non-splittable and they need to be processed as a whole. Here, I have the following question
for which I need answers to proceed further in this.

1.  I wish to schedule the map process in task tracker where data is already available. How
can I do it? Currently, I have a file that contains list of filenames. Each map get one line
of it via NLineInputFormat. The map process then accesses the file via FSDataInputStream and
work with it. Is there a way to ensure this map process is running on the node where the file
is available?. 

2.  Since the files are not large and it can be called as 'small' files by hadoop standard.
Now, I came across CombineFileInputFormat that can process more than one file in a single
map process.  What I need here is a format that can process more than one files in a single
map but does not have to read the files, and either in key or value, it has the filenames.
In map process then, I can run a loop to process these files. Any help?

3. Any othe alternatives?



View raw message