[Hadoop-common-user] Processing xml files

Hello,

We have the same type of data, we currently convert it to tab delimited file and use it as input for streaming

Regards,
Aleksandr

--- On Tue, 5/24/11, Mohit Anchlia wrote:

From: Mohit Anchlia <mohitanchlia@gmail.com>
Subject: Processing xml files
To: common-user@hadoop.apache.org
Date: Tuesday, May 24, 2011, 4:16 PM

I just started learning hadoop and got done with wordcount mapreduce
example. I also briefly looked at hadoop streaming.

Some questions
1) What should  be my first step now? Are there more examples
somewhere that I can try out?
2) Second question is around pracitcal usability using xml files. Our
xml files are not big they are around 120k in size but hadoop is
really meant for big files so how do I go about processing these xml
files?
3) Are there any samples or advise on how to processing with xml files?


Looking for help and pointers.

Can you please give more info?
Do you append multiple xml files data as a line into one file? Or
someother way? If so then how big do you let files to be.
how do you create these files assuming your xml is stored somewhere
else in the DB or filesystem? read them one by one?
what are your experiences using text files instead of xml?
Reason why xml files can't be directly used in hadoop or shouldn't be used?
Any performance implications?
Any readings suggested in this area?

Our xml is something like:

<column id="Name" security="sensitive" xsi:type="Text">
<value>free a last</value>
</column>
<column id="age" security="no" xsi:type="Text">
<value>40</value>
</column>

And we would for eg want to know how many customers above certain age
or certain age with certain income etc.

Sorry for all the questions. I am new and trying to get a grasp and
also learn how would I actually solve our use case.

Can you please give more info?
Do you append multiple xml files data as a line into one file? Or
someother way? If so then how big do you let files to be.


We currently feed to our process folder with converted files. We don't size it any way we let hadoop to handle it.

how do you create these files assuming your xml is stored somewhere
else in the DB or filesystem? read them one by one?


what are your experiences using text files instead of xml?
If you are using streaming job it is easier to build your logic if you have one file, you can actually try to parse xml in your mapper and convert it for reducer but, why you just don't write small app which will convert it?

Reason why xml files can't be directly used in hadoop or shouldn't be used?
Any performance implications?
If you are using Pig there is XML reader http://pig.apache.org/docs/r0.8.1/api/org/apache/pig/piggybank/storage/XMLLoader.html

If you have well define schema it is easier to work with big data :)

Any readings suggested in this area?
Try look into Pig it has lots of useful stuff, which will make your experience with hadoop nicer

Our xml  is something like:

<column id="Name" security="sensitive" xsi:type="Text">
<value>free a last</value>
</column>
<column id="age" security="no" xsi:type="Text">
<value>40</value>
</column>

And we would for eg want to know how many customers above certain age
or certain age with certain income etc.

Hadoop has build in counter, did you look into word count example from hadoop?


Regards,
Aleksandr

--- On Tue, 5/24/11, Mohit Anchlia wrote:

From: Mohit Anchlia <mohitanchlia@gmail.com>
Subject: Re: Processing xml files
To: common-user@hadoop.apache.org
Date: Tuesday, May 24, 2011, 4:41 PM
Can you please give more info?
Do you append multiple xml files data as a line into one file? Or
someother way? If so then how big do you let files to be.
how do you create these files assuming your xml is stored somewhere
else in the DB or filesystem? read them one by one?
what are your experiences using text files instead of xml?
Reason why xml files can't be directly used in hadoop or shouldn't be used?
Any performance implications?
Any readings suggested in this area?

Our xml  is something like:

<column id="Name" security="sensitive" xsi:type="Text">
<value>free a last</value>
</column>
<column id="age" security="no" xsi:type="Text">
<value>40</value>
</column>

And we would for eg want to know how many customers above certain age
or certain age with certain income etc.

Sorry for all the questions. I am new and trying to get a grasp and
also learn how would I actually solve our use case.

Thanks some more questions :)
Do you convert it into a flat file and write it to HDFS? Do you write
all the files to the same directory in DFS or do you group directories
based on days for eg? So like 2011/01/01 contains 10 files. store
results of 10 files somewhere and then on 2011/02/02 store another say
20 files. Now analyze 20 files and use the results from 10 files to do
the aggregation. If so then how do you do it. Or how should I do it
since it will be overhead processing those files again.

Please point me to examples so that you don't have to teach me Hadoop
or pig processing :)
Didn't think about it. I was just thinking in terms of using big
files. So when using small files hadoop will automatically distribute
the files accross cluster I am assuming based on some hashing.
Which one is better? Converting files to flat files or using xml as
is? How do I make that decision?

I will download pig tutorial and see how that works. Is there any
other xml related examples you can point me to?

Thanks a lot!

Hello,

We currently have complicated process which has more then 20 jobs piped to each other.
We are using shell script to control the flow, I saw some other company they were using spring batch. We use pig, streaming and hive

Not one thing if you are using ec2 for your jobs all local files need to be stored in /mnt Currently our cluster is organized this way in hdfs, and we process our data hourly and rotate final result to the beginning of pipeline for next process. Each process output is next process input, so we keep all data for current execution in the same dated folder so if you run daily it will be eg 20111212 if hourly 201112121416, and add subfolder for each subprocess into it
Example
/user/{domain}/{date}/input
/user/{domain}/{date}/process1
/user/{domain}/{date}/process2
/user/{domain}/{date}/process3
/user/{domain}/{date}/process4
our process1 takes input for current converted files and output from last process.
After we start the job we load converted files into input location and move them out from local space so we will not reprocess them.

Not sure if there is examples, this all will depend on architecture of the project you are doing, I bet if you will put all you need to do on whiteboard you will find best folder structure for yourself :)

Regards,
Aleksandr


--- On Tue, 5/24/11, Mohit Anchlia wrote:

From: Mohit Anchlia <mohitanchlia@gmail.com>
Subject: Re: Processing xml files
To: common-user@hadoop.apache.org
Date: Tuesday, May 24, 2011, 5:20 PM

Thanks some more questions :)
Do you convert it into a flat file and write it to HDFS? Do you write
all the files to the same directory in DFS or do you group directories
based on days for eg? So like 2011/01/01 contains 10 files. store
results of 10 files somewhere and then on 2011/02/02 store another say
20 files. Now analyze 20 files and use the results from 10 files to do
the aggregation. If so then how do you do it. Or how should I do it
since it will be overhead processing those files again.

Please point me to examples so that you don't have to teach me Hadoop
or pig processing :)
Didn't think about it. I was just thinking in terms of using big
files. So when using small files hadoop will automatically distribute
the files accross cluster I am assuming based on some hashing.
Which one is better? Converting files to flat files or using xml as
is? How do I make that decision?

I will download pig tutorial and see how that works. Is there any
other xml related examples you can point me to?

Thanks a lot!

Thanks for the info! Is there a way to point me to an example where
output of one process is then piped to other. For eg: Results of day 1
is stored in the day1/output directory then how do I input these
results to day2's job? I am assuming these results will directly go to
the reducer since it has already gone through map-reduce cycle. Just
looking for some example to get little more feel of how to do it. So
far I have installed hadoop and gone through basic hadoop tutorial of
word count but I still lack knowledge of some important features.