Are you allowed to change the order of the data in the output? If you want to calculate the
cr/dr indicator cumulative sum value, then it will easy if the business allow you to change
the order of your data group by CR/DR indicator in the output.
For example, you can do it very easy with the way I described in my original email if you
CAN change the output like following:
Txn ID Cr/Dr Indicator Amount CR cumulative Amount Dr Cumulative
Amount1001 CR 1000 1000
01004 CR 2000
3000 01002 DR
500 0 5001003
DR 1500 0
2000
As you can see, you have to group out your output by the Cr/Dr Indicator. If you want to keep
the original order, then it is hard, at least I cannot think a way in short time.
But if you allow to change the order of the output, then it is called cumulative sum with
grouping (in this case, it is group1 for CR, group 2 for DR).
1) In the mapper, omit your data by Cr/Dr indicator, which will group the data by CR/DR. So
all CR data will go to one reducer, then all DR data will go to one reducer.2) Besides grouping
the data, if you want the output sorted by the Amount (for example) in each group, then you
have to do the 2nd sorting. Google 2nd sort. Then for each group, the data arriving into each
reducer will be sorted by amount. Otherwise, if you don't need that sorting, then just ignore
the 2nd sorting.3) In each reducer, the data arriving should be already grouped. The default
partitioner for MR job is Hash Partitioner. Depending on the hashCode() return for 'CR' and
'DR', these 2 groups data could go to different reducers (assuming you are running with multi
reducers), or they could go to the same reducers. But even they are going to the same reducer,
they will be arrived into 2 groups. So the output of your reducers will be grouped, which
is sorted by the way.4) In your reducers, for the same group data, you will get an array of
values. For CR, you will get all the CR records in the array. What you need to do is to Iterating
your array, for every element, calculating the cumulative sum, and omit the cumulative sum
with the each record out.5) In the end, your output could be multi files, as each file generated
from one reducer. You can merge them into one file, or just leave them as that in the HDFS.6)
For best performance, if you have huge data, AND you know all your possible value for THE
Indicator, you may want to consider use your own custom Partitioner, instead of HashPartitioner.
What you want is like a RoundRobin distribution of your keys inside the available reducers,
instead of Random distribution by hash value(). Keep in mind that random distribution DOES
NOT work well if the distinct count of your keys is small enough.
Yong
Date: Fri, 5 Oct 2012 10:26:43 +0530
From: sarathchandra.josyam@algofusiontech.com
To: user@hadoop.apache.org
Subject: Re: Cumulative value using mapreduce
Thanks for all your responses. As
suggested will go through the documentation once again.
But just to clarify, this is not my first map-reduce program. I've
already written a map-reduce for our product which does filtering
and transformation of the financial data. This is a new
requirement we've got. I have also did the logic of calculating
the cumulative sums. But the output is not coming as desired and I
feel I'm not doing it right way and missing something. So thought
of taking a quick help from the mailing list.
As an example, say we have records as below -
Txn ID
Txn Date
Cr/Dr Indicator
Amount
1001
9/22/2012
CR
1000
1002
9/25/2012
DR
500
1003
10/1/2012
DR
1500
1004
10/4/2012
CR
2000
When this file passed the logic should append the below 2 columns
to the output for each record above -
CR Cumulative Amount
DR Cumulative Amount
1000
0
1000
500
1000
2000
3000
2000
Hope the problem is clear now. Please provide your suggestions on
the approach to the solution.
Regards,
Sarath.
On Friday 05 October 2012 02:51 AM, Bertrand Dechoux wrote:
I indeed didn't catch the cumulative sum part. Then I
guess it begs for what-is-often-called-a-secondary-sort, if you
want to compute different cumulative sums during the same job. It
can be more or less easy to implement depending on which
API/library/tool you are using. Ted comments on performance are
spot on.
Regards
Bertrand
On Thu, Oct 4, 2012 at 9:02 PM,
java8964 java8964 <java8964@hotmail.com> wrote:
I did the cumulative sum in the HIVE UDF, as one of the
project for my employer.
1) You need to decide the grouping elements for
your cumulative. For example, an account, a department
etc. In the mapper, combine these information as your
omit key.
2) If you don't have any grouping requirement, you
just want a cumulative sum for all your data, then
send all the data to one common key, so they will all
go to the same reducer.
3) When you calculate the cumulative sum, does the
output need to have a sorting order? If so, you need
to do the 2nd sorting, so the data will be sorted as
the order you want in the reducer.
4) In the reducer, just do the sum, omit every
value per original record (Not per key).
I will suggest you do this in the UDF of HIVE, as
it is much easy, if you can build a HIVE schema on top
of your data.
Yong
From: tdunning@maprtech.com
Date: Thu, 4 Oct 2012 18:52:09 +0100
Subject: Re: Cumulative value using mapreduce
To: user@hadoop.apache.org
Bertrand is almost right.
The only difference is that the original
poster asked about cumulative sum.
This can be done in reducer exactly as
Bertrand described except for two points that
make it different from word count:
a) you can't use a combiner
b) the output of the program is as large as
the input so it will have different
performance characteristics than aggregation
programs like wordcount.
Bertrand's key recommendation to go read a
book is the most important advice.
On Thu, Oct 4, 2012 at 5:20 PM, Bertrand
Dechoux <dechouxb@gmail.com>
wrote:
Hi,
It sounds like a
1) group information by account
2) compute sum per account
If that not the case, you should
precise a bit more about your context.
This computing looks like a small
variant of wordcount. If you do not know
how to do it, you should read books
about Hadoop MapReduce and/or online
tutorial. Yahoo's is old but still a
nice read to begin with : http://developer.yahoo.com/hadoop/tutorial/
Regards,
Bertrand
On Thu, Oct 4, 2012 at 3:58 PM,
Sarath <sarathchandra.josyam@algofusiontech.com>
wrote:
Hi,
I have a file which has some
financial transaction data. Each
transaction will have amount and
a credit/debit indicator.
I want to write a mapreduce
program which computes
cumulative credit & debit
amounts at each record
and append these values to the
record before dumping into the
output file.
Is this possible? How can I
achieve this? Where should i put
the logic of computing the
cumulative values?
Regards,
Sarath.
--
Bertrand Dechoux
--
Bertrand Dechoux
|