I think it will help for Ouch to clarify what is meant by "in order". If one
JSON file must be completely processed before the next file starts, there is
not much point to using MapReduce at all, since your problem cannot be
partitioned. On the other hand, there may be ways around this, for example:
* Use the file timestamps as keys into whatever the target data store
is, so that JSON->records->insert can proceed in parallel.
* Perform a pass on the JSON files, generating "sequence numbers" based
on the file times. Then process MR jobs where the {Sequence,Filename} is the
tuple into mapper, and some stream of {Sequence,FileData} is on output. The
sort/shuffle by Sequence would present a time-ordered stream to the reducers.
Then if absolute order is needed, use a single reducer.
john
Caveat: I may not know what I am talking about ;-)
From: Mahesh Balija [mailto:balijamahesh....@gmail.com]
Sent: Tuesday, January 15, 2013 7:47 AM
To: user@hadoop.apache.org
Subject: Re: Hadoop execution sequence
As per the Mapreduce behavior, mapper will process all the input file(s) in
parallel. i.e., no order is guaranteed among the input files.
If you want to process each file separately and maintain the order then you
need to process each file separately (in an independent mapreduce job) so that
your client is responsible for processing individual file in order.
Best,
Mahesh Balija,
Calsoft Labs.
On Tue, Jan 15, 2013 at 7:55 PM, Panshul Whisper
<ouchwhis...@gmail.com<mailto:ouchwhis...@gmail.com>> wrote:
Hello,
I was wondering if hadoop performs the map reduce operations on the data in
maintaining he order or sequence of data in which it received the data.
I have a hadoop cluster that is receiving json files.. Which are processed and
then stored on base.
For correct calculation it is essential for the json files to be processed on
the order they are received. How can I make sure this happens.
Thanking you,
Regards,
Ouch Whisper
01010101010
