Perhaps finetuning it for EC2, maybe even hosting the dataset there? I can see how this can be very useful! Otherwise... well... It seems like Hadoop gives you a lot of overhead, and it is just not practical to do parsing this way.  

With a straightforward implementation in Python, on a single Core2 Duo you can parse the dump (7z), compute diffs, md5, etc and store everything into a binary form in about 6-7 days.
For example an implementation here: http://code.google.com/p/pymwdat/  can do exactly that. I imagine that with faster C++ code and with modern i7 box it can be done within a day.
And after that this precomputed binary form (diffs+metadata+stats take about several times of the .7z dump ~ 100Gb) can be serialized very efficiently (just about an hour on a single box).

Saying that, I still think using Hadoop/EC2 could be really nice. Particularly if the dump can be made available on the S3/EC2.

-- Best, Dmitry


On Wed, Aug 17, 2011 at 3:07 PM, Diederik van Liere <dvanliere@gmail.com> wrote:
So the 14 day task included xml parsing and creating diffs. We might gain performance improvements by fine-tuning the Hadoop configuration although that seems to be more of  an art than science.
Diederik


On Wed, Aug 17, 2011 at 5:28 PM, Dmitry Chichkov <dchichkov@gmail.com> wrote:
Hello,

This is an excellent news! 

Have you tried running it on Amazon EC2? It would be really nice to know how well WikiHadoop scale up with the number of nodes.
Also, this timing - '3 x Quad Core / 14 days / full wikipedia dump", on what kind of task (xml parsing, diffs, md5, etc?) was it obtained?

-- Best, Dmitry

On Wed, Aug 17, 2011 at 9:58 AM, Diederik van Liere <dvanliere@gmail.com> wrote:
Hello!

Over the last few weeks, Yusuke Matsubara, Shawn Walker, Aaron Halfaker and Fabian Kaelin (who are all Summer of Research fellows)[0] have worked hard on a customized stream-based InputFormatReader that allows parsing of both bz2 compressed and uncompressed files of the full Wikipedia dump (dump file with the complete edit histories) using Hadoop. Prior to WikiHadoop and the accompanying InputFormatReader it was not possible to use Hadoop to analyze the full Wikipedia dump files (see the detailed tutorial / background for an explanation why that was not possible). 

This means:
1) We can now harness Hadoop's distributed computing capabilities in analyzing the full dump files.
2) You can send either one or two revisions to a single mapper so it's possible to diff two revisions and see what content has been addded / removed. 
3) You can exclude namespaces by supplying a regular expression. 
4) We are using Hadoop's Streaming interface which means people can use this InputFormat Reader using different languages such as Java, Python, Ruby and PHP.

The source code is available at: https://github.com/whym/wikihadoop
A more detailed tutorial and installation guide is available at: https://github.com/whym/wikihadoop/wiki


(Apologies for cross-posting to wikitech-l and wiki-research-l)
Best,

Diederik


_______________________________________________
Wiki-research-l mailing list
Wiki-research-l@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/wiki-research-l



_______________________________________________
Wiki-research-l mailing list
Wiki-research-l@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/wiki-research-l




--
<a href="http://about.me/diederik">Check out my about.me profile!</a>

_______________________________________________
Wiki-research-l mailing list
Wiki-research-l@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/wiki-research-l