[Hadoop-common-dev] Re: MapReduce Usage in Search Engines

Hello Yuhendar,I'll add as much as I can at a high level from what I have learned so far about map-reduce to answer your questions:
1) The goal behind map-reduce is to perform a distributed computation which breaks up a large computation intensive problem into smaller chunks and solve those individual chunks and finally combine the result, the problem in this case being search, in this problem you have a master node and a set of slave nodes, the master (or in the hadoop domain I believe its known as the name node) takes input from the client in the form of a job and forwards this job out to the slaves which go off and solve smaller pieces of the problem and return the results. The master then uses a combine approach to gather the results from all the slaves and present it back to the client. A more concrete example is the distributed grep problem which is a form of searching for a particular word (or document) in a huge dataset. Take a look at the hadoop examples or the hadoop webpage to learn more about this.
2) Google to my understanding is using their internal implementation of the general algorithm for mapreduce to store data in their datastore known as bigtable which is a multi-dimensional sorted map.

My 2 cents.Regards.

As my understanding, google may use mapred to build index, but won't use
mapred in the search phase.
Because search phase need to be low latency which is not mapred's feature.

MapReduce tends to be used for massive (re)indexing.
See http://search-lucene.com/?q=hadoop+mapreduce&fc_project=Solr&fc_project=Lucene
for how Lucene/Solr people are using MapReduce.

For example, in a recent project we used MapReduce (streaming with jruby,
actually) together with Solr (Embedded version, to be more precise) to speed up
indexing of a 20 GB index that used to take a couple of hours. Now it takes 7
minutes, because it's parallelized to Nth degree.


MapReduce can also be used for various Machine Learning data crunching, say for
query log analysis, for content analysis, for NLP, for building of better
relevance models for search, etc. etc. See http://mahout.apache.org .

Otis
----Sematext :: http://sematext.com/ :: Solr - Lucene - Nutch
Lucene ecosystem search :: http://search-lucene.com/