Processing 1 Terabyte of Text in 7 Seconds without Hadoop

Interesting post from Silvius Rus from the Cluster Team at Quantcast.

He implemented a simple Sawzall program to process 1 TB of text in 7 seconds starting from disk and did not use Hadoop (though he mentions that Quantcast’s proprietary MapReduce cluster is loosely based on Hadoop).

He also made a very interesting design decision –  to drop the sort phase of MapReduce and run the Reducer concurrently with the Mapper.