Like all good lessons, this one starts with a little lie and tells you the truth later on...
"Supriya : But how I will create different types of Chutneys?
Shekhar : Now you will see the missing phase of MapReduce — Shuffle phase. MapReduce will group all the outputs written by every Map based on the key. This will be automatically done for you. You can assume key as just a name of ingredient like Onion. So all the onion keys will be grouped together and will be transferred to a grinder which will just process onions. So, you will get onion Chutney. Similarly all the tomatoes will be transferred to the grinder marked for tomato and will produce tomato Chutney."
You see, women love cooking. The article is kind of hilarious in a chauvinistic sense, but it _does_ do a decent job of describing MapReduce. The only problem I see with simple overviews like this is that they leave out the fact that, often times, the reduce step is far more complicated than adding numbers together. This explanation pretty much applies for any basic framework for parallelism.
On a side note, I was surprised when MapReduce made such waves, as it's nothing _new_, per se. I still can't believe Google was granted a patent for it. I wrote "MapReduce"-esque algorithms in early high school when I made a distributed system to compute Pi. All it really does is provide a framework for creating, dispatching and joining jobs.
