Parallel running MapReduce on Hadoop - hadoop

Question

I'm trying to understand how the parallel processing works with Hadoop & MapReduce.
I understand how Map can be run in parallel but I don't understand how Reduce can. For example if I want to find the average of the following list:
COMPUTER | YEAR | RUNS
‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾
compA | 1989 | 20
compA | 1990 | 10
compB | 1991 | 300
Where compA & compB are two data nodes
If the average function in the Reduce is run on compA & compB and then the result of the two data noes is averaged it'll be wrong.

Answer 1

Multiple reducer tasks will not be spawned for all type of logic. For aggregation, averaging logic the data will run on a single reducer.
Mapper class gets data as key-value pair. It sends the output to reducer class. Before reaching the reducer, the keys will be sorted and the values corresponding to the same key will reach the same reducer. In this way, the results will be correct. If your requirement involves averaging the entire data, then the output from all the map tasks will be send to a single reducer.
Eg: I want to find the average number of vehicles sold per month over an year.
I have data in 12 files, each file contains the sales details corresponding to a month.
I have to write the following logic.
mapper class
Mapper will get every record as input.
Parse the number of sales and month details from every record.
Write a constant as key and the sales as value (eg: 2000 cars sold in january will be something like [sales,2000]. Similarly for 1000 sale in february will be something like [sales,1000]). I used the key sales so that all the values with the same key will reach the same reducer.
Sample Input data
jan.txt
car-sold,2000
feb.txt
cars-sold,1000
.....
.....
dec.txt
cars-sold,5000
Mapper output
[sales,2000]
[sales,1000]
....
....
[sales,5000]
Reducer Input
{sales,[1000,2000,.....,5000]}
Ie (key, list of values)