Hadoop Combiners

In the last post and in the preceding one we saw how to write a MapReduce program for finding the top-n items of a data set. The difference between the two was that the first program (which we call basic) emitted to the reducers every single item read from input, while the second (which we call enhanced) made a partial computation and emitted only a subset of the input. The enhanced top-n optimizes network transmissions (the less the key-value pairs emitted, the less network is used for transmitting them from mapper to reducer) and reduces the number of keys shuffled and sorted; but this is obtained at the cost of rewriting of the mapper. 

If we look at the code of the mapper of the enhanced top-n , we can see that it implements the idea behind the reducer: it uses a Map for making a partial count of the words and emits every word only once; looking at the reducer's code, we see that it implements the same idea. If we could execute the code of the reducer of the basic top-n after the mapper has run on every machine (with its subset of data), we would obtain exactly the same result than rewriting the mapper as in the enhanced. This is exactly what Hadoop combiners do: they're executed just after the mapper on every machine for improving performance. For telling Hadoop which class to use as a combiner, we can use the Job.setCombinerClass() method. 

Caution: using the reducer as a combiner works only if the function we're computing is both commutative (a + b = b + a) and associative (a + (b + c) = (a + b) + c). 
Let's make an example. Suppose we're analyzing the traffic of a website and we have an input file with the number of visits per day like this (YYYYMMDD value):

20140401 100
20140331 1000
20140330 1300
20140329 5100
20140328 1200

We want to find which is the day with the highest number of visits. 
Let's say that we have two mappers; the first one receives the first three lines and the second receives the last two. If we write the mapper to emit every line, the reducer will evaluate something like this:

max(100, 1000, 1300, 5100, 1200) -> 5100

and the max is 5100. 
If we use the reducer as a combiner, the reducer will evaluate something like this:

max( max(100, 1000, 1300), max(5100, 1200)) -> max( 1300, 5100) -> 5100

because each of the two mapper will evaluate locally the max function. In this case the result will be 5100 as well, since the function we're evaluating (the max function) is both commutative and associative. 

Let's say that now we need to compute the average number of visits per day. If we write the mapper to emit every line of the input file, the reducer will evaluate this:

mean(100, 1000, 1300, 5100, 1200) -> 1740

which is 1740. 
If we use the reducer as a combiner, the reducer will evaluate something like this:

mean( mean(100, 1000, 1300), mean(5100, 1200)) -> mean( 800, 3150) -> 1975

because each of the two mapper will evaluate locally the max function. In this case the result will be 1975, which is obviously wrong. 

So, if we're computing a commutative and associative function and we want to improve the performance of our job, we can use our reducer as a combiner; if we want to improve performance but we're computing a function that is not commutative and associative, we have to rewrite the mapper or to write a new combiner from stratch.

from: http://andreaiacono.blogspot.com/2014/03/hadoop-combiners.html