Where is Sort used in MapReduce phase and why?

Question

I am new to hadoop here. It is not clear why we need to be able to sort by keys while using hadoop mapreduce ? After map phase, we need to distribute the data corresponding to e

Answer 1

The "sort by key" is best understood if we consider the fact that hadoop DISTRIBUTES processes for you by sending different keys to different machines. The basic (simplified) version of the idea is this :

The reducer which a (k,v) pair is sent to = k.hashCode()%num_of_machines. 

So, if my key's hashcode is 10, and I have 2 machines, the key will be sent to machine #0, for example.

Thus, the key will (first) gives us a simple way to distribute computation.

In addition to simplifying distribution of computation, keys give us a way to join records from disparate data files into a single cluster. This is how we can do things like word_count, for example.

In fact, if you are finding that you don't need keys --- you probably don't need hadoop either !

The classic example (word count):

In the hadoop "word count" example, we emit keys (one key = one word) with values (# times that word was seen in a segment of text). This allows a SINGLE reduce function to receive a SINGLE word, and thus add all the times it was seen, creating an accurate word count.

Thus, aggregation of keys are what allow the "map" phase to be distributed across multiple machines independently. Without aggregating keys to the same reducer, in the word count example, we might get several word counts for a given word, since there is no gaurantee that a single reducer would receive all word counts from all files.

Another example:

Now... lets say we have social security numbers as ids and we want to output an aggregation of personal data. Lets say we have 2 massive files.

ssn->name

ssn->shoe_size

In this case, we can leverage the power of key grouping, such that an individuals name and shoe size are BOTH sent to the SAME reduce function.

The reducer(2) will receive 2 records here:

ssn->name,shoe_size

The idea here is that when writing map/reduce jobs, you have to encode your "tuples" that are outputted in such a way that they can be joined together in a meaningful way, in the reduce phase. Any distributed computing environments will likely, at some point, need to combine records calculated in different nodes. Keys give us a convenient and scalable methodology for doing this.

So -- the fact that we are gauranteed that the SAME keys go to the SAME reducer function confirms that EACH reducer for this particular social secuirty number will receive ALL data associated with that number, allowing us to join and output data records which include ssn, name, and shoe size.

Conclusion

Without distributing by key, joining data in such a way would require painfully complex logic involving some kind of intermediary data storage / caching. Hadoop simply generalizes and abstracts the common need to "join" data results from parallel computations by using a familiar pardigm : keys and values.