[Hadoop-common-user] Architectural question

My understanding is you have two sets of strings S1, and S2 and you want to mark all strings that
belong to both sets. If this is correct, then:

Mapper: for all strings K in Si (i is 1 or 2) emit: key K and value i.
Reducer: For key K, if the list of values includes both 1 and 2, you have a match, emit: K MATCH, else emit: K NO_MATCH (or nothing).

I assume that the load is not terribly unbalanced. The logic goes for intersection of any number of sets. Mark the members with their sets, reduce over them to see if they belong to every set.

Good luck.

The original poster said that there was no common key. Your suggestion
presupposes that such a key exists.

The original posting said - "The app does simple match every line of input data
with every line of persistent data." Hence the "key" should be replaced by a
String from the 10 GB store or a hash of it. Hence, we can match it with the
hash or String from the persistent Store.



________________________________
From: Ted Dunning <tdunning@maprtech.com>
To: common-user@hadoop.apache.org
Sent: Mon, 11 April, 2011 7:38:04 AM
Subject: Re: Architectural question

The original poster said that there was no common key. Your suggestion
presupposes that such a key exists.

That is how I interpreted it, but if by "simple" some other matching function
then the most obvious one is meant, then it still is possible to extend theText
class and overwrite the hashCode and equals functions to accommodate for
this new sort of equality.

There are no subtle ways to deal with quadratic problems like this. They
just don't scale.

Your suggestions are roughly on course. When matching 10GB against 50GB,
the choice of which input to use as input to the mapper depends a lot on how
much you can buffer in memory and how long such a buffer takes to build.

If you can't store the entire 10GB of data in memory at once, then consider
a program like this:

a) split the 50GB of data across as many mappers as you have using standard
methods

b) in the mapper, emit each record several times with keys of the form (i,
j) where i cycles through [0,n) and is incremented once for each record read
and j cycles through [0, m) and is incremented each time you emit a record.
Choose m so that 1/m of your 10GB data will fit in your reducers memory.
Choose n so that n x m is as large as your desired number of reducers.

c) in the reducer, you will get some key (i,j) and an iterator for a number
of records. Read the i-th segment of your 10GB data and compare each of
the records that the iterator gives you to that data. If you made n = 1 in
step (b), then you will have at most m-way parallelism in this step. If n
is large, however, your reducer may need to read the same segment of your
10GB data more than once. In such conditions you may want to sort the
records and remember which segment you have already read.

In general, though, as I mentioned this is not a scalable process and as
your data grows it is likely to become untenable.

If you can split your data into pieces and estimate which piece each record
should be matched to then you might be able to make the process more
scalable. Consider indexing techniques to do this rough targeting. For
instance, if you are trying to find the closes few strings based on edit
distance, you might be able to use n-grams to get approximate matches via a
text retrieval index. This can substantially reduce the cost of your
algorithm.

Dear oleksiy,

You can read and compare from file without fully loading the file into
memory on each map process. This will save memory on your cluster.
Otherwise, there is no quick solution as every other solution still
requires you to have the 10GB in memory (in pieces or in whole) for
every entry of the 50 GB file.

Daniel.