Big Data and Analytics Summit 2014

Big Data and Analytics Summit 2014 – NASSCOM

When: 27^th June 2014, 9:00 AM – 5:15 PM

Where: Hotel Trident, Hyderabad

Sponsored By: Accenture, CapGemini, Paypal, Airtel, EXL, Virtusa, Genpact, Contact Singapore, Target

Highlights:

1.       The Nasscom event was attended by thousands of analysts, architects, data scientists and managers across India. It provided a platform to interact with the market leaders like Mu Sigma, Amazon, P&G, Target, Paypal, Indix, Simplify360 and many more. It was a full day event with lots of presentations, tech talks, panel discussions, Q&A sessions, and product demos.

2.       There were many stalls at the event.  For instance, at the Accenture stall were two products exhibited – one was for automobile insurance the other one was for video analytics using Google glass. A car will be fitted with sensors that measure the speed, friction, wear and tear of the vehicle. These recordings are used to calculate the risk in real-time. Insurance premium is determined by the risk; for a safely driven car the premium is lower.

3.       Marc Chemin from CapGemini spoke about the advent of Big Data in Europe.

·         Big data is used in 3 areas – enhancement (e.g. to boost sales), disruption & new business.

·         With the advent of Big Data and Analytics, a lot of Public agencies are being replaced by Private agencies. The public agencies work in silos and don’t usually share any public data. However very intelligent systems can be built by sharing data, and winning people’s confidence.

·         Lastly it was discussion around Data privacy in Europe. There are 29 authorities in Europe and each of them defines their own privacy policies around data regulation. Hence it gets tough to execute some of the Big Data projects.

4.       Michael Barnes from Forrester Research provided a lot of insights to Analytics market.

·         In the earlier days data was used as a system of record, now it is used as system of engagement to touch people. More and more companies are coming out and investing in Business Technologies to provide better customer experience.

·         In the earlier days data was scarce, expensive, easy to manage, neat-rich-structured, manageable number of formats and structures. With Big data we have plenty of data, cheap, impossible to manage, messy & detailed, nearly infinite in format & structure.

·         Getting the right & sufficient data, skills, technology and experience are key requirements to work in Analytics.

·         Currently Big Data is used for real-time reporting. In future it will be used for process automation, adhoc decision support, operational planning, and strategy planning.

·         Michael gave this wonderful definition – “Big Data is not a technology, a project, an objective with a single compute. It is a journey, process, strategic effect. The practices & technologies that close the gap between the data available and able to turn data into business insights”.

5.       This was followed by presentation by Michael Svilar from Accenture Analytics. Some of the projects that they are executing – Connected Cars, Network Analytics, Safe City & Smart Water (to stop wastage of water in UK using Big Data). The statistics shows that in India 94% of the companies that use Big Data are satisfied with the results. Out of this, 65% are using to analyze customer behavior.

6.       The real crowd puller was by Dheeraj Rajaram, founder of one of the most successful startup in Analytics – Mu Sigma. He discussed some of the innovations in Analytics.

·         Innovation 1 – How some of the companies are using Social Score to sanction loans. They analyze how people are spending on and on what they are spending, by analyzing their Facebook posts.

·         Innovation 2 – Real-time video analytics used to determine possibility of fraud in betting industry.

·         Innovation 3 – Field engineers constantly face the risk of their lives. Big Data is being used to put an end to this. By 2015, 6 Billion devices will be connected.

7.       The next session was on Fraud Detection & Risk Management in online and offline space.

·         Online ecommerce companies constantly face the issue of Fraud & Risk (Cash On Delivery - customer might not pay for the item). Money-back guarantee is another big blow for these companies with the U.S dealing with 2.3 b$ worth of returns annually. Amazon quoted about this issue saying, companies need to be customer centric and always start with the customer in mind & work backwards. Consider a scenario where a lady purchases a lot of baby accessories and while billing she forgot to bill a baby wipe which was at the bottom of the basket. Instead of treating it as a theft, the store can be polite in asking her if she would like to bill anything else.

·         In the online space when a customer makes a purchase, all that the company has to predict risk is 50 milli second. In an offline mode user has an identity. However in the online mode, all that the company knows is the customer’s email id. Within 50 ms, 1000s of rules and predictive models have to be executed and validated.

·         In the online space even a tiniest loophole can be disastrous in lightning viral environment. So companies cannot create any space for mistake.

·         Do not just rely on a supervised learning method for your business. You would need unsupervised learning, random forest, neural networks, decision trees and many more.

·         One instance of fraud was discussed by the panel. There were 20 banks in Somalia and in a span of few weeks, 15 of the banks were mugged. The Somalia government approached the analysts to predict when the remaining 5 banks will be mugged, using statistical analysis.

8.       The panel discussion by the startups was the most motivating of all the talks. Bhupendra from Simplify360 & Sridhar Venkatesh from Indix kept the whole crowd on the toes. They spoke about their journey from nowhere to becoming the most promising startup.  Bhupendra shared his experiences of starting a product with no roadmap in 2009, when there was a global recession.  And how the big surprises awaited him at the end, to become the most influential technologist in India. Dedication, Perseverance, Innovation and Team play a big role in startups.

9.       James from the Boston Consulting Group gave the concluding talk and shared light on a lot of key areas –

·         Myth1 – Bigger Data is Better Data.  Actual fact - Size doesn’t matter

·         Myth2 – It’s all about correlation. Actual fact – Causation remains the key

·         Myth3 – Focus on new ideas. Actual fact – Often do the same things better

·         Myth4 – Requires big investments. Actual fact – More capability much cheaper

·         Myth5 – It’s only for advanced firms. Actual fact – All businesses can benefit.

10.   Check out some of the stalls put up by a few cool startups - Veda, Nanobi, Germin8, and DataWeave. One thing in common in all these startups was the factor of Social, Sentiment, and Semantic aspects. The market looks very promising with all these startups providing a hint towards it.

My No SQL - Big Data Presentation to Wellington - NZ

Raju Rama Krishna
Wellington, NZ
17th May
Topics:- High Scalability, MongoDB, Redis

WJUG 2013-04-17: Raju Ramakrishna - NoSQL Family from John Hurst on Vimeo.

Neo4j Graph Case Study - Facebook & Trip Planner

NoSQL Graph Database - Neo4j

Graphs are everywhere - social networks, routing algorithms, payment flow, and many more. Facebook friends, Twitter followers, LinkedIn graph are a few scenarios for Graph. Quite recently Facebook released Social Graph search. Neo4j is an open source NOSQL Graph database, immensely popular and being used by companies like Adobe, Intuit. Twitter uses FlockDB for the Graphs. 

Download and extract Neo4j Community edition from http://www.neo4j.org/download. Start the server by running /bin/Neo4j.bat. The Neo4j monitoring tool can be accessed from http://localhost:7474/webadmin/. The monitoring tool provides several utilities as shown below. 

Gremlin - A Groovy tool for querying Graph data

Neo4j Monitoring console

Neo4j Data Browser

Apart from the above utilities, Neo4j provides support for Graph Query language called Cypher. Since Ne04j is built on Java, it runs on JVM and can be used as a Embedded database. Or it could be used as a standalone database, through REST api. There is another interesting tool called Neoclipse which is very rich in visualizing Graphs. The following Graphs for facebook friends and railway routes are generated with Neoclipse.

Applications of Neo4j

1. Facebook friend connect

• Login to Facebook and follow the instructions from here to download the Facebook Graph data.
• Next, run the following program to download the Profile photos from Facebook to your local machine. The photos are used only for visualization purpose for neoclipse. This step uses the data downloaded from the previous step. change the "file" variable.
• The next step is to import the graph data to Neo4j. Run the below program. I have used the data for two Facebook profiles and loaded all the friends and interconnection data.
• Now the data is ready, I implemented a search utility which - given 2 Facebook IDs, it will provide all posible ways for the 2 people to get connected. 

2. Railway Trip Planner

• For this application, I used data from Indian Railways site. A few trains were randomly chosen and their routes and information were stored in local file system as a tab-delimited text. The route information contains station code, names, distances, time taken etc.  I created directories with name as train number. Each directory contains info.csv and schedule.csv. Download it and save the directories and its contents locally. It can be downloaded from here
• Next, run the below program to import the graph data to Neo4j.
• Now the data is available, I created a trip planner. It provides 2 options - Search for a path between 2 stations with least distance, Search for a path between 2 stations with least number of stops. In each case print the total distance. 
• The output of the above search program is

3. Facebook Graph Search

• Facebook Graph Search is released. It provides search based on your friends connections, likes and recommendations. So I could ask it for "Suggest me Indian Vegetarian restaurants in and around South London". There is a wonderful article on how to build this with Neo4j at http://jimwebber.org/2013/02/neo4j-facebook-graphsearch-for-the-rest-of-us/
• I implemented the code based on the above design.  First build the Graph <
• Next run search on restaurants

High Availability

Neo4j enterprise edition provides high availability in the form of replication and shard caching. I tried this on a 63 Million Node relationships with a cluster of 3 nodes. Planning to write up in next series....

Real-Time Data Analysis using Twitter Stream

Experiment

Twitter processes half a billion tweets per day. The size of a tweet is max 140 characters. For experimentation purpose, Twitter provides these tweets to the public. Initially they were providing it through REST API and then moved onto OAuth. Twitter Streaming API makes this possible, and provides real-time streaming tweets at 1% of the Twitter's load. So that means, on an average Twitter receives 5787  ( 500000000 Tweets / (24 Hours * 60 Mins * 60 Secs) ) Tweets per second and the Streaming API sends us 58 Tweets per second. And this data is arriving constantly, so we need a mechanism to store 58 tweets every second and run real-time analytics on them. Even though each Tweet is 140 characters or 280 Bytes ( Char is 2 bytes ), the Streaming-API sends us a lot of information for each Tweet (1 Tweet = 4 KB). This information is sent in the JSON format.

The Twitter data provides a very valuable tool in the field of marketing. Vendors can do sentiment analysis using Tweets for a specific hashTag. So if Samsung wants to know how content people are about their products, they can do so with the Tweet data. As a result a lot of NLP (Natural Language Processing) field researches have started. Apart from this, we can do a lot of machine learning tasks on these Tweets.

As part of this experiment I implemented a Consumer to read the stream of JSON tweets and persist in MongoDB. Since its a write-heavy application, I load-balanced (Sharded) my MongoDB. This application keeps running forever. Now the data starts filling up my MongoDB clusters. To keep the storage minimum, I extracted and stored only TweetID, Tweeter Name, Actual Tweet, the Source of the Tweet ( eg. twitter.com, facebook.com, blackberry, etc). Then I setup a MongoDB incremental Map-Reduce job to run every 10 minutes. This job gets the statistics of unique sources and their counts. From this I generated the top 10 statistics and create chart using JFreeChart. 

Architectural Overview

Execution

Setup Twitter Account

Goto https://dev.twitter.com/apps and click "Create a new application".

Fill up all mandatory information and submit the application

Goto the tab "OAuth Tool" and note down the Consumer Key and Consumer Secret.

Run the following program by changing the consumer key and consumer secret

Follow the instructions of the program to generate the Access Token and Access Token Secret.

It can later be obtained from the "OAuth Tool" tab.

Setup MongoDB

MongoDB provides sharding capability on database/collections. So I setup a simple MongoDB sharding setup with 2 Laptops - 1 TB, 4 GB RAM, Toshiba Windows 7

System-1 : 192.168.1.100
System-2 : 192.168.1.101

MongoDB 2.2.3 is installed in both the laptops at c:\\apps\mongodb.

Create directories in System-1
c:\\apps\mongodb\data1
c:\\apps\mongodb\data2
c:\\apps\mongodb\data3
c:\\apps\mongodb\conf

On System1
c:\apps\mongodb\bin> mongod --shardsvr --dbpath c:\apps\mongodb\data1 --port 27020

On System2
c:\apps\mongodb\bin> mongod --shardsvr --dbpath c:\apps\mongodb\data1 --port 27020

On System1
c:\apps\mongodb\bin> mongod --shardsvr --dbpath c:\apps\mongodb\data2 --port 27021

c:\apps\mongodb\bin> mongod --configsvr --dbpath c:\apps\mongodb\conf --port 27022

c:\apps\mongodb\bin> mongos --configsvr --configdb 192.168.1.100:27020,192.168.1.100:27021,192.168.1.101:27020 --port 27017

c:\apps\mongodb\bin> mongo 192.168.1.100:27017

mongos> use admin
switched to admin
mongos> db.runCommand({addShard:"192.168.1.100:27020"});
{"shardAdded": "shard0000", "ok": 1}
mongos> db.runCommand({addShard:"192.168.1.100:27021"});
{"shardAdded": "shard0001", "ok": 1}
mongos> db.runCommand({addShard:"192.168.1.101:27020"});
{"shardAdded": "shard0002", "ok": 1}
mongos> db.runCommand({listShards: 1})
{
 "shards" : [
   {
     "_id"  : "shard0000",
     "host" : "192.168.1.100:27020"
   },
   {
     "_id"  : "shard0001",
     "host" : "192.168.1.100:27021"
   },
   {
     "_id"  : "shard0002",
     "host" : "192.168.1.101:27020"
   }
 ]
 "ok" : 1
}

mongos> use twitterdb
switched to db twitterdb

mongos> db.createCollection("tweets")
{"ok" : 1}

mongos> use admin
switched to db admin

mongos> db.runCommand({enableSharding: "twitterdb"})
{ "ok" : 1}

mongos> db.runCommand({shardCollection: "twitterdb.tweets", key: {id_str: 1}})
{"collectionSharded" : "twitterdb.tweets", "ok" : 1}

mongos> use twitterdb
switched to db twitterdb

mongos> db.tweets.find().count()
0

Running the Application

So we have just finished setting up the Shards and the database setup. 

RUN The below Twitter Stream Application below (please change the appropriate values as per your settings). Data Starts pumping into MongoDB. Don't forget to stop the application when you are done, else twitter stream consumes network bandwidth and the mongodb storage will shoot up.

Keep a tab on the filesystem data:-
  System:1
c:\apps\mongodb\data1
c:\apps\mongodb\data2

  System:2
c:\apps\mongodb\data1

 The data files grows continously. Verify the counts of the tweets in the database.

  mongos> db.tweets.find().count()
25043

 So 25,000 tweets accumulated in 10 mins or 40 tweets per second Find out how many people tweeted in this 10 minutes using web

  mongos> db.tweets.find({"source" : "web"}).count()
4365

Now RUN the below MapReduce Job to run every 10 minutes and aggregate the results and generate the reports as Pie-Chart. These charts will be stored in your local file system.

Chart Utility

Download

The file can be downloaded here containing the entire project.

Massive Movie Recommendation System using MongoDB

MongoDB Basics & Internals

Big Data is a buzz word of today. It means Velocity, Volume and Variety of data. Google processes 24 PetaBytes of data per day. FaceBook users upload 300 Million photos and 3.2 Billion Likes/Comments per day. These are some examples of Velocity and Volume. "Variety" refers to different types of data handled by these applications. For instance, Log files have unstructured/semi-structured data. This is not what RDBMS is designed for. They are suitable for structured data and volumes not exceeding TeraBytes. Since RDBMS are disk-based they are relatively slow (IO-bound). Indexing is not a solution, when data grows inserts will get slow. Joins are big bottlenecks and hence you have to go for de-normalization.

High volume sites were mostly built on LAMP (Linux, Apache, MySQL, PHP) stack. Load-Balancing was done at web-tier and application-tier levels. However this was difficult with data-tier. So one solution people used was Memcached, acting as a caching layer in front of data-tier. They utilize a cluster of MySQL instances to split the data among these clusters. And these clusters were replicated as well. Further optimizations were done on this - MySQL Handler Socket was used to tune 500K requests per second. One of the solutions to address high-volumes was to use CDN(Content Delivery Networks) like Akamai. HTTP Accelerators like Varnish were used to handle web-tier load. 

When Google published papers on Big Table and Map-Reduce, it changed the game. Doug Cutting, who was working with Yahoo then, built Hadoop (an open-source implementation of big table and map-reduce). Several NoSQL databases appeared then. We can broadly categorize them as Document-Oriented (MongoDB, CouchDB), Graph (Neo4J), Columnar (HBase, Cassandra) Databases. MongoDB is a document-oriented database from 10Gen. FourSquare, ShutterFly, CodeAcademy are a few of the customers who uses MongoDB. 

MongoDB works on JSON datatypes and each document is a JSON Document, internally stored as BSON (Binary representation of JSON). Its void of Joins and Transactions. It supports Geo-Spatial queries. This is a USP for FourSquare to use MongoDB. Indexing, Auto-Sharding and Replication support are few more things in MongoDB. I could configure my 3 laptops to store Millions of User data. I configured laptop-1 and laptop-2 as Shards with users A-M stored on laptop-1 and N-Z on laptop-2. And laptop-3 I configured as a replica of laptop-1. Replication was immediate (less than a sec) and even offline-replication was simply awesome (bring down the replica and start again, automatically synced up). Replica Sets support voting for choosing Primary. There is support for Map-Reduce with Mongo (built on top of Hadoop). MongoDB provides several tools - important ones are mongod, mongos, mongo, mongostat. MongoDB uses Memory-mapped files to map the files to virtual memory. Hence 64-Bit machines are preferred with MongoDB. Internally the data is stored as a 2 level Linked-List. When a database (e.g. movieLens) is created it creates 3 files automatically - movieLens.ns (16 MB), movieLens.0 (64 MB) and movieLens.1 (128 MB). The movieLens.ns file is a giant HashMap that stores the metadata of the database and the collections ( tables are called collections in MongoDB). The movieLens.0 and movieLens.1 files are data files and the numbers are increased sequentially, sizes increased exponentially till it achieves 2 GB. The internal architecture and setup of MongoDB is a very interesting and a long story (for next post). 

Movie Recommendation

We will build a movie recommendation system with MongoDB to handle massive volume of data. I used the 10 Million movie ratings provided by MovieLens. Download the data (MovieLens 10M) from http://www.grouplens.org/node/73. Extract it to a location like C:\Project\Recommender. It consists of 3 main files - movies.dat, ratings.dat, and tags.dat. We will load this data to MongoDB. For this, download MongoDB from http://www.mongodb.org/downloads.  Install MongoDB in some location like c:\mongodb. Create a directory c:\mongodb\data. Start mongod with this command -

c:\mongodb\bin>  mongod --dbpath c:\mongodb\data

Now we will run the following Java application to import data onto MongoDB. For this we need MongoDB Java driver and set the Heap Size to 1 GB.

The output is as follows-

*********************
Inserted 10681 Movies
Inserted 95580 Tags
Inserted 10000054 Ratings
Completed in 685 Seconds
*********************

Run mongo using the following command

c:\mongodb\bin>  mongod

Now verify that the movieLens database is created. Some of the commands you can run are as below-

show dbs

show collections

db.movies.findOne()

db.ratings.find().count()

Next, create Indexes:

db.ratings.ensureIndex({user_id:1, item_id:1})

db.movies.ensureIndex({genres:1});

Next run map-reduce task to create a new Collection to store userId/RatingCount

map = function() { emit({user_id: this.user_id}, {count:1}) }

reduce = function(k, v) { var count = 0; v.forEach( function(x) { count += x['count']; }); return {count: count}; }

db.ratings.mapReduce(map, reduce, "user_ratings");

{

        "result" : "user_ratings",

        "timeMillis" : 520420,

        "counts" : {

                "input" : 10000054,

                "emit" : 10000054,

                "reduce" : 169150,

                "output" : 69878

        },

        "ok" : 1,

}

Now we are ready to create our Recommendation Engine. The one which we are going to implement is a "Collabarative Filtering" application, similar to Amazon or Netflix. The algorithms used by these sites are normally Weighted Slope-One or Singular Value Decomposition (SVD). Twitter for instance uses SVD. Apache Mahout (an open source Machine Learning solution for Big Data) provides Recommendation System. I have used a similar strategy here. 

The idea used here is- 

Users rate movies on a scale of 1-5. To know what movie to recommend for a given user (user_1), we use the ratings that he provided for movies. Find out k-neighbours (users who are similar to him in taste). From this list of users, find out the movies that they have rated. Get a weighted list of movies based on ratings, ignoring movies that he (user_1) has already rated. The basics for this are - Manhattan Distance, Eucledians Distance, Minkowski Distance, Pearson's Co-efficient and K-Nearest-Neighbors. 

The output of the program is-

*******************

Loaded 10000054 ratings
Loaded 10681 movies
Recommended for 16
Lord of the Rings: The Two Towers, The
Liar Liar
Rudy
Field of Dreams
Christmas Story, A
Gone in 60 Seconds
Remember the Titans
Striptease
Hoosiers
No Country for Old Men
Recommended for 127
Clockwork Orange, A
Fight Club
Reservoir Dogs
12 Monkeys (Twelve Monkeys)
Shawshank Redemption, The
Silence of the Lambs, The
Terminator 2: Judgment Day
One Flew Over the Cuckoo's Nest
Fargo
Shining, The

In the next post I will discuss more about how the Recommendation System works and the algorithm

Building an online Movie Library

This is more of a PoC (Proof of Concept) for the Movie Library Site. I have taken a couple of requirements to build a working code -

Requirements
The site maintains a list of tons of movies and each movie is identified by a title, price, release date and category. To assist the users in searching movies easily - a search field is provided. So if we enter "English" in the search box, it would return all English movies in the sorted order of title. It would also provide a "Faceted Navigation" - a concept which is used widely these days. So along with displaying the list of English movies, it would categories them on the left hand side section with search results based on "Categories". And the Categories will be displaying the title and will be sorted on the total occurrences found (desc order). Following is the result page for the application we are building (displaying Kannada movies)

The next requirement is to provide Pagination for the results as shown below. In the above page, if we click on "List All" we end up with the below page -

The next functionality we want to build is for the Suppliers. The suppliers want to add movies into our library and hence we provide them with a FTP location. Suppliers would create one or more CSV Files (Comma-Separated-File) of predefined format and put it onto the FTP location. We will create a scheduler application to poll the folder, pick up the files and extract and populate it in the database. On completion of this step, the movies will become available on the site. The format of the CSV is as below-

In a real-world scenario they would be creating a zip file containing CSV files and images.








Tools/Technologies Used-
JDK 1.6 (or above)
Maven 3.0.4
Eclipse Juno
Spring 3.1.1
Hibernate 4.1.4
Lucene/Hibernate Search 4.2.0
MySQL
Spring MVC 3.1.1
Jackson 1.9.4 (for JSON)
JQuery 1.7.2
Spring Integration 2.2.0 (Spring Batch support)
Mockito Test Framework for Test Driven Development (in progress...)

Implementation

1) Download and install JDK 1.6 from here
2) Download and install Eclipse IDE from here
3) Download and install Maven from here
4) Setup environment variables as follows-          

    JAVA_HOME=C:\Apps\Java\jdk1.7.0 

    M2_HOME=C:\Apps\apache-maven-3.0.4       

    Path=%JAVA_HOME%\bin;%M2_HOME%\bin
5) Download MYSQL and install from here 
6) Set the root username/password as root/root. Open MySQL Shell and create a database with name - "TPCH"
7) Create a table "CD" with initial data. Following SQL can be used-

https://gist.github.com/4572002

8) Next download the complete movie library project from here
9) Extract it to a folder, say C:\\Projects. This will create the following directory structure-

It is a maven project with 4 maven modules (highlighted in red). Apart from that there is a lucene folder, which is used for indexing.
The main project is a POM.

Web has dependency on Service and DAO
Service is dependent on DAO
Batch is independent stand-alone depends on DAO

Maven has a standard directory structure-
src/main/resources used for configuration
src/main/java used for source code
src/main/test used for test cases


10) Open Eclipse and import as maven project (root should be C:\\Projects\spring-mockito)
11) Configuration changes (if any)-

spring-mockito-dao/src/main/resources/hibernate.cfg.xml
Ensure the connection url, username and password are as specified by you.
The property "hibernate.search.default.indexBase" is used to specify where lucene has to store its index.

spring-mockito-dao/src/main/resources/spring-dao.xml
Ensure the values are right here

12) Run the following commands from a command prompt-
C:\Project\spring-mockito>mvn clean install
C:\Project\spring-mockito\spring-mockito-web>mvn jetty:run

The first command cleans up the maven modules and creates the targets.
The second command deploys the war file and starts-up jetty (light weight server). If you want to test on tomcat, please copy the war from spring-mockito-web\target folder and deploy to tomcat directory.

13) To access the application the url is http://localhost:8080/spring-mockito-web/index.jsp
14) To test the Supplier functionality first ensure that you have a directory C:\\FileServer\input created already. You can change the location in the file
spring-mockito-batch/src/main/resources/ApplicationContext-File.xml

Run the stand-alone java class com.raj.projects.batch.client.Main from spring-mockito-batch project. This will startup the poller and the integration environment.

15) Create files with .csv extension as shown below-

You can store as many files as you want. However ensure that the Ids are unique. 

Drop the files into the folder C:\\FileServer\input

They will be picked up and processed and movie entries written to database. Then the files will be deleted.

Count Distinct Values in huge text file

This is a common problem today. You are given a huge text file, containing tens of millions of text data. And you have to count how many distinct values exist in the file. The problem gets more complicated, if you cannot load the entire data into memory. How do you make your algorithm "Online". An "online algorithm" is one in which, the result is expected to be returned immediately - may be in one pass, and you won't have memory to hold all the data.

The usage of this algorithm is in various fields. In web analytics to return number of unique visitors last month, last week, so on. Another usage is in detecting a suspected attack. If the load on your server increases all of a sudden, and you notice a spike in the network, then the first thing to look for the possibility of an attack. In this case, the requests could be coming repeatedly from a system. So next we will discuss different implementations available. For the purpose of this article, I used Shakespeare's complete work as a huge text file. It can be downloaded from http://www.gutenberg.org/cache/epub/100/pg100.txt. The file is around 5.5 MB in size and contains 67802 distinct words.

APPROACH:1
Read the input file line by line, break it into words. Then store the word in a set. Once all the words are stored, then get the size of the set. This is the distinct count. However with this approach, we have to store the entire 5.5 MB data in the Set. Assuming that there are repeated words, and say 4 MB is the distinct words. Still we need to save 4 MB data in Set. And since a Set internally uses a dictionary, and assuming 0.7 load factor then we would need 4/0.7 = 5.7 MB. And if the file is extremely big then we cannot use this approach at all.

APPROACH:2
A small work-around for Approach 1 is here. Assume that you have a very good hashCode function and the hashCode for any 2 words are never the same. In this case, store the hashCode instead of the word. So hashCode being int and when stored in a set as Integer, it is 32 bits or 4-bytes long. So to store 67802 hashCodes we need 378 KB memory. This is a huge gain, but the problem is we should find a good hashCode function. And the size of the Set increases with the input file size.

APPROACH:3
We can use "Bloom Filters" as an alternative. It is a 2-dimensional boolean array. In this approach, we use 3 or 4 different hashCode functions - h1, h2, h3, h4. For each word, compute h1 to h4. The rows of our bloom filter are the hashCodes and columns are fixed based on the accuracy needed. One thing to note here is that, the Bloom Filter is a probabilistic algorithm. It doesn't give 100% accurate results always. There is a chance of error 1% to 4% or so. The bigger the size of the Bloom Filter array sizes, the more accurate the results are. Assume that the values for h1, h2, h3, and h4 are 22, 17, 35, and 19. If the bloom filter array is represented as a, we set a[22][0], a[17][1], a[35][2], a[19][3] to true. Using this approach, I could get a result of 67801 instead of 67802 with array 4*1M. Since boolean is a byte in Java, we would need 4*1M bytes or 3 MB memory. However with a memory of 300 KB, I got 3% error. The efficiency of these algorithms are - we don't need to store the entire data and we can get the result in one pass.

Solution

APPROACH:4
The next approach we discuss is a very efficient algorithm. And the same has been used in many applications - including Cassandra (by Facebook). It is called "HyperLogLog Counter". Its a slightly complex and mathematical algorithm. It needs very little memory and doesn't vary much on the input data size. The idea is as follows. Read each word and compute the hash of the word. Cassandra uses "MurMur Hash". In this algorithm we use something called "Stochastic Averaging". We define 'b' as a number between 4 and 16. And create an int (or byte) array of size 2^b. So if we use b=4, then we have 16 buckets. Assume that we have chosen b=15, hashCode of a word is 1379266913. We convert this hashCode to a 32-bit binary number (Integer is 32-bit). And then break it into first 15 bits and next 17 bits. The first 15-bits decide which bucket to use. The next 17-bit is used to determine the rank. The rank is nothing but, the position of the first set-bit (or position of 1), from left to right. At the bucket index, if the value is lesser than the rank, then rank is stored in the bucket. So basically at each bucket, we store the highest 1-bit position. Once all the words are read and the buckets populated, we use the harmonic mean to compute the result.
Using this approach with b-value of 15, meaning 32768 buckets ( or 32 KB) memory the result obtained was 67749 (against the actual 67802). Which means using just 32 KB memory we got a result which is 99.92% accurate!!

Solution