Learn By Example: Hadoop, MapReduce for Big Data problems
Taught by a 4 person team including 2 Stanford-educated, ex-Googlers and 2 ex-Flipkart Lead Analysts. This team has decades of practical experience in working with Java and with billions of rows of data.
This course is a zoom-in, zoom-out, hands-on workout involving Hadoop, MapReduce and the art of thinking parallel.
Let’s parse that.
Zoom-in, Zoom-Out: This course is both broad and deep. It covers the individual components of Hadoop in great detail, and also gives you a higher level picture of how they interact with each other.
Hands-on workout involving Hadoop, MapReduce : This course will get you hands-on with Hadoop very early on. You’ll learn how to set up your own cluster using both VMs and the Cloud. All the major features of MapReduce are covered – including advanced topics like Total Sort and Secondary Sort.
The art of thinking parallel: MapReduce completely changed the way people thought about processing Big Data. Breaking down any problem into parallelizable units is an art. The examples in this course will train you to “think parallel”.
Lot’s of cool stuff ..
- Using MapReduce to
- Recommend friends in a Social Networking site: Generate Top 10 friend recommendations using a Collaborative filtering algorithm.
- Build an Inverted Index for Search Engines: Use MapReduce to parallelize the humongous task of building an inverted index for a search engine.
- Generate Bigrams from text: Generate bigrams and compute their frequency distribution in a corpus of text.
- Build your Hadoop cluster:
- Install Hadoop in Standalone, Pseudo-Distributed and Fully Distributed modes
- Set up a hadoop cluster using Linux VMs.
- Set up a cloud Hadoop cluster on AWS with Cloudera Manager.
- Understand HDFS, MapReduce and YARN and their interaction
- Customize your MapReduce Jobs:
- Chain multiple MR jobs together
- Write your own Customized Partitioner
- Total Sort : Globally sort a large amount of data by sampling input files
- Secondary sorting
- Unit tests with MR Unit
- Integrate with Python using the Hadoop Streaming API
.. and of course all the basics:
- MapReduce : Mapper, Reducer, Sort/Merge, Partitioning, Shuffle and Sort
- HDFS & YARN: Namenode, Datanode, Resource manager, Node manager, the anatomy of a MapReduce application, YARN Scheduling, Configuring HDFS and YARN to performance tune your cluster.
We start off with an introduction on what this course is all about.
Why is Big Data a Big Deal
Big data may be a cliched term, but what does it really mean? Where does this data come from and why is it big?
Distributed computing makes processing very fast - but why? Let's take a simple example and see why distributed computing is so powerful.
What exactly is Hadoop? Its origins and its logical components explained.
HDFS based on GFS (The Google File System) is the storage layer within Hadoop. It stores files in blocks of 128MB.
MapReduce is the framework which allows developers to write massively parallel programs without worrying about the underlying details of distributed computing. The developer simply implements the map() and reduce() functions in order to crunch large input sets of data.
Yarn is responsible for managing resources in the Hadoop cluster. Yarn was introduced recently in Hadoop 2.0.
Installing Hadoop in a Local Environment
Hadoop has 3 different install modes - Standalone, Pseudo-distributed and Fully Distributed. Get an overview of when to use each
How to set up Hadoop in the standalone mode. Windows users need to install a Virtual Linux instance before this video.
Set up Hadoop in the Pseudo-Distributed mode. All Hadoop services will be up and running!
The MapReduce "Hello World"
In the world of MapReduce every problem can be thought of in terms of key values pairs. Map transforms the key-value pair in a meaningful way, they are sorted and merged and reduce combines key-value pairs in a meaningful way.
If you're learning MapReduce for the very first time - it's best to visualize what exactly it does before you get down into the little details.
What really goes on with a single record as it flows through the map and then reduce phase?
Counting the number of times a word occurs in input text is the Hello World of MapReduce. This was the very first example given in Jeff Dean and Sanjay Ghemawat's original paper on MapReduce.
Nothing is real unless it is on code. Setting up our very first Mapper.
Nothing is real unless it is on code. Setting up our very first Reducer.
Nothing is real unless it is on code. Setting up our very first MapReduce Job.
Run a MapReduce Job
Learn how to use HDFS's command line interface and add data to HDFS to run your jobs on.
Run your very first MapReduce Job. We'll also explore the Web interface for YARN and HDFS and see how to track your jobs.
Juicing your MapReduce - Combiners, Shuffle and Sort and The Streaming API
The reduce phase can be optimized by combining the output of the map phase at the map node itself. This is an optimization of the reduce phase to allow it to work on data that has been "partially reduced".
Using a Combiner should not change the output of the MapReduce. Which means not every Reducer can work as a combine function
The number of mapper processes depend on the number of input splits of your data. It's not really in your control. What you, as a developer, do control, is the number of reducers.
In order to have more than one Reducer work on your map data, you need partitions. Visualize how partitions and shuffle and sort work.
The Hadoop Streaming API uses the standard input and output to communicate with mapper and reducer functions in any language. Understand how Hadoop interacts with mappers and reducers in other languages.
It's not real till it's in code. Implement the word count MapReduce example in Python using the Streaming API.
HDFS and Yarn
Let's understand HDFS and it's data replication strategy in some detail.
Name nodes provide an index of what file is stored where in the data nodes. If the name node is lost the mapping of where the files are is lost. Which means even though the data is present in the data nodes, we'll have no idea how to access it!
Hadoop backs up name nodes using two strategies. Backing up the snapshot and edits to the file system and by setting up a secondary name node.
The Resource Manager assigns resources to processes based on policies and constraints of the cluster while the Node Manager manages memory, and other resource for a single node. These two form the basic components of Yarn.
What happens under the hood when you submit a job to Yarn? Resource Manager, Container, the Application Master and the Node Manager all work together to run your MapReduce job.
The Resource Manager acts as a pure scheduler and allows plugging in different policies to schedule jobs. Understand how the FIFO scheduler, the Capacity scheduler and the Fair scheduler work.
The user has a lot of leeway in configuring how the scheduler works. Let's study some of the options we can specify in the various config files.
MapReduce Customizations For Finer Grained Control
The Main class in your MapReduce needs some special set up before it can accept command line arguments.
The library classes and interfaces which allow parsing command line arguments. Learn what they are and how to use them.
The Job object allows you to plug in your own classes to control inputs, outputs and many intermediate steps in the MapReduce.
Between the Map phase and the Reduce phase lie a whole number of intermediate steps performed by the Hadoop framework. Partitioning, Sorting and Grouping are 3 specific operations and each of these can be customized to fit your problem statement.
The Inverted Index, Custom Data Types for Keys, Bigram Counts and Unit Tests!
The Inverted Index which provides a mapping from every word to the page on which that word occurs is at the heart of every search engine. This is one of the original use cases for MapReduce.
It's not real unless it's in code, generate the inverted index using a MR job.
Understand why we need the Writable and the WritableComparable interface and why the keys in the Mapper output implement these interfaces.
A Bigram is a pair of adjacent words, use a special data type to represent a Bigram, it needs to be a WritableComparable to be serialized across the network and sorted and merged by Hadoop.
Use the Bigram data type in your MapReduce to produce a count of all Bigrams in the input text file.
Follow these instructions to set up your Hadoop project.
No code is complete without unit tests. The MRUnit framework uses JUnit to test MapReduce jobs. Write test cases for the Bigram count code.
Input and Output Formats and Customized Partitioning
The Input Format specifies the kind of input data that feeds into the MapReduce. The FileInputFormat is the base class for all inputs which are files
The most common kind of files are text files and binary files and Hadoop has built in library classes to represent both of these.
What if you want to partition on something other than key hashes? Custom partitioners allow you to partition on whatever metric you, you just need to write a bit of code.
Total Order Partitioning is a mind bending concept in Hadoop. This allows you to locally sort data such that it's in globally sorted order. Sounds confusing? It is a hard concept to wrap one's head around but the results are pretty amazing!
Input sampling, samples the input data to produce a key to partition mapping. The total order partitioner uses this mapping to partition the data in such a manner that locally sorting the data results in a globally sorted result.
The Hadoop Sort/Merge operation sorts the output keys of the mapper. Here is a neat trick to sort the values for each key as well.
Recommendation Systems using Collaborative Filtering
At the heart of recommendation systems is a beautifully simple idea called collaborative filtering. If 2 users have a lot in common then the chances are that what one user likes the other will as well. You can recommend the users' likes to each other!
Recommend potential friends to all users of a social network. This involves using 2 MapReduce jobs and chaining them in such a way that the output of one MapReduce feeds into the second MapReduce
The first MapReduce finds the number of common friends for every pair of users. This requires special treatment for users who are already friends.
The second MapReduce takes in the common friends for every pair of users and generates the top 10 friend recommendations for every user of the social network.
Note there are 2 MR jobs chained together.
Hadoop as a Database
How is Hadoop different from a database? Can we leverage the power of Hadoop for structured data?
Let's see how to implement SQL Select , Where constructs using MapReduce
Select and Where constructs are implemented in the Mapper. Group By and having constructs are implemented in the reducer.
Joins can be surprisingly tricky to implement with MapReduce - Let's see what the Mapper looks like for a Join.
What should the Reducer do in a Join?
For the Join to work properly, you'll need to customize how the Sorting and Partitioning occurs in the MapReduce.
We continue with MapReduce joins. Let's put everything together in the Job.
K-Means Clustering is a popular machine learning algorithm. We start with an intro to Clustering and how the K-Means algorithm works.
We continue with understanding the K-Means algorithm. We'll break down the algorithm into a MapReduce task.
We'll start describing the code to implement K-Means with MapReduce. First some setup to represent data as points and measure the distance between them.
We need to set up a couple of Custom Writables that can be used by MapReduce for the input and output.
We're finally on to the MapReduce bits. We start by configuring the job and doing some setup that will be needed for the Mapper/Reducer.
The Mapper and Reducer for K-Means run once for each iteration and update the cluster centers.
Finally, we need to set it up so that the Jobs run iteratively and stop when convergence occurs.
Setting up a Hadoop Cluster
Manually configure a Hadoop cluster. You'll use Linux Virtual Machines to do this. Please go through the "Setting up a Virtual Linux Instance (in the Installing Hadoop on Local Environment section) before this video.
You can use a cloud service to setup a Hadoop Cluster. This video gets you started with AWS and the Elastic Compute 2 CLI tools
Install Cloudera Manager on AWS and use it to launch a Hadoop Cluster.
Hadoop is basically for Linux/Unix systems. If you are on Windows, you can set up a Linux Virtual Machine on your computer and use that for the install.
If you are unfamiliar with softwares that require working with a shell/command line environment, this video will be helpful for you. It explains how to update the PATH environment variable, which is needed to set up most Linux/Mac shell based softwares.