Hadoop | Big Data Projects

What Is Apache Hadoop?

The Apache Hadoop software library is a framework that allows for the distributed processing of large data sets across clusters of computers using simple programming models. It is designed to scale up from single servers to thousands of machines, each offering local computation and storage. Rather than rely on hardware to deliver high-avaiability, the library itself is designed to detect and handle failures at the application layer, so delivering a highly-availabile service on top of a cluster of computers, each of which may be prone to failures.

Apache Hadoop is 100% open source, and pioneered a fundamentally new way of storing and processing data. Instead of relying on expensive, proprietary hardware and different systems to store and process data, Hadoop enables distributed parallel processing of huge amounts of data across inexpensive, industry-standard servers that both store and process the data, and can scale without limits. With Hadoop, no data is too big. And in today’s hyper-connected world where more and more data is being created every day, Hadoop’s breakthrough advantages mean that businesses and organizations can now find value in data that was recently considered useless.

Hadoop makes it possible to run applications on systems with thousands of nodes involving thousands of terabytes. Its distributed file system facilitates rapid data transfer rates among nodes and allows the system to continue operating uninterrupted in case of a node failure. This approach lowers the risk of catastrophic system failure, even if a significant number of nodes become inoperative.

Hadoop was inspired by Google’s MapReduce, a software framework in which an application is broken down into numerous small parts. Any of these parts (also called fragments or blocks) can be run on any node in the cluster. Doug Cutting, Hadoop’s creator, named the framework after his child’s stuffed toy elephant. The current Apache Hadoop ecosystem consists of the Hadoop kernel, MapReduce, the Hadoop distributed file system (HDFS) and a number of related projects such as Apache Hive, HBase and Zookeeper.

The Hadoop framework is used by major players including Google, Yahoo and IBM, largely for applications involving search engines and advertising. The preferred operating systems are Windows and Linux but Hadoop can also work with BSD and OS X.

The project includes these modules:

Hadoop Common: The common utilities that support the other Hadoop modules.

Hadoop Distributed File System (HDFS™): A distributed file system that provides high-throughput access to application data.

Hadoop YARN: A framework for job scheduling and cluster resource management.

Hadoop MapReduce: A YARN-based system for parallel processing of large data sets.


Reference : http://hadoop.apache.org/

Yahoo! Hadoop Tutorial

Table of Contents

Welcome to the Yahoo! Hadoop Tutorial. This tutorial includes the following materials designed to teach you how to use the Hadoop distributed data processing environment:

  • Hadoop 0.18.0 distribution (includes full source code)
  • A virtual machine image running Ubuntu Linux and preconfigured with Hadoop
  • VMware Player software to run the virtual machine image
  • A tutorial which will guide you through many aspects of Hadoop’s installation and operation.

The tutorial is divided into seven modules, designed to be worked through in order. They can be accessed from the links below.

  1. Tutorial Introduction
  2. The Hadoop Distributed File System
  3. Getting Started With Hadoop
  4. MapReduce
  5. Advanced MapReduce Features
  6. Related Topics
  7. Managing a Hadoop Cluster
  8. Pig Tutorial

Module 1: Tutorial Introduction


Welcome to the Yahoo! Hadoop tutorial! This series of tutorial documents will walk you through many aspects of the Apache Hadoop system. You will be shown how to set up simple and advanced cluster configurations, use the distributed file system, and develop complex Hadoop MapReduce applications. Other related systems are also reviewed.

Goals for this Module:

  • Understand the scope of problems applicable to Hadoop
  • Understand how Hadoop addresses these problems differently from other distributed systems.


  1. Introduction
  2. Goals for this Module
  3. Outline
  4. Problem Scope
    1. Challenges at Large Scale
    2. Moore’s Law
  5. The Hadoop Approach
    1. Comparison to Existing Techniques
    2. Data Distribution
    3. MapReduce: Isolated Processes
    4. Flat Scalability
  6. The Rest of the Tutorial

Problem Scope

Hadoop is a large-scale distributed batch processing infrastructure. While it can be used on a single machine, its true power lies in its ability to scale to hundreds or thousands of computers, each with several processor cores. Hadoop is also designed to efficiently distribute large amounts of work across a set of machines.

How large an amount of work? Orders of magnitude larger than many existing systems work with. Hundreds of gigabytes of data constitute the low end of Hadoop-scale. Actually Hadoop is built to process “web-scale” data on the order of hundreds of gigabytes to terabytes or petabytes. At this scale, it is likely that the input data set will not even fit on a single computer’s hard drive, much less in memory. So Hadoop includes a distributed file system which breaks up input data and sends fractions of the original data to several machines in your cluster to hold. This results in the problem being processed in parallel using all of the machines in the cluster and computes output results as efficiently as possible.

Challenges at Large Scale

Performing large-scale computation is difficult. To work with this volume of data requires distributing parts of the problem to multiple machines to handle in parallel. Whenever multiple machines are used in cooperation with one another, the probability of failures rises. In a single-machine environment, failure is not something that program designers explicitly worry about very often: if the machine has crashed, then there is no way for the program to recover anyway.

In a distributed environment, however, partial failures are an expected and common occurrence. Networks can experience partial or total failure if switches and routers break down. Data may not arrive at a particular point in time due to unexpected network congestion. Individual compute nodes may overheat, crash, experience hard drive failures, or run out of memory or disk space. Data may be corrupted, or maliciously or improperly transmitted. Multiple implementations or versions of client software may speak slightly different protocols from one another. Clocks may become desynchronized, lock files may not be released, parties involved in distributed atomic transactions may lose their network connections part-way through, etc. In each of these cases, the rest of the distributed system should be able to recover from the component failure or transient error condition and continue to make progress. Of course, actually providing such resilience is a major software engineering challenge.

Different distributed systems specifically address certain modes of failure, while worrying less about others. Hadoop provides no security model, nor safeguards against maliciously inserted data. For example, it cannot detect a man-in-the-middle attack between nodes. On the other hand, it is designed to handle hardware failure and data congestion issues very robustly. Other distributed systems make different trade-offs, as they intend to be used for problems with other requirements (e.g., high security).

In addition to worrying about these sorts of bugs and challenges, there is also the fact that the compute hardware has finite resources available to it. The major resources include:

  • Processor time
  • Memory
  • Hard drive space
  • Network bandwidth

Individual machines typically only have a few gigabytes of memory. If the input data set is several terabytes, then this would require a thousand or more machines to hold it in RAM — and even then, no single machine would be able to process or address all of the data.

Hard drives are much larger; a single machine can now hold multiple terabytes of information on its hard drives. But intermediate data sets generated while performing a large-scale computation can easily fill up several times more space than what the original input data set had occupied. During this process, some of the hard drives employed by the system may become full, and the distributed system may need to route this data to other nodes which can store the overflow.

Finally, bandwidth is a scarce resource even on an internal network. While a set of nodes directly connected by a gigabit Ethernet may generally experience high throughput between them, if all of the machines were transmitting multi-gigabyte data sets, they can easily saturate the switch’s bandwidth capacity. Additionally if the machines are spread across multiple racks, the bandwidth available for the data transfer would be much less. Furthermore RPC requests and other data transfer requests using this channel may be delayed or dropped.

To be successful, a large-scale distributed system must be able to manage the above mentioned resources efficiently. Furthermore, it must allocate some of these resources toward maintaining the system as a whole, while devoting as much time as possible to the actual core computation.

Synchronization between multiple machines remains the biggest challenge in distributed system design. If nodes in a distributed system can explicitly communicate with one another, then application designers must be cognizant of risks associated with such communication patterns. It becomes very easy to generate more remote procedure calls (RPCs) than the system can satisfy! Performing multi-party data exchanges is also prone to deadlock or race conditions. Finally, the ability to continue computation in the face of failures becomes more challenging. For example, if 100 nodes are present in a system and one of them crashes, the other 99 nodes should be able to continue the computation, ideally with only a small penalty proportionate to the loss of 1% of the computing power. Of course, this will require re-computing any work lost on the unavailable node. Furthermore, if a complex communication network is overlaid on the distributed infrastructure, then determining how best to restart the lost computation and propagating this information about the change in network topology may be non trivial to implement.

Moore’s Law

So why use a distributed system at all? They seem like more trouble than they’re worth. And with the fast pace of computer hardware design, it seems inevitable that single-chip hardware will be able to “grow up” to handle the larger volumes of data. After all, Moore’s Law (named after Gordon Moore, the founder of Intel) states that the number of transistors that can be placed in a processor will double approximately every two years, for half the cost. But trends in chip design are changing to face new realities. While we can still double the number of transistors per unit area at this pace, this does not necessarily result in faster single-threaded performance. New processors such as Intel Core 2 and Itanium 2 architectures now focus on embedding many smaller CPUs or “cores” onto the same physical device. This allows multiple threads to process twice as much data in parallel, but at the same speed at which they operated previously.

Even if hundreds or thousands of CPU cores are placed on a single machine, it would not be possible to deliver input data to these cores fast enough for processing. Individual hard drives can only sustain read speeds between 60-100 MB/second. These speeds have been increasing over time, but not at the same breakneck pace as processors. Optimistically assuming the upper limit of 100 MB/second, and assuming four independent I/O channels are available to the machine, that provides 400 MB of data every second. A 4 terabyte data set would thus take over 10,000 seconds to read–about three hours just to load the data! With 100 separate machines each with two I/O channels on the job, this drops to three minutes.

The Hadoop Approach

Hadoop is designed to efficiently process large volumes of information by connecting many commodity computers together to work in parallel. The theoretical 1000-CPU machine described earlier would cost a very large amount of money, far more than 1,000 single-CPU or 250 quad-core machines. Hadoop will tie these smaller and more reasonably priced machines together into a single cost-effective compute cluster.

Comparison to Existing Techniques

Performing computation on large volumes of data has been done before, usually in a distributed setting. What makes Hadoop unique is its simplified programming model which allows the user to quickly write and test distributed systems, and its efficient, automatic distribution of data and work across machines and in turn utilizing the underlying parallelism of the CPU cores.

Grid scheduling of computers can be done with existing systems such as Condor. But Condor does not automatically distribute data: a separate SAN must be managed in addition to the compute cluster. Furthermore, collaboration between multiple compute nodes must be managed with a communication system such as MPI. This programming model is challenging to work with and can lead to the introduction of subtle errors.

Data Distribution

In a Hadoop cluster, data is distributed to all the nodes of the cluster as it is being loaded in. The Hadoop Distributed File System (HDFS) will split large data files into chunks which are managed by different nodes in the cluster. In addition to this each chunk is replicated across several machines, so that a single machine failure does not result in any data being unavailable. An active monitoring system then re-replicates the data in response to system failures which can result in partial storage. Even though the file chunks are replicated and distributed across several machines, they form a single namespace, so their contents are universally accessible.

Data is conceptually record-oriented in the Hadoop programming framework. Individual input files are broken into lines or into other formats specific to the application logic. Each process running on a node in the cluster then processes a subset of these records. The Hadoop framework then schedules these processes in proximity to the location of data/records using knowledge from the distributed file system. Since files are spread across the distributed file system as chunks, each compute process running on a node operates on a subset of the data. Which data operated on by a node is chosen based on its locality to the node: most data is read from the local disk straight into the CPU, alleviating strain on network bandwidth and preventing unnecessary network transfers. This strategy of moving computation to the data, instead of moving the data to the computation allows Hadoop to achieve high data locality which in turn results in high performance.


Figure 1.1: Data is distributed across nodes at load time.

MapReduce: Isolated Processes

Hadoop limits the amount of communication which can be performed by the processes, as each individual record is processed by a task in isolation from one another. While this sounds like a major limitation at first, it makes the whole framework much more reliable. Hadoop will not run just any program and distribute it across a cluster. Programs must be written to conform to a particular programming model, named “MapReduce.”

In MapReduce, records are processed in isolation by tasks called Mappers. The output from the Mappers is then brought together into a second set of tasks called Reducers, where results from different mappers can be merged together.

mapreduce-processFigure 1.2: Mapping and reducing tasks run on nodes where individual records of data are already present.

Separate nodes in a Hadoop cluster still communicate with one another. However, in contrast to more conventional distributed systems where application developers explicitly marshal byte streams from node to node over sockets or through MPI buffers, communication in Hadoop is performed implicitly. Pieces of data can be tagged with key names which inform Hadoop how to send related bits of information to a common destination node. Hadoop internally manages all of the data transfer and cluster topology issues.

By restricting the communication between nodes, Hadoop makes the distributed system much more reliable. Individual node failures can be worked around by restarting tasks on other machines. Since user-level tasks do not communicate explicitly with one another, no messages need to be exchanged by user programs, nor do nodes need to roll back to pre-arranged checkpoints to partially restart the computation. The other workers continue to operate as though nothing went wrong, leaving the challenging aspects of partially restarting the program to the underlying Hadoop layer.

Flat Scalability

One of the major benefits of using Hadoop in contrast to other distributed systems is its flat scalability curve. Executing Hadoop on a limited amount of data on a small number of nodes may not demonstrate particularly stellar performance as the overhead involved in starting Hadoop programs is relatively high. Other parallel/distributed programming paradigms such as MPI (Message Passing Interface) may perform much better on two, four, or perhaps a dozen machines. Though the effort of coordinating work among a small number of machines may be better-performed by such systems, the price paid in performance and engineering effort (when adding more hardware as a result of increasing data volumes) increases non-linearly.

A program written in distributed frameworks other than Hadoop may require large amounts of refactoring when scaling from ten to one hundred or one thousand machines. This may involve having the program be rewritten several times; fundamental elements of its design may also put an upper bound on the scale to which the application can grow.

Hadoop, however, is specifically designed to have a very flat scalability curve. After a Hadoop program is written and functioning on ten nodes, very little–if any–work is required for that same program to run on a much larger amount of hardware. Orders of magnitude of growth can be managed with little re-work required for your applications. The underlying Hadoop platform will manage the data and hardware resources and provide dependable performance growth proportionate to the number of machines available.

The Rest of the Tutorial

This module of the tutorial has highlighted the major benefits of using a system such as Hadoop. The rest of the tutorial is designed to show you how to effectively use it.

  • In Module 2, you’ll learn how the Hadoop Distributed File System (HDFS) stores vast quantities of information, how to configure HDFS, and how to use it to store and retrieve your data.
  • Module 3 shows you how to get started setting up a Hadoop environment to experiment with. It reviews how to install a Hadoop virtual machine (included in this resource CD) so that you can run Hadoop regardless of what operating system you are running.
  • Module 4 explains the Hadoop MapReduce programming model itself, and how to write some MapReduce programs.
  • Module 5 goes into further detail about the specifics of Hadoop MapReduce, and how to use advanced features for more powerful control over a program’s execution.
  • Module 6 describes some other components of the Hadoop ecosystem which can add further capabilities to your distributed system.
  • Module 7 describes how to configure Hadoop clusters of different sizes. It describes what particular parameters of Hadoop need to be tuned for setting up clusters of various sizes. In addition it describes the various performance monitoring tools available in Hadoop for monitoring the health of your cluster.
  • And to expand upon the Pig section described in Module 6, a separate Pig Tutorial is included in this package at the end as Module 8.

Good luck!


Module 2: The Hadoop Distributed File System


HDFS, the Hadoop Distributed File System, is a distributed file system designed to hold very large amounts of data (terabytes or even petabytes), and provide high-throughput access to this information. Files are stored in a redundant fashion across multiple machines to ensure their durability to failure and high availability to very parallel applications. This module introduces the design of this distributed file system and instructions on how to operate it.

Goals for this Module:

  • Understand the basic design of HDFS and how it relates to basic distributed file system concepts
  • Learn how to set up and use HDFS from the command line
  • Learn how to use HDFS in your applications


  1. Introduction
  2. Goals for this Module
  3. Outline
  4. Distributed File System Basics
  5. Configuring HDFS
  6. Interacting With HDFS
    1. Common Example Operations
    2. HDFS Command Reference
    3. DFSAdmin Command Reference
  7. Using HDFS in MapReduce
  8. Using HDFS Programmatically
  9. HDFS Permissions and Security
  10. Additional HDFS Tasks
    1. Rebalancing Blocks
    2. Copying Large Sets of Files
    3. Decommissioning Nodes
    4. Verifying File System Health
    5. Rack Awareness
  11. HDFS Web Interface
  12. References

Distributed File System Basics

A distributed file system is designed to hold a large amount of data and provide access to this data to many clients distributed across a network. There are a number of distributed file systems that solve this problem in different ways.

NFS, the Network File System, is the most ubiquitous distributed file system. It is one of the oldest still in use. While its design is straightforward, it is also very constrained. NFS provides remote access to a single logical volume stored on a single machine. An NFS server makes a portion of its local file system visible to external clients. The clients can then mount this remote file system directly into their own Linux file system, and interact with it as though it were part of the local drive.

One of the primary advantages of this model is its transparency. Clients do not need to be particularly aware that they are working on files stored remotely. The existing standard library methods like open(), close(), fread(), etc. will work on files hosted over NFS.

But as a distributed file system, it is limited in its power. The files in an NFS volume all reside on a single machine. This means that it will only store as much information as can be stored in one machine, and does not provide any reliability guarantees if that machine goes down (e.g., by replicating the files to other servers). Finally, as all the data is stored on a single machine, all the clients must go to this machine to retrieve their data. This can overload the server if a large number of clients must be handled. Clients must also always copy the data to their local machines before they can operate on it.

HDFS is designed to be robust to a number of the problems that other DFS’s such as NFS are vulnerable to. In particular:

  • HDFS is designed to store a very large amount of information (terabytes or petabytes). This requires spreading the data across a large number of machines. It also supports much larger file sizes than NFS.
  • HDFS should store data reliably. If individual machines in the cluster malfunction, data should still be available.
  • HDFS should provide fast, scalable access to this information. It should be possible to serve a larger number of clients by simply adding more machines to the cluster.
  • HDFS should integrate well with Hadoop MapReduce, allowing data to be read and computed upon locally when possible.

But while HDFS is very scalable, its high performance design also restricts it to a particular class of applications; it is not as general-purpose as NFS. There are a large number of additional decisions and trade-offs that were made with HDFS. In particular:

  • Applications that use HDFS are assumed to perform long sequential streaming reads from files. HDFS is optimized to provide streaming read performance; this comes at the expense of random seek times to arbitrary positions in files.
  • Data will be written to the HDFS once and then read several times; updates to files after they have already been closed are not supported. (An extension to Hadoop will provide support for appending new data to the ends of files; it is scheduled to be included in Hadoop 0.19 but is not available yet.)
  • Due to the large size of files, and the sequential nature of reads, the system does not provide a mechanism for local caching of data. The overhead of caching is great enough that data should simply be re-read from HDFS source.
  • Individual machines are assumed to fail on a frequent basis, both permanently and intermittently. The cluster must be able to withstand the complete failure of several machines, possibly many happening at the same time (e.g., if a rack fails all together). While performance may degrade proportional to the number of machines lost, the system as a whole should not become overly slow, nor should information be lost. Data replication strategies combat this problem.

The design of HDFS is based on the design of GFS, the Google File System. Its design was described in a paper published by Google.

HDFS is a block-structured file system: individual files are broken into blocks of a fixed size. These blocks are stored across a cluster of one or more machines with data storage capacity. Individual machines in the cluster are referred to as DataNodes. A file can be made of several blocks, and they are not necessarily stored on the same machine; the target machines which hold each block are chosen randomly on a block-by-block basis. Thus access to a file may require the cooperation of multiple machines, but supports file sizes far larger than a single-machine DFS; individual files can require more space than a single hard drive could hold.

If several machines must be involved in the serving of a file, then a file could be rendered unavailable by the loss of any one of those machines. HDFS combats this problem by replicating each block across a number of machines (3, by default).


Figure 2.1: DataNodes holding blocks of multiple files with a replication factor of 2. The NameNode maps the filenames onto the block ids.

Most block-structured file systems use a block size on the order of 4 or 8 KB. By contrast, the default block size in HDFS is 64MB — orders of magnitude larger. This allows HDFS to decrease the amount of metadata storage required per file (the list of blocks per file will be smaller as the size of individual blocks increases). Furthermore, it allows for fast streaming reads of data, by keeping large amounts of data sequentially laid out on the disk. The consequence of this decision is that HDFS expects to have very large files, and expects them to be read sequentially. Unlike a file system such as NTFS or EXT, which see many very small files, HDFS expects to store a modest number of very large files: hundreds of megabytes, or gigabytes each. After all, a 100 MB file is not even two full blocks. Files on your computer may also frequently be accessed “randomly,” with applications cherry-picking small amounts of information from several different locations in a file which are not sequentially laid out. By contrast, HDFS expects to read a block start-to-finish for a program. This makes it particularly useful to the MapReduce style of programming described in Module 4. That having been said, attempting to use HDFS as a general-purpose distributed file system for a diverse set of applications will be suboptimal.

Because HDFS stores files as a set of large blocks across several machines, these files are not part of the ordinary file system. Typing ls on a machine running a DataNode daemon will display the contents of the ordinary Linux file system being used to host the Hadoop services — but it will not include any of the files stored inside the HDFS. This is because HDFS runs in a separate namespace, isolated from the contents of your local files. The files inside HDFS (or more accurately: the blocks that make them up) are stored in a particular directory managed by the DataNode service, but the files will named only with block ids. You cannot interact with HDFS-stored files using ordinary Linux file modification tools (e.g., ls, cp, mv, etc). However, HDFS does come with its own utilities for file management, which act very similar to these familiar tools. A later section in this tutorial will introduce you to these commands and their operation.

It is important for this file system to store its metadata reliably. Furthermore, while the file data is accessed in a write once and read many model, the metadata structures (e.g., the names of files and directories) can be modified by a large number of clients concurrently. It is important that this information is never desynchronized. Therefore, it is all handled by a single machine, called the NameNode. The NameNode stores all the metadata for the file system. Because of the relatively low amount of metadata per file (it only tracks file names, permissions, and the locations of each block of each file), all of this information can be stored in the main memory of the NameNode machine, allowing fast access to the metadata.

To open a file, a client contacts the NameNode and retrieves a list of locations for the blocks that comprise the file. These locations identify the DataNodes which hold each block. Clients then read file data directly from the DataNode servers, possibly in parallel. The NameNode is not directly involved in this bulk data transfer, keeping its overhead to a minimum.

Of course, NameNode information must be preserved even if the NameNode machine fails; there are multiple redundant systems that allow the NameNode to preserve the file system’s metadata even if the NameNode itself crashes irrecoverably. NameNode failure is more severe for the cluster than DataNode failure. While individual DataNodes may crash and the entire cluster will continue to operate, the loss of the NameNode will render the cluster inaccessible until it is manually restored. Fortunately, as the NameNode’s involvement is relatively minimal, the odds of it failing are considerably lower than the odds of an arbitrary DataNode failing at any given point in time.

A more thorough overview of the architectural decisions involved in the design and implementation of HDFS is given in the official Hadoop HDFS documentation. Before continuing in this tutorial, it is advisable that you read and understand the information presented there.

Configuring HDFS

The HDFS for your cluster can be configured in a very short amount of time. First we will fill out the relevant sections of the Hadoop configuration file, then format the NameNode.


Cluster configuration

These instructions for cluster configuration assume that you have already downloaded and unzipped a copy of Hadoop. Module 3 discusses getting started with Hadoop for this tutorial. Module 7 discusses how to set up a larger cluster and provides preliminary setup instructions for Hadoop, including downloading prerequisite software.

The HDFS configuration is located in a set of XML files in the Hadoop configuration directory; conf/ under the main Hadoop install directory (where you unzipped Hadoop to). The conf/hadoop-defaults.xml file contains default values for every parameter in Hadoop. This file is considered read-only. You override this configuration by setting new values in conf/hadoop-site.xml. This file should be replicated consistently across all machines in the cluster. (It is also possible, though not advisable, to host it on NFS.)

Configuration settings are a set of key-value pairs of the format:



Adding the line <final>true</final> inside the property body will prevent properties from being overridden by user applications. This is useful for most system-wide configuration options.

The following settings are necessary to configure HDFS:

key value example
fs.default.name protocol://servername:port hdfs://alpha.milkman.org:9000
dfs.data.dir pathname /home/username/hdfs/data
dfs.name.dir pathname /home/username/hdfs/name

These settings are described individually below:

fs.default.name – This is the URI (protocol specifier, hostname, and port) that describes the NameNode for the cluster. Each node in the system on which Hadoop is expected to operate needs to know the address of the NameNode. The DataNode instances will register with this NameNode, and make their data available through it. Individual client programs will connect to this address to retrieve the locations of actual file blocks.

dfs.data.dir – This is the path on the local file system in which the DataNode instance should store its data. It is not necessary that all DataNode instances store their data under the same local path prefix, as they will all be on separate machines; it is acceptable that these machines are heterogeneous. However, it will simplify configuration if this directory is standardized throughout the system. By default, Hadoop will place this under /tmp. This is fine for testing purposes, but is an easy way to lose actual data in a production system, and thus must be overridden.

dfs.name.dir – This is the path on the local file system of the NameNode instance where the NameNode metadata is stored. It is only used by the NameNode instance to find its information, and does not exist on the DataNodes. The caveat above about /tmp applies to this as well; this setting must be overridden in a production system.

Another configuration parameter, not listed above, is dfs.replication. This is the default replication factor for each block of data in the file system. For a production cluster, this should usually be left at its default value of 3. (You are free to increase your replication factor, though this may be unnecessary and use more space than is required. Fewer than three replicas impact the high availability of information, and possibly the reliability of its storage.)

The following information can be pasted into the hadoop-site.xml file for a single-node configuration:





Of course, your.server.name.com needs to be changed, as does username. Using port 9000 for the NameNode is arbitrary.

After copying this information into your conf/hadoop-site.xml file, copy this to the conf/ directories on all machines in the cluster.

The master node needs to know the addresses of all the machines to use as DataNodes; the startup scripts depend on this. Also in the conf/ directory, edit the file slaves so that it contains a list of fully-qualified hostnames for the slave instances, one host per line. On a multi-node setup, the master node (e.g., localhost) is not usually present in this file.

Then make the directories necessary:

  user@EachMachine$ mkdir -p $HOME/hdfs/data

  user@namenode$ mkdir -p $HOME/hdfs/name

The user who owns the Hadoop instances will need to have read and write access to each of these directories. It is not necessary for all users to have access to these directories. Set permissions with chmod as appropriate. In a large-scale environment, it is recommended that you create a user named “hadoop” on each node for the express purpose of owning and running Hadoop tasks. For a single individual’s machine, it is perfectly acceptable to run Hadoop under your own username. It is not recommended that you run Hadoop as root.


Starting HDFS

Now we must format the file system that we just configured:

  user@namenode:hadoop$ bin/hadoop namenode -format

This process should only be performed once. When it is complete, we are free to start the distributed file system:

  user@namenode:hadoop$ bin/start-dfs.sh

This command will start the NameNode server on the master machine (which is where the start-dfs.sh script was invoked). It will also start the DataNode instances on each of the slave machines. In a single-machine “cluster,” this is the same machine as the NameNode instance. On a real cluster of two or more machines, this script will ssh into each slave machine and start a DataNode instance.

Interacting With HDFS

This section will familiarize you with the commands necessary to interact with HDFS, loading and retrieving data, as well as manipulating files. This section makes extensive use of the command-line.

The bulk of commands that communicate with the cluster are performed by a monolithic script named bin/hadoop. This will load the Hadoop system with the Java virtual machine and execute a user command. The commands are specified in the following form:

  user@machine:hadoop$ bin/hadoop moduleName -cmd args...

The moduleName tells the program which subset of Hadoop functionality to use. -cmd is the name of a specific command within this module to execute. Its arguments follow the command name.

Two such modules are relevant to HDFS: dfs and dfsadmin. Their use is described in the sections below.

Common Example Operations

The dfs module, also known as “FsShell,” provides basic file manipulation operations. Their usage is introduced here.

A cluster is only useful if it contains data of interest. Therefore, the first operation to perform is loading information into the cluster. For purposes of this example, we will assume an example user named “someone” — but substitute your own username where it makes sense. Also note that any operation on files in HDFS can be performed from any node with access to the cluster, whose conf/hadoop-site.xml is configured to set fs.default.name to your cluster’s NameNode. We will call the fictional machine on which we are operating anynode. Commands are being run from the “hadoop” directory where you installed Hadoop. This may be /home/someone/src/hadoop on your machine, or /home/foo/hadoop on someone else’s. These initial commands are centered around loading information into HDFS, checking that it’s there, and getting information back out of HDFS.

Listing files

If we attempt to inspect HDFS, we will not find anything interesting there:

  someone@anynode:hadoop$ bin/hadoop dfs -ls

The “-ls” command returns silently. Without any arguments, -ls will attempt to show the contents of your “home” directory inside HDFS. Don’t forget, this is not the same as /home/$USER (e.g., /home/someone) on the host machine (HDFS keeps a separate namespace from the local files). There is no concept of a “current working directory” or cd command in HDFS.

If you provide -ls with an argument, you may see some initial directory contents:

  someone@anynode:hadoop$ bin/hadoop dfs -ls /
  Found 2 items
  drwxr-xr-x   - hadoop supergroup          0 2008-09-20 19:40 /hadoop
  drwxr-xr-x   - hadoop supergroup          0 2008-09-20 20:08 /tmp

These entries are created by the system. This example output assumes that “hadoop” is the username under which the Hadoop daemons (NameNode, DataNode, etc) were started. “supergroup” is a special group whose membership includes the username under which the HDFS instances were started (e.g., “hadoop”). These directories exist to allow the Hadoop MapReduce system to move necessary data to the different job nodes; this is explained in more detail in Module 4.

So we need to create our home directory, and then populate it with some files.

class=”sectionSubH”>Inserting data into the cluster

Whereas a typical UNIX or Linux system stores individual users’ files in /home/$USER, the Hadoop DFS stores these in /user/$USER. For some commands like ls, if a directory name is required and is left blank, this is the default directory name assumed. (Other commands require explicit source and destination paths.) Any relative paths used as arguments to HDFS, Hadoop MapReduce, or other components of the system are assumed to be relative to this base directory.

Step 1: Create your home directory if it does not already exist.

  someone@anynode:hadoop$ bin/hadoop dfs -mkdir /user

If there is no /user directory, create that first. It will be automatically created later if necessary, but for instructive purposes, it makes sense to create it manually ourselves this time.

Then we are free to add our own home directory:

  someone@anynode:hadoop$ bin/hadoop dfs -mkdir /user/someone

Of course, replace /user/someone with /user/yourUserName.

Step 2: Upload a file. To insert a single file into HDFS, we can use the put command like so:

  someone@anynode:hadoop$ bin/hadoop dfs -put /home/someone/interestingFile.txt /user/yourUserName/

This copies /home/someone/interestingFile.txt from the local file system into /user/yourUserName/interestingFile.txt on HDFS.

Step 3: Verify the file is in HDFS. We can verify that the operation worked with either of the two following (equivalent) commands:

  someone@anynode:hadoop$ bin/hadoop dfs -ls /user/yourUserName
  someone@anynode:hadoop$ bin/hadoop dfs -ls

You should see a listing that starts with Found 1 items and then includes information about the file you inserted.

The following table demonstrates example uses of the put command, and their effects:

Command: Assuming: Outcome:
bin/hadoop dfs -put foo bar No file/directory named /user/$USER/bar exists in HDFS Uploads local file foo to a file named /user/$USER/bar
bin/hadoop dfs -put foo bar /user/$USER/bar is a directory Uploads local file foo to a file named /user/$USER/bar/foo
bin/hadoop dfs -put foo somedir/somefile /user/$USER/somedir does not exist in HDFS Uploads local file foo to a file named /user/$USER/somedir/somefile, creating the missing directory
bin/hadoop dfs -put foo bar /user/$USER/bar is already a file in HDFS No change in HDFS, and an error is returned to the user.

When the put command operates on a file, it is all-or-nothing. Uploading a file into HDFS first copies the data onto the DataNodes. When they all acknowledge that they have received all the data and the file handle is closed, it is then made visible to the rest of the system. Thus based on the return value of the put command, you can be confident that a file has either been successfully uploaded, or has “fully failed;” you will never get into a state where a file is partially uploaded and the partial contents are visible externally, but the upload disconnected and did not complete the entire file contents. In a case like this, it will be as though no upload took place.

Step 4: Uploading multiple files at once. The put command is more powerful than moving a single file at a time. It can also be used to upload entire directory trees into HDFS.

Create a local directory and put some files into it using the cp command. Our example user may have a situation like the following:

  someone@anynode:hadoop$ ls -R myfiles
  file1.txt  file2.txt  subdir/


This entire myfiles/ directory can be copied into HDFS like so:

  someone@anynode:hadoop$ bin/hadoop -put myfiles /user/myUsername
  someone@anynode:hadoop$ bin/hadoop -ls
  Found 1 items
  /user/someone/myfiles   <dir>    2008-06-12 20:59    rwxr-xr-x    someone    supergroup
  user@anynode:hadoop bin/hadoop -ls myfiles
  Found 3 items
  /user/someone/myfiles/file1.txt   <r 1>   186731  2008-06-12 20:59        rw-r--r--       someone   supergroup
  /user/someone/myfiles/file2.txt   <r 1>   168     2008-06-12 20:59        rw-r--r--       someone   supergroup
  /user/someone/myfiles/subdir      <dir>           2008-06-12 20:59        rwxr-xr-x       someone   supergroup

Thus demonstrating that the tree was correctly uploaded recursively. You’ll note that in addition to the file path, ls also reports the number of replicas of each file that exist (the “1” in <r 1>), the file size, upload time, permissions, and owner information.

Another synonym for -put is -copyFromLocal. The syntax and functionality are identical.

Retrieving data from HDFS

There are multiple ways to retrieve files from the distributed file system. One of the easiest is to use cat to display the contents of a file on stdout. (It can, of course, also be used to pipe the data into other applications or destinations.)

Step 1: Display data with cat.

If you have not already done so, upload some files into HDFS. In this example, we assume that a file named “foo” has been loaded into your home directory on HDFS.

  someone@anynode:hadoop$ bin/hadoop dfs -cat foo
  (contents of foo are displayed here)

Step 2: Copy a file from HDFS to the local file system.

The get command is the inverse operation of put; it will copy a file or directory (recursively) from HDFS into the target of your choosing on the local file system. A synonymous operation is called -copyToLocal.

  someone@anynode:hadoop$ bin/hadoop dfs -get foo localFoo
  someone@anynode:hadoop$ ls
  someone@anynode:hadoop$ cat localFoo
  (contents of foo are displayed here)

Like the put command, get will operate on directories in addition to individual files.

Shutting Down HDFS

If you want to shut down the HDFS functionality of your cluster (either because you do not want Hadoop occupying memory resources when it is not in use, or because you want to restart the cluster for upgrading, configuration changes, etc.), then this can be accomplished by logging in to the NameNode machine and running:

  someone@namenode:hadoop$ bin/stop-dfs.sh

This command must be performed by the same user who started HDFS with bin/start-dfs.sh.

HDFS Command Reference

There are many more commands in bin/hadoop dfs than were demonstrated here, although these basic operations will get you started. Running bin/hadoop dfs with no additional arguments will list all commands which can be run with the FsShell system. Furthermore, bin/hadoop dfs -help commandName will display a short usage summary for the operation in question, if you are stuck.

A table of all operations is reproduced below. The following conventions are used for parameters:

  • italics denote variables to be filled out by the user.
  • “path” means any file or directory name.
  • “path…” means one or more file or directory names.
  • “file” means any filename.
  • “src” and “dest” are path names in a directed operation.
  • “localSrc” and “localDest” are paths as above, but on the local file system. All other file and path names refer to objects inside HDFS.
  • Parameters in [brackets] are optional.
Command Operation
-ls path Lists the contents of the directory specified by path, showing the names, permissions, owner, size and modification date for each entry.
-lsr path Behaves like -ls, but recursively displays entries in all subdirectories of path.
-du path Shows disk usage, in bytes, for all files which match path; filenames are reported with the full HDFS protocol prefix.
-dus path Like -du, but prints a summary of disk usage of all files/directories in the path.
-mv src dest Moves the file or directory indicated by src to dest, within HDFS.
-cp src dest Copies the file or directory identified by src to dest, within HDFS.
-rm path Removes the file or empty directory identified by path.
-rmr path Removes the file or directory identified by path. Recursively deletes any child entries (i.e., files or subdirectories of path).
-put localSrc dest Copies the file or directory from the local file system identified by localSrc to dest within the DFS.
-copyFromLocal localSrc dest Identical to -put
-moveFromLocal localSrc dest Copies the file or directory from the local file system identified by localSrc to dest within HDFS, then deletes the local copy on success.
-get [-crc] src localDest Copies the file or directory in HDFS identified by src to the local file system path identified by localDest.
-getmerge src localDest [addnl] Retrieves all files that match the path src in HDFS, and copies them to a single, merged file in the local file system identified by localDest.
-cat filename Displays the contents of filename on stdout.
-copyToLocal [-crc] src localDest Identical to -get
-moveToLocal [-crc] src localDest Works like -get, but deletes the HDFS copy on success.
-mkdir path Creates a directory named path in HDFS. Creates any parent directories in path that are missing (e.g., like mkdir -p in Linux).
-setrep [-R] [-w] rep path Sets the target replication factor for files identified by path to rep. (The actual replication factor will move toward the target over time)
-touchz path Creates a file at path containing the current time as a timestamp. Fails if a file already exists at path, unless the file is already size 0.
-test -[ezd] path Returns 1 if path exists; has zero length; or is a directory, or 0 otherwise.
-stat [format] path Prints information about path. format is a string which accepts file size in blocks (%b), filename (%n), block size (%o), replication (%r), and modification date (%y, %Y).
-tail [-f] file Shows the lats 1KB of file on stdout.
-chmod [-R] mode,mode,… path… Changes the file permissions associated with one or more objects identified by path…. Performs changes recursively with -R. mode is a 3-digit octal mode, or {augo}+/-{rwxX}. Assumes a if no scope is specified and does not apply a umask.
-chown [-R] [owner][:[group]] path… Sets the owning user and/or group for files or directories identified by path…. Sets owner recursively if -R is specified.
-chgrp [-R] group path… Sets the owning group for files or directories identified by path…. Sets group recursively if -R is specified.
-help cmd Returns usage information for one of the commands listed above. You must omit the leading ‘-‘ character in cmd

DFSAdmin Command Reference

While the dfs module for bin/hadoop provides common file and directory manipulation commands, they all work with objects within the file system. The dfsadmin module manipulates or queries the file system as a whole. The operation of the commands in this module is described in this section.

Getting overall status: A brief status report for HDFS can be retrieved with bin/hadoop dfsadmin -report. This returns basic information about the overall health of the HDFS cluster, as well as some per-server metrics.

More involved status: If you need to know more details about what the state of the NameNode’s metadata is, the command bin/hadoop dfsadmin -metasave filename will record this information in filename. The metasave command will enumerate lists of blocks which are under-replicated, in the process of being replicated, and scheduled for deletion. NB: The help for this command states that it “saves NameNode’s primary data structures,” but this is a misnomer; the NameNode’s state cannot be restored from this information. However, it will provide good information about how the NameNode is managing HDFS’s blocks.

Safemode: Safemode is an HDFS state in which the file system is mounted read-only; no replication is performed, nor can files be created or deleted. This is automatically entered as the NameNode starts, to allow all DataNodes time to check in with the NameNode and announce which blocks they hold, before the NameNode determines which blocks are under-replicated, etc. The NameNode waits until a specific percentage of the blocks are present and accounted-for; this is controlled in the configuration by the dfs.safemode.threshold.pct parameter. After this threshold is met, safemode is automatically exited, and HDFS allows normal operations. The bin/hadoop dfsadmin -safemode what command allows the user to manipulate safemode based on the value of what, described below:

  • enter – Enters safemode
  • leave – Forces the NameNode to exit safemode
  • get – Returns a string indicating whether safemode is ON or OFF
  • wait – Waits until safemode has exited and returns

Changing HDFS membership – When decommissioning nodes, it is important to disconnect nodes from HDFS gradually to ensure that data is not lost. See the section on decommissioning later in this document for an explanation of the use of the -refreshNodes dfsadmin command.

Upgrading HDFS versions – When upgrading from one version of Hadoop to the next, the file formats used by the NameNode and DataNodes may change. When you first start the new version of Hadoop on the cluster, you need to tell Hadoop to change the HDFS version (or else it will not mount), using the command: bin/start-dfs.sh -upgrade. It will then begin upgrading the HDFS version. The status of an ongoing upgrade operation can be queried with the bin/hadoop dfsadmin -upgradeProgress status command. More verbose information can be retrieved with bin/hadoop dfsadmin -upgradeProgress details. If the upgrade is blocked and you would like to force it to continue, use the command: bin/hadoop dfsadmin -upgradeProgress force. (Note: be sure you know what you are doing if you use this last command.)

When HDFS is upgraded, Hadoop retains backup information allowing you to downgrade to the original HDFS version in case you need to revert Hadoop versions. To back out the changes, stop the cluster, re-install the older version of Hadoop, and then use the command: bin/start-dfs.sh -rollback. It will restore the previous HDFS state.

Only one such archival copy can be kept at a time. Thus, after a few days of operation with the new version (when it is deemed stable), the archival copy can be removed with the command bin/hadoop dfsadmin -finalizeUpgrade. The rollback command cannot be issued after this point. This must be performed before a second Hadoop upgrade is allowed.

Getting help – As with the dfs module, typing bin/hadoop dfsadmin -help cmd will provide more usage information about the particular command.

Using HDFS in MapReduce

The HDFS is a powerful companion to Hadoop MapReduce. By setting the fs.default.name configuration option to point to the NameNode (as was done above), Hadoop MapReduce jobs will automatically draw their input files from HDFS. Using the regular FileInputFormat subclasses, Hadoop will automatically draw its input data sources from file paths within HDFS, and will distribute the work over the cluster in an intelligent fashion to exploit block locality where possible. The mechanics of Hadoop MapReduce are discussed in much greater detail in Module 4.

Using HDFS Programmatically

While HDFS can be manipulated explicitly through user commands, or implicitly as the input to or output from a Hadoop MapReduce job, you can also work with HDFS inside your own Java applications. (A JNI-based wrapper, libhdfs also provides this functionality in C/C++ programs.)

This section provides a short tutorial on using the Java-based HDFS API. It will be based on the following code listing:

1:  import java.io.File;
2:  import java.io.IOException;
4:  import org.apache.hadoop.conf.Configuration;
5:  import org.apache.hadoop.fs.FileSystem;
6:  import org.apache.hadoop.fs.FSDataInputStream;
7:  import org.apache.hadoop.fs.FSDataOutputStream;
8:  import org.apache.hadoop.fs.Path;
10: public class HDFSHelloWorld {
12:   public static final String theFilename = "hello.txt";
13:   public static final String message = "Hello, world!\n";
15:   public static void main (String [] args) throws IOException {
17:     Configuration conf = new Configuration();
18:     FileSystem fs = FileSystem.get(conf);
20:     Path filenamePath = new Path(theFilename);
22:     try {
23:       if (fs.exists(filenamePath)) {
24:         // remove the file first
25:         fs.delete(filenamePath);
26:       }
28:       FSDataOutputStream out = fs.create(filenamePath);
29:       out.writeUTF(message;
30:       out.close();
32:       FSDataInputStream in = fs.open(filenamePath);
33:       String messageIn = in.readUTF();
34:       System.out.print(messageIn);
35:       in.close();
46:     } catch (IOException ioe) {
47:       System.err.println("IOException during operation: " + ioe.toString());
48:       System.exit(1);
49:     }
40:   }
41: }

This program creates a file named hello.txt, writes a short message into it, then reads it back and prints it to the screen. If the file already existed, it is deleted first.

First we get a handle to an abstract FileSystem object, as specified by the application configuration. The Configuration object created uses the default parameters.

17:     Configuration conf = new Configuration();
18:     FileSystem fs = FileSystem.get(conf);

The FileSystem interface actually provides a generic abstraction suitable for use in several file systems. Depending on the Hadoop configuration, this may use HDFS or the local file system or a different one altogether. If this test program is launched via the ordinary ‘java classname‘ command line, it may not find conf/hadoop-site.xml and will use the local file system. To ensure that it uses the proper Hadoop configuration, launch this program through Hadoop by putting it in a jar and running:

$HADOOP_HOME/bin/hadoop jar yourjar HDFSHelloWorld

Regardless of how you launch the program and which file system it connects to, writing to a file is done in the same way:

28:       FSDataOutputStream out = fs.create(filenamePath);
29:       out.writeUTF(message);
30:       out.close();

First we create the file with the fs.create() call, which returns an FSDataOutputStream used to write data into the file. We then write the information using ordinary stream writing functions; FSDataOutputStream extends the java.io.DataOutputStream class. When we are done with the file, we close the stream with out.close().

This call to fs.create() will overwrite the file if it already exists, but for sake of example, this program explicitly removes the file first anyway (note that depending on this explicit prior removal is technically a race condition). Testing for whether a file exists and removing an existing file are performed by lines 23-26:

23:       if (fs.exists(filenamePath)) {
24:         // remove the file first
25:         fs.delete(filenamePath);
26:       }

Other operations such as copying, moving, and renaming are equally straightforward operations on Path objects performed by the FileSystem.

Finally, we re-open the file for read, and pull the bytes from the file, converting them to a UTF-8 encoded string in the process, and print to the screen:

32:       FSDataInputStream in = fs.open(filenamePath);
33:       String messageIn = in.readUTF();
34:       System.out.print(messageIn);
35:       in.close();

The fs.open() method returns an FSDataInputStream, which subclasses java.io.DataInputStream. Data can be read from the stream using the readUTF() operation, as on line 33. When we are done with the stream, we call close() to free the handle associated with the file.

More information:

Complete JavaDoc for the HDFS API is provided at http://hadoop.apache.org/common/docs/r0.20.2/api/index.html.

A direct link to the FileSystem interface is: http://hadoop.apache.org/common/docs/r0.20.2/api/org/apache/hadoop/fs/FileSystem.html.

Another example HDFS application is available on the Hadoop wiki. This implements a file copy operation.

HDFS Permissions and Security

Starting with Hadoop 0.16.1, HDFS has included a rudimentary file permissions system. This permission system is based on the POSIX model, but does not provide strong security for HDFS files. The HDFS permissions system is designed to prevent accidental corruption of data or casual misuse of information within a group of users who share access to a cluster. It is not a strong security model that guarantees denial of access to unauthorized parties.

HDFS security is based on the POSIX model of users and groups. Each file or directory has 3 permissions (read, write and execute) associated with it at three different granularities: the file’s owner, users in the same group as the owner, and all other users in the system. As the HDFS does not provide the full POSIX spectrum of activity, some combinations of bits will be meaningless. For example, no file can be executed; the +x bits cannot be set on files (only directories). Nor can an existing file be written to, although the +w bits may still be set.

Security permissions and ownership can be modified using the bin/hadoop dfs -chmod, -chown, and -chgrp operations described earlier in this document; they work in a similar fashion to the POSIX/Linux tools of the same name.

Determining identity – Identity is not authenticated formally with HDFS; it is taken from an extrinsic source. The Hadoop system is programmed to use the user’s current login as their Hadoop username (i.e., the equivalent of whoami). The user’s current working group list (i.e, the output of groups) is used as the group list in Hadoop. HDFS itself does not verify that this username is genuine to the actual operator.

Superuser status – The username which was used to start the Hadoop process (i.e., the username who actually ran bin/start-all.sh or bin/start-dfs.sh) is acknowledged to be the superuser for HDFS. If this user interacts with HDFS, he does so with a special username superuser. This user’s operations on HDFS never fail, regardless of permission bits set on the particular files he manipulates. If Hadoop is shutdown and restarted under a different username, that username is then bound to the superuser account.

Supergroup – There is also a special group named supergroup, whose membership is controlled by the configuration parameter dfs.permissions.supergroup.

Disabling permissions – By default, permissions are enabled on HDFS. The permission system can be disabled by setting the configuration option dfs.permissions to false. The owner, group, and permissions bits associated with each file and directory will still be preserved, but the HDFS process does not enforce them, except when using permissions-related operations such as -chmod.

Additional HDFS Tasks

Rebalancing Blocks

New nodes can be added to a cluster in a straightforward manner. On the new node, the same Hadoop version and configuration (conf/hadoop-site.xml) as on the rest of the cluster should be installed. Starting the DataNode daemon on the machine will cause it to contact the NameNode and join the cluster. (The new node should be added to the slaves file on the master server as well, to inform the master how to invoke script-based commands on the new node.)

But the new DataNode will have no data on board initially; it is therefore not alleviating space concerns on the existing nodes. New files will be stored on the new DataNode in addition to the existing ones, but for optimum usage, storage should be evenly balanced across all nodes.

This can be achieved with the automatic balancer tool included with Hadoop. The Balancer class will intelligently balance blocks across the nodes to achieve an even distribution of blocks within a given threshold, expressed as a percentage. (The default is 10%.) Smaller percentages make nodes more evenly balanced, but may require more time to achieve this state. Perfect balancing (0%) is unlikely to actually be achieved.

The balancer script can be run by starting bin/start-balancer.sh in the Hadoop directory. The script can be provided a balancing threshold percentage with the -threshold parameter; e.g., bin/start-balancer.sh -threshold 5. The balancer will automatically terminate when it achieves its goal, or when an error occurs, or it cannot find more candidate blocks to move to achieve better balance. The balancer can always be terminated safely by the administrator by running bin/stop-balancer.sh.

The balancing script can be run either when nobody else is using the cluster (e.g., overnight), but can also be run in an “online” fashion while many other jobs are on-going. To prevent the rebalancing process from consuming large amounts of bandwidth and significantly degrading the performance of other processes on the cluster, the dfs.balance.bandwidthPerSec configuration parameter can be used to limit the number of bytes/sec each node may devote to rebalancing its data store.

Copying Large Sets of Files

When migrating a large number of files from one location to another (either from one HDFS cluster to another, from S3 into HDFS or vice versa, etc), the task should be divided between multiple nodes to allow them all to share in the bandwidth required for the process. Hadoop includes a tool called distcp for this purpose.

By invoking bin/hadoop distcp src dest, Hadoop will start a MapReduce task to distribute the burden of copying a large number of files from src to dest. These two parameters may specify a full URL for the the path to copy. e.g., "hdfs://SomeNameNode:9000/foo/bar/" and "hdfs://OtherNameNode:2000/baz/quux/" will copy the children of /foo/bar on one cluster to the directory tree rooted at /baz/quux on the other. The paths are assumed to be directories, and are copied recursively. S3 URLs can be specified with s3://bucket-name/key.

Decommissioning Nodes

In addition to allowing nodes to be added to the cluster on the fly, nodes can also be removed from a cluster while it is running, without data loss. But if nodes are simply shut down “hard,” data loss may occur as they may hold the sole copy of one or more file blocks.

Nodes must be retired on a schedule that allows HDFS to ensure that no blocks are entirely replicated within the to-be-retired set of DataNodes.

HDFS provides a decommissioning feature which ensures that this process is performed safely. To use it, follow the steps below:

Step 1: Cluster configuration. If it is assumed that nodes may be retired in your cluster, then before it is started, an excludes file must be configured. Add a key named dfs.hosts.exclude to your conf/hadoop-site.xml file. The value associated with this key provides the full path to a file on the NameNode’s local file system which contains a list of machines which are not permitted to connect to HDFS.

Step 2: Determine hosts to decommission. Each machine to be decommissioned should be added to the file identified by dfs.hosts.exclude, one per line. This will prevent them from connecting to the NameNode.

Step 3: Force configuration reload. Run the command bin/hadoop dfsadmin -refreshNodes. This will force the NameNode to reread its configuration, including the newly-updated excludes file. It will decommission the nodes over a period of time, allowing time for each node’s blocks to be replicated onto machines which are scheduled to remain active.

Step 4: Shutdown nodes. After the decommission process has completed, the decommissioned hardware can be safely shutdown for maintenance, etc. The bin/hadoop dfsadmin -report command will describe which nodes are connected to the cluster.

Step 5: Edit excludes file again. Once the machines have been decommissioned, they can be removed from the excludes file. Running bin/hadoop dfsadmin -refreshNodes again will read the excludes file back into the NameNode, allowing the DataNodes to rejoin the cluster after maintenance has been completed, or additional capacity is needed in the cluster again, etc.

Verifying File System Health

After decommissioning nodes, restarting a cluster, or periodically during its lifetime, you may want to ensure that the file system is healthy–that files are not corrupted or under-replicated, and that blocks are not missing.

Hadoop provides an fsck command to do exactly this. It can be launched at the command line like so:

  bin/hadoop fsck [path] [options]

If run with no arguments, it will print usage information and exit. If run with the argument /, it will check the health of the entire file system and print a report. If provided with a path to a particular directory or file, it will only check files under that path. If an option argument is given but no path, it will start from the file system root (/). The options may include two different types of options:

Action options specify what action should be taken when corrupted files are found. This can be -move, which moves corrupt files to /lost+found, or -delete, which deletes corrupted files.

Information options specify how verbose the tool should be in its report. The -files option will list all files it checks as it encounters them. This information can be further expanded by adding the -blocks option, which prints the list of blocks for each file. Adding -locations to these two options will then print the addresses of the DataNodes holding these blocks. Still more information can be retrieved by adding -racks to the end of this list, which then prints the rack topology information for each location. (See the next subsection for more information on configuring network rack awareness.) Note that the later options do not imply the former; you must use them in conjunction with one another. Also, note that the Hadoop program uses -files in a “common argument parser” shared by the different commands such as dfsadmin, fsck, dfs, etc. This means that if you omit a path argument to fsck, it will not receive the -files option that you intend. You can separate common options from fsck-specific options by using -- as an argument, like so:

  bin/hadoop fsck -- -files -blocks

The -- is not required if you provide a path to start the check from, or if you specify another argument first such as -move.

By default, fsck will not operate on files still open for write by another client. A list of such files can be produced with the -openforwrite option.

Rack Awareness

For small clusters in which all servers are connected by a single switch, there are only two levels of locality: “on-machine” and “off-machine.” When loading data from a DataNode’s local drive into HDFS, the NameNode will schedule one copy to go into the local DataNode, and will pick two other machines at random from the cluster.

For larger Hadoop installations which span multiple racks, it is important to ensure that replicas of data exist on multiple racks. This way, the loss of a switch does not render portions of the data unavailable due to all replicas being underneath it.

HDFS can be made rack-aware by the use of a script which allows the master node to map the network topology of the cluster. While alternate configuration strategies can be used, the default implementation allows you to provide an executable script which returns the “rack address” of each of a list of IP addresses.

The network topology script receives as arguments one or more IP addresses of nodes in the cluster. It returns on stdout a list of rack names, one for each input. The input and output order must be consistent.

To set the rack mapping script, specify the key topology.script.file.name in conf/hadoop-site.xml. This provides a command to run to return a rack id; it must be an executable script or program. By default, Hadoop will attempt to send a set of IP addresses to the file as several separate command line arguments. You can control the maximum acceptable number of arguments with the topology.script.number.args key.

Rack ids in Hadoop are hierarchical and look like path names. By default, every node has a rack id of /default-rack. You can set rack ids for nodes to any arbitrary path, e.g., /foo/bar-rack. Path elements further to the left are higher up the tree. Thus a reasonable structure for a large installation may be /top-switch-name/rack-name.

Hadoop rack ids are not currently expressive enough to handle an unusual routing topology such as a 3-d torus; they assume that each node is connected to a single switch which in turn has a single upstream switch. This is not usually a problem, however. Actual packet routing will be directed using the topology discovered by or set in switches and routers. The Hadoop rack ids will be used to find “near” and “far” nodes for replica placement (and in 0.17, MapReduce task placement).

The following example script performs rack identification based on IP addresses given a hierarchical IP addressing scheme enforced by the network administrator. This may work directly for simple installations; more complex network configurations may require a file- or table-based lookup process. Care should be taken in that case to keep the table up-to-date as nodes are physically relocated, etc. This script requires that the maximum number of arguments be set to 1.

# Set rack id based on IP address.
# Assumes network administrator has complete control
# over IP addresses assigned to nodes and they are
# in the 10.x.y.z address space. Assumes that
# IP addresses are distributed hierarchically. e.g.,
# 10.1.y.z is one data center segment and 10.2.y.z is another;
# 10.1.1.z is one rack, 10.1.2.z is another rack in
# the same segment, etc.)
# This is invoked with an IP address as its only argument

# get IP address from the input

# select "x.y" and convert it to "x/y"
segments=`echo $ipaddr | cut --delimiter=. --fields=2-3 --output-delimiter=/`
echo /${segments}

HDFS Web Interface

HDFS exposes a web server which is capable of performing basic status monitoring and file browsing operations. By default this is exposed on port 50070 on the NameNode. Accessing http://namenode:50070/ with a web browser will return a page containing overview information about the health, capacity, and usage of the cluster (similar to the information returned by bin/hadoop dfsadmin -report).

The address and port where the web interface listens can be changed by setting dfs.http.address in conf/hadoop-site.xml. It must be of the form address:port. To accept requests on all addresses, use

From this interface, you can browse HDFS itself with a basic file-browser interface. Each DataNode exposes its file browser interface on port 50075. You can override this by setting the dfs.datanode.http.address configuration key to a setting other than Log files generated by the Hadoop daemons can be accessed through this interface, which is useful for distributed debugging and troubleshooting.


Ghemawat, S. Gobioff, H. and Leung, S.-T. The Google File System. Proceedings of the 19th ACM Symposium on Operating Systems Principles. pp 29–43. Bolton Landing, NY, USA. 2003. © 2003, ACM.
Borthakur, Dhruba. The Hadoop Distributed File System: Architecture and Design. © 2007, The Apache Software Foundation.
Hadoop DFS User Guide. © 2007, The Apache Software Foundation.
HDFS: Permissions User and Administrator Guide. © 2007, The Apache Software Foundation.
HDFS API Javadoc © 2008, The Apache Software Foundation.

Module 3: Getting Started With Hadoop


Hadoop is an open source implementation of the MapReduce platform and distributed file system, written in Java. This module explains the basics of how to begin using Hadoop to experiment and learn from the rest of this tutorial. It covers setting up the platform and connecting other tools to use it.

Goals for this Module:

  • Set up a pre-configured Hadoop virtual machine
  • Verify that you can connect to the virtual machine
  • Understand tools available to help you use Hadoop


  1. Introduction
  2. Goals for this Module
  3. Outline
  4. Prerequisites
  5. A Virtual Machine Hadoop Environment
    1. Installing VMware Player
    2. Setting up the Virtual Environment
    3. Virtual Machine User Accounts
    4. Running a Hadoop Job
    5. Accessing the VM via ssh
    6. Shutting Down the VM
  6. Getting Started With Eclipse
    1. Downloading and Installing
    2. Installing the Hadoop MapReduce Plugin
    3. Making a Copy of Hadoop
    4. Running Eclipse
    5. Configuring the MapReduce Plugin
  7. Interacting With HDFS
    1. Using the Command Line
    2. Using the MapReduce Plugin For Eclipse
  8. Running a Sample Program
    1. Creating the Project
    2. Creating the Source Files
    3. Launching the Job
  9. References & Resources
  10. Complete Tools List


Developing for Hadoop requires a Java programming environment. You can download a Java Development Kit (JDK) for a wide variety of operating systems from http://java.sun.com. Hadoop requires the Java Standard Edition (Java SE), version 6, which is the most current version at the time of this writing.

A Virtual Machine Hadoop Environment

This section explains how to configure a virtual machine to run Hadoop within your host computer. After installing the virtual machine software and the virtual machine image, you will learn how to log in and run jobs within the Hadoop environment.

Users of Linux, Mac OSX, or other Unix-like environments are able to install Hadoop and run it on one (or more) machines with no additional software beyond Java. If you are interested in doing this, there are instructions available on the Hadoop web site in the Getting Started document.

Running Hadoop on top of Windows requires installing cygwin, a Linux-like environment that runs within Windows. Hadoop works reasonably well on cygwin, but it is officially for “development purposes only.” Hadoop on cygwin may be unstable, and installing cygwin itself can be cumbersome.

To aid developers in getting started easily with Hadoop, we have provided a virtual machine image containing a preconfigured Hadoop installation. The virtual machine image will run inside of a “sandbox” environment in which we can run another operating system. The OS inside the sandbox does not know that there is another operating environment outside of it; it acts as though it is on its own computer. This sandbox environment is referred to as the “guest machine” running a “guest operating system.” The actual physical machine running the VM software is referred to as the “host machine” and it runs the “host operating system.” The virtual machine provides other host-machine applications with the appearance that another physical computer is available on the same network. Applications running on the host machine see the VM as a separate machine with its own IP address, and can interact with the programs inside the VM in this fashion.


Figure 3.1: A virtual machine encapsulates one operating system within another. Applications in the VM believe they run on a separate physical host from other applications in the external operating system. Here we demonstrate a Windows host machine and a Linux guest (virtual) machine.

Application developers do not need to use the virtual machine to run Hadoop. Developers on Linux typically use Hadoop in their native development environment, and Windows users often install cygwin for Hadoop development. The virtual machine provided with this tutorial allows users a convenient alternative development platform with a minimum of configuration required. Another advantage of the virtual machine is its easy reset functionality. If your experiments break the Hadoop configuration or render the operating system unusable, you can always simply copy the virtual machine image from the CD back to where you installed it on your computer, and start from a known-good state.

Our virtual machine will run Linux, and comes preconfigured to run Hadoop in pseudo-distributed mode on this system. (It is configured like a fully distributed system, but is actually running on a single machine instance.) We can write Hadoop programs using editors and other applications of the host platform, and run them on our “cluster” consisting of just the virtual machine. We will connect our host environment to the virtual machine through the network.

It should be noted that the virtual machine will also run inside of another instance of Linux. Linux users can install the virtual machine software and run the Hadoop VM as well; the same separation between host processes and guest processes applies here.

Installing VMware Player

The virtual machine is designed to run inside of the VMware Player. A copy of the VMware player installer (version 2.5) for both 32-bit Windows and Linux is included here (linux-rpm, linux-bundle, windows-exe). A Getting Started guide for VMware player provides instructions for installing the VMware player. Review the license information for VMware player before using it..

If you are running on a different operating system, or would prefer to download a more recent version of the player, an alternate installation strategy is to navigate to http://info.vmware.com/content/GLP_VMwarePlayer. You will need to register for a “virtualization starter kit.” You will receive an email with a link to “Download VMware Player.” Click the link, then click the “download now” button at the top of the screen under “most recent version” and follow the instructions. VMware Player is available for Windows or Linux. The latter is available in both 32- and 64-bit versions.

VMware Player itself is approximately a 170 MB download. When the download has completed, run the installer program to set up VMware Player, and follow the prompts as directed. Installation in Windows is performed by a typical Windows installation process.

Setting up the Virtual Environment

Next, copy the Hadoop Virtual Machine into a location on your hard drive. It is a zipped vmware folder (hadoop-vm-appliance-0-18-0) which includes a few files; a .vmdk file that is a snapshot of the virtual machine’s hard drive, and a .vmx file which contains the configuration information to start the virtual machine. After unzipping the vmware folder zip file, to start the virtual machine, double-click on the hadoop-appliance-0.18.0.vmx file in Windows Explorer.


Figure 3.2: When you start the virtual machine for the first time, tell VMware Player that you have copied the VM image.

When you start the virtual machine for the first time, VMware Player will recognize that the virtual machine image is not in the same location it used to be. You should inform VMware Player that you copied this virtual machine image. VMware Player will then generate new session identifiers for this instance of the virtual machine. If you later move the VM image to a different location on your own hard drive, you should tell VMware Player that you have moved the image.

If you ever corrupt the VM image (e.g., by inadvertently deleting or overwriting important files), you can always restore a pristine copy of the virtual machine by copying a fresh VM image off of this tutorial CD. (So don’t be shy about exploring! You can always reset it to a functioning state.)

After you select this option and click OK, the virtual machine should begin booting normally. You will see it perform the standard boot procedure for a Linux system. It will bind itself to an IP address on an unused network segment, and then display a prompt allowing a user to log in.

Virtual Machine User Accounts

The virtual machine comes preconfigured with two user accounts: “root” and “hadoop-user”. The hadoop-user account has sudo permissions to perform system management functions, such as shutting down the virtual machine. The vast majority of your interaction with the virtual machine will be as hadoop-user.

To log in as hadoop-user, first click inside the virtual machine’s display. The virtual machine will take control of your keyboard and mouse. To escape back into Windows at any time, press CTRL+ALT at the same time. The hadoop-user user’s password is hadoop. To log in as root, the password is root.

Running a Hadoop Job

Now that the VM is started, or you have installed Hadoop on your own system in pseudo-distributed mode, let us make sure that Hadoop is properly configured.

If you are using the VM, log in as hadoop-user, as directed above. You will start in your home directory: /home/hadoop-user. Typing ls, you will see a directory named hadoop/, as well as a set of scripts to manage the server. The virtual machine’s hostname is hadoop-desk.

First, we must start the Hadoop system. Type the following command:

hadoop-user@hadoop-desk:~$ ./start-hadoop

If you installed Hadoop on your host system, use the following commands to launch hadoop (assuming you installed to ~/hadoop):

you@your-machine:~$ cd hadoop
you@your-machine:~/hadoop$ bin/start-all.sh

You will see a set of status messages appear as the services boot. If prompted whether it is okay to connect to the current host, type “yes”. Try running an example program to ensure that Hadoop is correctly configured:

hadoop-user@hadoop-desk:~$ cd hadoop
hadoop-user@hadoop-desk:~/hadoop$ bin/hadoop jar hadoop-0.18.0-examples.jar pi 10 1000000

This should provide output that looks something like this:

Wrote input for Map #1
Wrote input for Map #2
Wrote input for Map #3
Wrote input for Map #10
Starting Job
INFO mapred.FileInputFormat: Total input paths to process: 10
INFO mapred.JobClient: Running job: job_200806230804_0001
INFO mapred.JobClient: map 0% reduce 0%
INFO mapred.JobClient: map 10% reduce 0%
INFO mapred.JobClient: map 100% reduce 100%
INFO mapred.JobClient: Job complete: job_200806230804_0001
Job Finished in 25.841 second
Estimated value of PI is 3.141688

This task runs a simulation to estimate the value of pi based on sampling. The test first wrote out a number of points to a list of files, one per map task. It then calculated an estimate of pi based on these points, in the MapReduce task itself. How MapReduce works and how to write such a program are discussed in the next module. The Hadoop client program you used to launch the pi test launched the job, displayed some progress update information as to how the job is proceeding, and then displayed some final performance counters and the job-specific output: an estimate for the value of pi.

Accessing the VM via ssh

Rather than directly use the terminal of the virtual machine, you can also log in “remotely” over ssh from the host environment. Using an ssh client like putty (in Windows), log in with username “hadoop-user” (password hadoop) to the IP address displayed in the virtual machine terminal when it starts up. You can now interact with this virtual machine as if it were another Linux machine on the network.

This can only be done from the host machine. The VMware image is, by default, configured to use host-only networking; only the host machine can talk to the virtual machine over its network interface. The virtual machine does not appear on the actual external network. This is done for security purposes.

If you need to find the virtual machine’s IP address later, the ifconfig command will display this under the “inet addr” field.

Important security note: In the VMware settings, you can reconfigure the virtual machine for networked access rather than host-only networking. If you enable network access, you can access the virtual machine from anywhere else on the network via its IP address. In this case, you should change the passwords associated with the accounts on the virtual machine to prevent unauthorized users from logging in with the default password.

Shutting Down the VM

When you are done with the virtual machine, you can turn it off by logging in as hadoop-user and typing sudo poweroff. The virtual machine will shut itself down in an orderly fashion and the window it runs in will disappear.

Getting Started With Eclipse

A powerful development environment for Java-based programming is Eclipse. Eclipse is a free, open-source IDE. It supports multiple languages through a plugin interface, with special attention paid to Java. Tools designed for working with Hadoop can be integrated into Eclipse, making it an attractive platform for Hadoop development. In this section we will review how to obtain, configure, and use Eclipse.

Downloading and Installing

Note: The most current release of Eclipse is called Ganymede. Our testing shows that Ganymede is currently incompatible with the Hadoop MapReduce plugin. The most recent version which worked properly with the Hadoop plugin is version 3.3.1, “Europa.” To download Europa, do not visit the main Eclipse website; it can be found in the archive site http://archive.eclipse.org/eclipse/downloads/ as the “Archived Release (3.3.1).”

The Eclipse website has several versions available for download; choose either “Eclipse Classic” or “Eclipse IDE for Java Developers.”

Because it is written in Java, Eclipse is very cross-platform. Eclipse is available for Windows, Linux, and Mac OSX.

Installing Eclipse is very straightforward. Eclipse is packaged as a .zip file. Windows itself can natively unzip the compressed file into a directory. If you encounter errors using the Windows decompression tool (see [1]), try using a third-party unzip utility such as 7-zip or WinRAR.

After you have decompressed Eclipse into a directory, you can run it straight from that directory with no modifications or other “installation” procedure. You may want to move it into C:\Program Files\Eclipse to keep consistent with your other applications, but it can reside in the Desktop or elsewhere as well.

Installing the Hadoop MapReduce Plugin

Hadoop comes with a plugin for Eclipse that makes developing MapReduce programs easier. In the hadoop-0.18.0/contrib/eclipse-plugin directory on this CD, you will find a file named hadoop-0.18.0-eclipse-plugin.jar. Copy this into the plugins/ subdirectory of wherever you unzipped Eclipse.

Making a Copy of Hadoop

While we will be running MapReduce programs on the virtual machine, we will be compiling them on the host machine. The host therefore needs a copy of the Hadoop jars to compile your code against. Copy the /hadoop-0.18.0 directory from the CD into a location on your local drive, and remember where this is. You do not need to configure this copy of Hadoop in any way.

Running Eclipse

Navigate into the Eclipse directory and run eclipse.exe to start the IDE. Eclipse stores all of your source projects and their related settings in a directory called a workspace.

Upon starting Eclipse, it will prompt you for a directory to act as the workspace. Choose a directory name that makes sense to you and click OK.


Figure 3.3: When you first start Eclipse, you must choose a directory to act as your workspace.

Configuring the MapReduce Plugin

In this section, we will walk through the process of configuring Eclipse to switch to the MapReduce perspective and connect to the Hadoop virtual machine.

Step 1: If you have not already done so, start Eclipse and choose a workspace directory. If you are presented with a “welcome” screen, click the button that says “Go to the Workbench.” The Workbench is the main view of Eclipse, where you can write source code, launch programs, and manage your projects.

Step 2: Start the virtual machine. Double-click on the image.vmx file in the virtual machine’s installation directory to launch the virtual machine. It should begin the Linux boot process.

Step 3: Switch to the MapReduce perspective. In the upper-right corner of the workbench, click the “Open Perspective” button, as shown in Figure 3.4:


Figure 3.4: Changing the Perspective

Select “Other,” followed by “Map/Reduce” in the window that opens up. At first, nothing may appear to change. In the menu, choose Window * Show View * Other. Under “MapReduce Tools,” select “Map/Reduce Locations.” This should make a new panel visible at the bottom of the screen, next to Problems and Tasks.

Step 4: Add the Server. In the Map/Reduce Locations panel, click on the elephant logo in the upper-right corner to add a new server to Eclipse.

Figure 3.5: Adding a New Server

You will now be asked to fill in a number of parameters identifying the server. To connect to the VMware image, the values are:

Location name: (Any descriptive name you want; e.g., "VMware server")
Map/Reduce Master Host: (The IP address printed at startup)
Map/Reduce Master Port: 9001
DFS Master Port: 9000
User name: hadoop-user

Next, click on the “Advanced” tab. There are two settings here which must be changed.

Scroll down to hadoop.job.ugi. It contains your current Windows login credentials. Highlight the first comma-separated value in this list (your username) and replace it with hadoop-user.

Next, scroll further down to mapred.system.dir. Erase the current value and set it to /hadoop/mapred/system.

When you are done, click “Finish.” Your server will now appear in the Map/Reduce Locations panel. If you look in the Project Explorer (upper-left corner of Eclipse), you will see that the MapReduce plugin has added the ability to browse HDFS. Click the [+] buttons to expand the directory tree to see any files already there. If you inserted files into HDFS yourself, they will be visible in this tree.

Figure 3.6: Files Visible in the HDFS Viewer

Now that your system is configured, the following sections will introduce you to the basic features and verify that they work correctly.

Interacting With HDFS

The VMware image will expose a single-node HDFS instance for your use in MapReduce applications. If you are logged in to the virtual machine, you can interact with HDFS using the command-line tools described in Module 2. You can also manipulate HDFS through the MapReduce plugin.

Using the Command Line

An interesting MapReduce task will require some external data to process: log files, web crawl results, etc. Before you can begin processing with MapReduce, data must be loaded into its distributed file system. In Module 2, you learned how to copy files from the local file system into HDFS. But this will copy files from the local file system of the VM into HDFS – not from the file system of your host computer.

To load data into HDFS in the virtual machine, you have several options available to you:

  1. scp the files to the virtual machine, and then use the bin/hadoop fs -put ... syntax to copy the files from the VM’s local file system into HDFS,
  2. pipe the data from the local machine into a put command reading from stdin,
  3. or install the Hadoop tools on the host system and configure it to communicate directly with the guest instance

We will review each of these in turn.

To load data into HDFS using the command line within the virtual machine, you can first send the data to the VM’s local disk, then insert it into HDFS. You can send files to the VM using an scp client, such as the pscp component of putty, or WinSCP.

scp will allow you to copy files from one machine to another over the network. The scp command takes two arguments, both of the form [[username@]hostname]:filename. The scp command itself is of the form scp source dest, where source and dest are formatted as described above. By default, it will assume that paths are on the local host, and should be accessed using the current username. You can override the username and hostname to perform remote copies.

So supposing you have a file named foo.txt, and you would like to copy this into the virtual machine which has IP address, you can perform this operation with the command:

  $ scp foo.txt hadoop-user@

If you are using the pscp program, substitute pscp instead of scp above. A copy of the “regular” scp can be run under cygwin by downloading the OpenSSH package. pscp is a utility by the makers of putty and does not require cygwin.

Note that since we did not specify a destination directory, it will go in /home/hadoop-user by default. To change the target directory, specify it after the hostname (e.g., hadoop-user@ You can also omit the destination filename, if you want it to be identical to the source filename. However, if you omit both the target directory and filename, you must not forget the colon (“:”) that follows the target hostname. Otherwise it will make a local copy of the file, with the name An equivalent correct command to copy foo.txt to /home/hadoop-user on the remote machine is:

  $ scp foo.txt hadoop-user@

Windows users may be more inclined to use a GUI tool to perform scp commands. The free WinSCP program provides an FTP-like GUI interface over scp.

After you have copied files into the local disk of the virtual machine, you can log in to the virtual machine as hadoop-user and insert the files into HDFS using the standard Hadoop commands. For example,

hadoop-user@vm-instance:hadoop$ bin/hadoop dfs -put ~/foo.txt \

A second option available to upload individual files to HDFS from the host machine is to echo the file contents into a put command running via ssh. e.g., assuming you have the cat program (which comes with Linux or cygwin) to echo the contents of a file to the terminal output, you can connect its output to the input of a put command running over ssh like so:

you@host-machine$ cat somefile | ssh hadoop-user@vm-ip-addr \
  "hadoop/bin/hadoop fs -put - destinationfile

The - as an argument to the put command instructs the system to use stdin as its input file. This will copy somefile on the host machine to destinationfile in HDFS on the virtual machine.

Finally, if you are running either Linux or cygwin, you can copy the /hadoop-0.18.0 directory on the CD to your local instance. You can then configure hadoop-site.xml to use the virtual machine as the default distributed file system (by setting the fs.default.name parameter). If you then run bin/hadoop fs -put ... commands on this machine (or any other hadoop commands, for that matter), they will interact with HDFS as served by the virtual machine. See the Hadoop getting started for instructions on configuring a Hadoop installation, or Module 7 for a more thorough treatment.

Using the MapReduce Plugin For Eclipse

An easier way to manipulate files in HDFS may be through the Eclipse plugin. In the DFS location viewer, right-click on any folder to see a list of actions available. You can create new subdirectories, upload individual files or whole subdirectories, or download files and directories to the local disk.

If /user/hadoop-user does not exist, create that first. Right-click on the top-level directory and select “Create New Directory”. Type “user” and click OK. You will then need to refresh the current directory view by right-clicking and selecting “Refresh” from the pop-up menu. Repeat this process to create the “hadoop-user” directory under “user.”

Now, prepare some local files to upload. Somewhere on your hard drive, create a directory named “input” and find some text files to copy there. In the DFS explorer, right-click the “hadoop-user” directory and click “Upload Directory to DFS.” Select your new input folder and click OK. Eclipse will copy the files directly into HDFS, bypassing the local drive of the virtual machine. You may have to refresh the directory view to see your changes. You should now have a directory hierarchy containing the /user/hadoop-user/input directory, which has at least one text file in it.

Running a Sample Program

While we have not yet formally introduced the programming style for Hadoop, we can still test whether a MapReduce program will run on our Hadoop virtual machine. This section walks you through the steps required to verify this.

The program that we will run is a word count utility. The program will read the files you uploaded to HDFS in the previous section, and determine how many times each word in the files appears.

If you have not already done so, start the virtual machine and Eclipse, and switch Eclipse to use the MapReduce perspective. Instructions are in the previous section.

Creating the Project

In the menu, click File * New * Project. Select “Map/Reduce Project” from the list and click Next.

You now need to select a project name. Any name will do, e.g., “WordCount”. You will also need to specify the Hadoop Library Installation Path. This is the path where you made a copy of the /hadoop-0.18.0 folder on the CD.  Since we have not yet configured this part of Eclipse, do so now by clicking “Configure Hadoop install directory…” and choosing the path where you copied Hadoop to. There should be a file named hadoop-0.18.0-core.jar in this directory. Creating a MapReduce Project instead of a generic Java project automatically adds the prerequisite jar files to the build path. If you create a regular Java project, you must add the Hadoop jar (and its dependencies) to the build path manually.

When you have completed these steps, click Finish.

Creating the Source Files

Our program needs three classes to run: a Mapper, a Reducer, and a Driver. The Driver tells Hadoop how to run the MapReduce process. The Mapper and Reducer operate on your data.

Right-click on the “src” folder under your project and select New * Other…. In the “Map/Reduce” folder on the resulting window, we can create Mapper, Reducer, and Driver classes based on pre-written stub code. Create classes named WordCountMapper, WordCountReducer, and WordCount that use the Mapper, Reducer, and Driver stubs respectively.

The code for each of these classes is shown here. You can copy this code into your files.


import java.io.IOException;
import java.util.StringTokenizer;

import org.apache.hadoop.io.IntWritable;
import org.apache.hadoop.io.LongWritable;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.io.Writable;
import org.apache.hadoop.io.WritableComparable;
import org.apache.hadoop.mapred.MapReduceBase;
import org.apache.hadoop.mapred.Mapper;
import org.apache.hadoop.mapred.OutputCollector;
import org.apache.hadoop.mapred.Reporter;

public class WordCountMapper extends MapReduceBase
    implements Mapper<LongWritable, Text, Text, IntWritable> {

  private final IntWritable one = new IntWritable(1);
  private Text word = new Text();

  public void map(WritableComparable key, Writable value,
      OutputCollector output, Reporter reporter) throws IOException {

    String line = value.toString();
    StringTokenizer itr = new StringTokenizer(line.toLowerCase());
    while(itr.hasMoreTokens()) {
      output.collect(word, one);


import java.io.IOException;
import java.util.Iterator;

import org.apache.hadoop.io.IntWritable;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.io.WritableComparable;
import org.apache.hadoop.mapred.MapReduceBase;
import org.apache.hadoop.mapred.OutputCollector;
import org.apache.hadoop.mapred.Reducer;
import org.apache.hadoop.mapred.Reporter;

public class WordCountReducer extends MapReduceBase
    implements Reducer<Text, IntWritable, Text, IntWritable> {

  public void reduce(Text key, Iterator values,
      OutputCollector output, Reporter reporter) throws IOException {

    int sum = 0;
    while (values.hasNext()) {
      IntWritable value = (IntWritable) values.next();
      sum += value.get(); // process value

    output.collect(key, new IntWritable(sum));


import org.apache.hadoop.fs.Path;
import org.apache.hadoop.io.IntWritable;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapred.FileInputFormat;
import org.apache.hadoop.mapred.FileOutputFormat;
import org.apache.hadoop.mapred.JobClient;
import org.apache.hadoop.mapred.JobConf;

public class WordCount {

  public static void main(String[] args) {
    JobClient client = new JobClient();
    JobConf conf = new JobConf(WordCount.class);

    // specify output types

    // specify input and output dirs
    FileInputPath.addInputPath(conf, new Path("input"));
    FileOutputPath.addOutputPath(conf, new Path("output"));

    // specify a mapper

    // specify a reducer

    try {
    } catch (Exception e) {

For now, don’t worry about how these functions work; we will introduce how to write MapReduce programs in Module 4. We currently just want to establish that we can run jobs on the virtual machine.

Launching the Job

After the code has been entered, it is time to run it. You have already created a directory named input below /user/hadoop-user in HDFS. This will serve as the input files to this process. In the Project Explorer, right-click on the driver class, WordCount.java. In the pop-up menu, select Run As * Run On Hadoop. A window will appear asking you to select a Hadoop location to run on. Select the VMware server that you configured earlier, and click Finish.

If all goes well, the progress output from Hadoop should appear in the console in Eclipse; it should look something like:

08/06/25 12:14:22 INFO mapred.FileInputFormat: Total input paths to process : 3
08/06/25 12:14:23 INFO mapred.JobClient: Running job: job_200806250515_0002
08/06/25 12:14:24 INFO mapred.JobClient:  map 0% reduce 0%
08/06/25 12:14:31 INFO mapred.JobClient:  map 50% reduce 0%
08/06/25 12:14:33 INFO mapred.JobClient:  map 100% reduce 0%
08/06/25 12:14:42 INFO mapred.JobClient:  map 100% reduce 100%
08/06/25 12:14:43 INFO mapred.JobClient: Job complete: job_200806250515_0002
08/06/25 12:14:43 INFO mapred.JobClient: Counters: 12
08/06/25 12:14:43 INFO mapred.JobClient:   Job Counters
08/06/25 12:14:43 INFO mapred.JobClient:     Launched map tasks=4
08/06/25 12:14:43 INFO mapred.JobClient:     Launched reduce tasks=1
08/06/25 12:14:43 INFO mapred.JobClient:     Data-local map tasks=4
08/06/25 12:14:43 INFO mapred.JobClient:   Map-Reduce Framework
08/06/25 12:14:43 INFO mapred.JobClient:     Map input records=211
08/06/25 12:14:43 INFO mapred.JobClient:     Map output records=1609
08/06/25 12:14:43 INFO mapred.JobClient:     Map input bytes=11627
08/06/25 12:14:43 INFO mapred.JobClient:     Map output bytes=16918
08/06/25 12:14:43 INFO mapred.JobClient:     Combine input records=1609
08/06/25 12:14:43 INFO mapred.JobClient:     Combine output records=682
08/06/25 12:14:43 INFO mapred.JobClient:     Reduce input groups=568
08/06/25 12:14:43 INFO mapred.JobClient:     Reduce input records=682
08/06/25 12:14:43 INFO mapred.JobClient:     Reduce output records=568

In the DFS Explorer, right-click on /user/hadoop-user and select “Refresh.” You should now see an “output” directory containing a file named part-00000. This is the output of the job. Double-clicking this file will allow you to view it in Eclipse; you can see each word and its frequency in the documents. (You may receive a warning that this file is larger than 1 MB, first. Click OK.)

If you want to run the job again, you will need to delete the output directory first. Right-click the output directory in the DFS Explorer and click “Delete.”

Congratulations! You should now have a functioning Hadoop development environment. In the next module, we will learn how to use it to perform powerful programming tasks.

References & Resources

These resources are links to general Hadoop sites. They should be your first stop for troubleshooting or more information.

  • Hadoop site – Central location for downloads, documentation and information
  • Hadoop wiki – User-powered documentation for Hadoop
  • JavaDoc – Current Hadoop API documentation
  • Mailing list info – Hadoop community discussion & advice

Appendix: Complete Tools List

Included in this section is a complete list of programs necessary to run Hadoop, and optional programs which may be helpful in installing or using it. Some of these assume a Windows development environment (though not necessarily a Windows-based cluster).

Necessary for Hadoop:

Useful for this tutorial:

Module 4: MapReduce


MapReduce is a programming model designed for processing large volumes of data in parallel by dividing the work into a set of independent tasks. MapReduce programs are written in a particular style influenced by functional programming constructs, specifically idioms for processing lists of data. This module explains the nature of this programming model and how it can be used to write programs which run in the Hadoop environment.

Goals for this Module:

  • Understand functional programming as it applies to MapReduce
  • Understand the MapReduce program flow
  • Understand how to write programs for Hadoop MapReduce
  • Learn about additional features of Hadoop designed to aid software development.


  1. Introduction
  2. Goals for this Module
  3. Outline
  4. Prerequisites
  5. MapReduce Basics
    1. Functional Programming Concepts
    2. List Processing
    3. Mapping Lists
    4. Reducing Lists
    5. Putting them Together in MapReduce
    6. An Example Application: Word Count
    7. The Driver Method
  6. MapReduce Data Flow
    1. A Closer Look
    2. Additional MapReduce Functionality
    3. Fault Tolerance
  7. Checkpoint
  8. More Tips
    1. Chaining Jobs
    2. Troubleshooting: Debugging MapReduce
    3. Listing and Killing Jobs
  9. Additional Language Support
    1. Pipes
    2. Hadoop Streaming
  10. Conclusions
  11. Solution to Inverted Index Code


This module requires that you have set up a build environment as described in Module 3. If you have not already configured Hadoop and successfully run the example applications, go back and do so now.

MapReduce Basics

Functional Programming Concepts

MapReduce programs are designed to compute large volumes of data in a parallel fashion. This requires dividing the workload across a large number of machines. This model would not scale to large clusters (hundreds or thousands of nodes) if the components were allowed to share data arbitrarily. The communication overhead required to keep the data on the nodes synchronized at all times would prevent the system from performing reliably or efficiently at large scale.

Instead, all data elements in MapReduce are immutable, meaning that they cannot be updated. If in a mapping task you change an input (key, value) pair, it does not get reflected back in the input files; communication occurs only by generating new output (key, value) pairs which are then forwarded by the Hadoop system into the next phase of execution.

List Processing

Conceptually, MapReduce programs transform lists of input data elements into lists of output data elements. A MapReduce program will do this twice, using two different list processing idioms: map, and reduce. These terms are taken from several list processing languages such as LISP, Scheme, or ML.

Mapping Lists

The first phase of a MapReduce program is called mapping. A list of data elements are provided, one at a time, to a function called the Mapper, which transforms each element individually to an output data element.


Figure 4.1: Mapping creates a new output list by applying a function to individual elements of an input list.

As an example of the utility of map: Suppose you had a function toUpper(str) which returns an uppercase version of its input string. You could use this function with map to turn a list of strings into a list of uppercase strings. Note that we are not modifying the input string here: we are returning a new string that will form part of a new output list.

Reducing Lists

Reducing lets you aggregate values together. A reducer function receives an iterator of input values from an input list. It then combines these values together, returning a single output value.


Figure 4.2: Reducing a list iterates over the input values to produce an aggregate value as output.

Reducing is often used to produce “summary” data, turning a large volume of data into a smaller summary of itself. For example, “+” can be used as a reducing function, to return the sum of a list of input values.

Putting Them Together in MapReduce:

The Hadoop MapReduce framework takes these concepts and uses them to process large volumes of information. A MapReduce program has two components: one that implements the mapper, and another that implements the reducer. The Mapper and Reducer idioms described above are extended slightly to work in this environment, but the basic principles are the same.

Keys and values: In MapReduce, no value stands on its own. Every value has a key associated with it. Keys identify related values. For example, a log of time-coded speedometer readings from multiple cars could be keyed by license-plate number; it would look like:

AAA-123   65mph, 12:00pm
ZZZ-789   50mph, 12:02pm
AAA-123   40mph, 12:05pm
CCC-456   25mph, 12:15pm

The mapping and reducing functions receive not just values, but (key, value) pairs. The output of each of these functions is the same: both a key and a value must be emitted to the next list in the data flow.

MapReduce is also less strict than other languages about how the Mapper and Reducer work. In more formal functional mapping and reducing settings, a mapper must produce exactly one output element for each input element, and a reducer must produce exactly one output element for each input list. In MapReduce, an arbitrary number of values can be output from each phase; a mapper may map one input into zero, one, or one hundred outputs. A reducer may compute over an input list and emit one or a dozen different outputs.

Keys divide the reduce space: A reducing function turns a large list of values into one (or a few) output values. In MapReduce, all of the output values are not usually reduced together. All of the values with the same key are presented to a single reducer together. This is performed independently of any reduce operations occurring on other lists of values, with different keys attached.


Figure 4.3: Different colors represent different keys. All values with the same key are presented to a single reduce task.

An Example Application: Word Count

A simple MapReduce program can be written to determine how many times different words appear in a set of files. For example, if we had the files:

foo.txt: Sweet, this is the foo file

bar.txt: This is the bar file

We would expect the output to be:


sweet 1
this  2
is    2
the   2
foo   1
bar   1
file  2

Naturally, we can write a program in MapReduce to compute this output. The high-level structure would look like this:

mapper (filename, file-contents):
  for each word in file-contents:
    emit (word, 1)

reducer (word, values):
  sum = 0
  for each value in values:
    sum = sum + value
  emit (word, sum)

Listing 4.1: High-Level MapReduce Word Count

Several instances of the mapper function are created on the different machines in our cluster. Each instance receives a different input file (it is assumed that we have many such files). The mappers output (word, 1) pairs which are then forwarded to the reducers. Several instances of the reducer method are also instantiated on the different machines. Each reducer is responsible for processing the list of values associated with a different word. The list of values will be a list of 1’s; the reducer sums up those ones into a final count associated with a single word. The reducer then emits the final (word, count) output which is written to an output file.

We can write a very similar program to this in Hadoop MapReduce; it is included in the Hadoop distribution in src/examples/org/apache/hadoop/examples/WordCount.java. It is partially reproduced below:

  public static class MapClass extends MapReduceBase
    implements Mapper<LongWritable, Text, Text, IntWritable> {

    private final static IntWritable one = new IntWritable(1);
    private Text word = new Text();

    public void map(LongWritable key, Text value,
                    OutputCollector<Text, IntWritable> output,
                    Reporter reporter) throws IOException {
      String line = value.toString();
      StringTokenizer itr = new StringTokenizer(line);
      while (itr.hasMoreTokens()) {
        output.collect(word, one);

   * A reducer class that just emits the sum of the input values.
  public static class Reduce extends MapReduceBase
    implements Reducer<Text, IntWritable, Text, IntWritable> {

    public void reduce(Text key, Iterator<IntWritable> values,
                       OutputCollector<Text, IntWritable> output,
                       Reporter reporter) throws IOException {
      int sum = 0;
      while (values.hasNext()) {
        sum += values.next().get();
      output.collect(key, new IntWritable(sum));

Listing 4.2: Hadoop MapReduce Word Count Source

There are some minor differences between this actual Java implementation and the pseudo-code shown above. First, Java has no native emit keyword; the OutputCollector object you are given as an input will receive values to emit to the next stage of execution. And second, the default input format used by Hadoop presents each line of an input file as a separate input to the mapper function, not the entire file at a time. It also uses a StringTokenizer object to break up the line into words. This does not perform any normalization of the input, so “cat”, “Cat” and “cat,” are all regarded as different strings. Note that the class-variable word is reused each time the mapper outputs another (word, 1) pairing; this saves time by not allocating a new variable for each output. The output.collect() method will copy the values it receives as input, so you are free to overwrite the variables you use.

The Driver Method

There is one final component of a Hadoop MapReduce program, called the Driver. The driver initializes the job and instructs the Hadoop platform to execute your code on a set of input files, and controls where the output files are placed. A cleaned-up version of the driver from the example Java implementation that comes with Hadoop is presented below:

  public void run(String inputPath, String outputPath) throws Exception {
    JobConf conf = new JobConf(WordCount.class);

    // the keys are words (strings)
    // the values are counts (ints)


    FileInputFormat.addInputPath(conf, new Path(inputPath));
    FileOutputFormat.setOutputPath(conf, new Path(outputPath));


Listing 4.3: Hadoop MapReduce Word Count Driver

This method sets up a job to execute the word count program across all the files in a given input directory (the inputPath argument). The output from the reducers are written into files in the directory identified by outputPath. The configuration information to run the job is captured in the JobConf object. The mapping and reducing functions are identified by the setMapperClass() and setReducerClass() methods. The data types emitted by the reducer are identified by setOutputKeyClass() and setOutputValueClass(). By default, it is assumed that these are the output types of the mapper as well. If this is not the case, the methods setMapOutputKeyClass() and setMapOutputValueClass() methods of the JobConf class will override these. The input types fed to the mapper are controlled by the InputFormat used. Input formats are discussed in more detail below. The default input format, “TextInputFormat,” will load data in as (LongWritable, Text) pairs. The long value is the byte offset of the line in the file. The Text object holds the string contents of the line of the file.

The call to JobClient.runJob(conf) will submit the job to MapReduce. This call will block until the job completes. If the job fails, it will throw an IOException. JobClient also provides a non-blocking version called submitJob().

MapReduce Data Flow

Now that we have seen the components that make up a basic MapReduce job, we can see how everything works together at a higher level:


Figure 4.4: High-level MapReduce pipeline

MapReduce inputs typically come from input files loaded onto our processing cluster in HDFS. These files are evenly distributed across all our nodes. Running a MapReduce program involves running mapping tasks on many or all of the nodes in our cluster. Each of these mapping tasks is equivalent: no mappers have particular “identities” associated with them. Therefore, any mapper can process any input file. Each mapper loads the set of files local to that machine and processes them.

When the mapping phase has completed, the intermediate (key, value) pairs must be exchanged between machines to send all values with the same key to a single reducer. The reduce tasks are spread across the same nodes in the cluster as the mappers. This is the only communication step in MapReduce. Individual map tasks do not exchange information with one another, nor are they aware of one another’s existence. Similarly, different reduce tasks do not communicate with one another. The user never explicitly marshals information from one machine to another; all data transfer is handled by the Hadoop MapReduce platform itself, guided implicitly by the different keys associated with values. This is a fundamental element of Hadoop MapReduce’s reliability. If nodes in the cluster fail, tasks must be able to be restarted. If they have been performing side-effects, e.g., communicating with the outside world, then the shared state must be restored in a restarted task. By eliminating communication and side-effects, restarts can be handled more gracefully.

A Closer Look

The previous figure described the high-level view of Hadoop MapReduce. From this diagram, you can see where the mapper and reducer components of the Word Count application fit in, and how it achieves its objective. We will now examine this system in a bit closer detail.


Figure 4.5: Detailed Hadoop MapReduce data flow

Figure 4.5 shows the pipeline with more of its mechanics exposed. While only two nodes are depicted, the same pipeline can be replicated across a very large number of nodes. The next several paragraphs describe each of the stages of a MapReduce program more precisely.

Input files: This is where the data for a MapReduce task is initially stored. While this does not need to be the case, the input files typically reside in HDFS. The format of these files is arbitrary; while line-based log files can be used, we could also use a binary format, multi-line input records, or something else entirely. It is typical for these input files to be very large — tens of gigabytes or more.

InputFormat: How these input files are split up and read is defined by the InputFormat. An InputFormat is a class that provides the following functionality:

  • Selects the files or other objects that should be used for input
  • Defines the InputSplits that break a file into tasks
  • Provides a factory for RecordReader objects that read the file

Several InputFormats are provided with Hadoop. An abstract type is called FileInputFormat; all InputFormats that operate on files inherit functionality and properties from this class. When starting a Hadoop job, FileInputFormat is provided with a path containing files to read. The FileInputFormat will read all files in this directory. It then divides these files into one or more InputSplits each. You can choose which InputFormat to apply to your input files for a job by calling the setInputFormat() method of the JobConf object that defines the job. A table of standard InputFormats is given below.

InputFormat: Description: Key: Value:
TextInputFormat Default format; reads lines of text files The byte offset of the line The line contents
KeyValueInputFormat Parses lines into key, val pairs Everything up to the first tab character The remainder of the line
SequenceFileInputFormat A Hadoop-specific high-performance binary format user-defined user-defined

Table 4.1: InputFormats provided by MapReduce

The default InputFormat is the TextInputFormat. This treats each line of each input file as a separate record, and performs no parsing. This is useful for unformatted data or line-based records like log files. A more interesting input format is the KeyValueInputFormat. This format also treats each line of input as a separate record. While the TextInputFormat treats the entire line as the value, the KeyValueInputFormat breaks the line itself into the key and value by searching for a tab character. This is particularly useful for reading the output of one MapReduce job as the input to another, as the default OutputFormat (described in more detail below) formats its results in this manner. Finally, the SequenceFileInputFormat reads special binary files that are specific to Hadoop. These files include many features designed to allow data to be rapidly read into Hadoop mappers. Sequence files are block-compressed and provide direct serialization and deserialization of several arbitrary data types (not just text). Sequence files can be generated as the output of other MapReduce tasks and are an efficient intermediate representation for data that is passing from one MapReduce job to anther.

InputSplits: An InputSplit describes a unit of work that comprises a single map task in a MapReduce program. A MapReduce program applied to a data set, collectively referred to as a Job, is made up of several (possibly several hundred) tasks. Map tasks may involve reading a whole file; they often involve reading only part of a file. By default, the FileInputFormat and its descendants break a file up into 64 MB chunks (the same size as blocks in HDFS). You can control this value by setting the mapred.min.split.size parameter in hadoop-site.xml, or by overriding the parameter in the JobConf object used to submit a particular MapReduce job. By processing a file in chunks, we allow several map tasks to operate on a single file in parallel. If the file is very large, this can improve performance significantly through parallelism. Even more importantly, since the various blocks that make up the file may be spread across several different nodes in the cluster, it allows tasks to be scheduled on each of these different nodes; the individual blocks are thus all processed locally, instead of needing to be transferred from one node to another. Of course, while log files can be processed in this piece-wise fashion, some file formats are not amenable to chunked processing. By writing a custom InputFormat, you can control how the file is broken up (or is not broken up) into splits. Custom input formats are described in Module 5.

The InputFormat defines the list of tasks that make up the mapping phase; each task corresponds to a single input split. The tasks are then assigned to the nodes in the system based on where the input file chunks are physically resident. An individual node may have several dozen tasks assigned to it. The node will begin working on the tasks, attempting to perform as many in parallel as it can. The on-node parallelism is controlled by the mapred.tasktracker.map.tasks.maximum parameter.

RecordReader: The InputSplit has defined a slice of work, but does not describe how to access it. The RecordReader class actually loads the data from its source and converts it into (key, value) pairs suitable for reading by the Mapper. The RecordReader instance is defined by the InputFormat. The default InputFormat, TextInputFormat, provides a LineRecordReader, which treats each line of the input file as a new value. The key associated with each line is its byte offset in the file. The RecordReader is invoke repeatedly on the input until the entire InputSplit has been consumed. Each invocation of the RecordReader leads to another call to the map() method of the Mapper.

Mapper: The Mapper performs the interesting user-defined work of the first phase of the MapReduce program. Given a key and a value, the map() method emits (key, value) pair(s) which are forwarded to the Reducers. A new instance of Mapper is instantiated in a separate Java process for each map task (InputSplit) that makes up part of the total job input. The individual mappers are intentionally not provided with a mechanism to communicate with one another in any way. This allows the reliability of each map task to be governed solely by the reliability of the local machine. The map() method receives two parameters in addition to the key and the value:

  • The OutputCollector object has a method named collect() which will forward a (key, value) pair to the reduce phase of the job.
  • The Reporter object provides information about the current task; its getInputSplit() method will return an object describing the current InputSplit. It also allows the map task to provide additional information about its progress to the rest of the system. The setStatus() method allows you to emit a status message back to the user. The incrCounter() method allows you to increment shared performance counters. You may define as many arbitrary counters as you wish. Each mapper can increment the counters, and the JobTracker will collect the increments made by the different processes and aggregate them for later retrieval when the job ends.

Partition & Shuffle: After the first map tasks have completed, the nodes may still be performing several more map tasks each. But they also begin exchanging the intermediate outputs from the map tasks to where they are required by the reducers. This process of moving map outputs to the reducers is known as shuffling. A different subset of the intermediate key space is assigned to each reduce node; these subsets (known as “partitions”) are the inputs to the reduce tasks. Each map task may emit (key, value) pairs to any partition; all values for the same key are always reduced together regardless of which mapper is its origin. Therefore, the map nodes must all agree on where to send the different pieces of the intermediate data. The Partitioner class determines which partition a given (key, value) pair will go to. The default partitioner computes a hash value for the key and assigns the partition based on this result. Custom partitioners are described in more detail in Module 5.

Sort: Each reduce task is responsible for reducing the values associated with several intermediate keys. The set of intermediate keys on a single node is automatically sorted by Hadoop before they are presented to the Reducer.

Reduce: A Reducer instance is created for each reduce task. This is an instance of user-provided code that performs the second important phase of job-specific work. For each key in the partition assigned to a Reducer, the Reducer’s reduce() method is called once. This receives a key as well as an iterator over all the values associated with the key. The values associated with a key are returned by the iterator in an undefined order. The Reducer also receives as parameters OutputCollector and Reporter objects; they are used in the same manner as in the map() method.

OutputFormat: The (key, value) pairs provided to this OutputCollector are then written to output files. The way they are written is governed by the OutputFormat. The OutputFormat functions much like the InputFormat class described earlier. The instances of OutputFormat provided by Hadoop write to files on the local disk or in HDFS; they all inherit from a common FileOutputFormat. Each Reducer writes a separate file in a common output directory. These files will typically be named part-nnnnn, where nnnnn is the partition id associated with the reduce task. The output directory is set by the FileOutputFormat.setOutputPath() method. You can control which particular OutputFormat is used by calling the setOutputFormat() method of the JobConf object that defines your MapReduce job. A table of provided OutputFormats is given below.

OutputFormat: Description
TextOutputFormat Default; writes lines in “key \t value” form
SequenceFileOutputFormat Writes binary files suitable for reading into subsequent MapReduce jobs
NullOutputFormat Disregards its inputs

Table 4.2: OutputFormats provided by Hadoop

Hadoop provides some OutputFormat instances to write to files. The basic (default) instance is TextOutputFormat, which writes (key, value) pairs on individual lines of a text file. This can be easily re-read by a later MapReduce task using the KeyValueInputFormat class, and is also human-readable. A better intermediate format for use between MapReduce jobs is the SequenceFileOutputFormat which rapidly serializes arbitrary data types to the file; the corresponding SequenceFileInputFormat will deserialize the file into the same types and presents the data to the next Mapper in the same manner as it was emitted by the previous Reducer. The NullOutputFormat generates no output files and disregards any (key, value) pairs passed to it by the OutputCollector. This is useful if you are explicitly writing your own output files in the reduce() method, and do not want additional empty output files generated by the Hadoop framework.

RecordWriter: Much like how the InputFormat actually reads individual records through the RecordReader implementation, the OutputFormat class is a factory for RecordWriter objects; these are used to write the individual records to the files as directed by the OutputFormat.

The output files written by the Reducers are then left in HDFS for your use, either by another MapReduce job, a separate program, for for human inspection.

Additional MapReduce Functionality


Figure 4.6: Combiner step inserted into the MapReduce data flow

Combiner: The pipeline showed earlier omits a processing step which can be used for optimizing bandwidth usage by your MapReduce job. Called the Combiner, this pass runs after the Mapper and before the Reducer. Usage of the Combiner is optional. If this pass is suitable for your job, instances of the Combiner class are run on every node that has run map tasks. The Combiner will receive as input all data emitted by the Mapper instances on a given node. The output from the Combiner is then sent to the Reducers, instead of the output from the Mappers. The Combiner is a “mini-reduce” process which operates only on data generated by one machine.

Word count is a prime example for where a Combiner is useful. The Word Count program in listings 1–3 emits a (word, 1) pair for every instance of every word it sees. So if the same document contains the word “cat” 3 times, the pair ("cat", 1) is emitted three times; all of these are then sent to the Reducer. By using a Combiner, these can be condensed into a single ("cat", 3) pair to be sent to the Reducer. Now each node only sends a single value to the reducer for each word — drastically reducing the total bandwidth required for the shuffle process, and speeding up the job. The best part of all is that we do not need to write any additional code to take advantage of this! If a reduce function is both commutative and associative, then it can be used as a Combiner as well. You can enable combining in the word count program by adding the following line to the driver:


The Combiner should be an instance of the Reducer interface. If your Reducer itself cannot be used directly as a Combiner because of commutativity or associativity, you might still be able to write a third class to use as a Combiner for your job.

Fault Tolerance

One of the primary reasons to use Hadoop to run your jobs is due to its high degree of fault tolerance. Even when running jobs on a large cluster where individual nodes or network components may experience high rates of failure, Hadoop can guide jobs toward a successful completion.

The primary way that Hadoop achieves fault tolerance is through restarting tasks. Individual task nodes (TaskTrackers) are in constant communication with the head node of the system, called the JobTracker. If a TaskTracker fails to communicate with the JobTracker for a period of time (by default, 1 minute), the JobTracker will assume that the TaskTracker in question has crashed. The JobTracker knows which map and reduce tasks were assigned to each TaskTracker.

If the job is still in the mapping phase, then other TaskTrackers will be asked to re-execute all map tasks previously run by the failed TaskTracker. If the job is in the reducing phase, then other TaskTrackers will re-execute all reduce tasks that were in progress on the failed TaskTracker.

Reduce tasks, once completed, have been written back to HDFS. Thus, if a TaskTracker has already completed two out of three reduce tasks assigned to it, only the third task must be executed elsewhere. Map tasks are slightly more complicated: even if a node has completed ten map tasks, the reducers may not have all copied their inputs from the output of those map tasks. If a node has crashed, then its mapper outputs are inaccessible. So any already-completed map tasks must be re-executed to make their results available to the rest of the reducing machines. All of this is handled automatically by the Hadoop platform.

This fault tolerance underscores the need for program execution to be side-effect free. If Mappers and Reducers had individual identities and communicated with one another or the outside world, then restarting a task would require the other nodes to communicate with the new instances of the map and reduce tasks, and the re-executed tasks would need to reestablish their intermediate state. This process is notoriously complicated and error-prone in the general case. MapReduce simplifies this problem drastically by eliminating task identities or the ability for task partitions to communicate with one another. An individual task sees only its own direct inputs and knows only its own outputs, to make this failure and restart process clean and dependable.

Speculative execution: One problem with the Hadoop system is that by dividing the tasks across many nodes, it is possible for a few slow nodes to rate-limit the rest of the program. For example if one node has a slow disk controller, then it may be reading its input at only 10% the speed of all the other nodes. So when 99 map tasks are already complete, the system is still waiting for the final map task to check in, which takes much longer than all the other nodes.

By forcing tasks to run in isolation from one another, individual tasks do not know where their inputs come from. Tasks trust the Hadoop platform to just deliver the appropriate input. Therefore, the same input can be processed multiple times in parallel, to exploit differences in machine capabilities. As most of the tasks in a job are coming to a close, the Hadoop platform will schedule redundant copies of the remaining tasks across several nodes which do not have other work to perform. This process is known as speculative execution. When tasks complete, they announce this fact to the JobTracker. Whichever copy of a task finishes first becomes the definitive copy. If other copies were executing speculatively, Hadoop tells the TaskTrackers to abandon the tasks and discard their outputs. The Reducers then receive their inputs from whichever Mapper completed successfully, first.

Speculative execution is enabled by default. You can disable speculative execution for the mappers and reducers by setting the mapred.map.tasks.speculative.execution and mapred.reduce.tasks.speculative.execution JobConf options to false, respectively.


You now know about all of the basic operations of the Hadoop MapReduce platform. Try the following exercise, to see if you understand the MapReduce programming concepts.

Exercise: Given the code for WordCount in listings 2 and 3, modify this code to produce an inverted index of its inputs. An inverted index returns a list of documents that contain each word in those documents. Thus, if the word “cat” appears in documents A and B, but not C, then the line:

cat    A, B

should appear in the output. If the word “baseball” appears in documents B and C, then the line:

baseball    B, C

should appear in the output as well.

If you get stuck, read the section on troubleshooting below. The working solution is provided at the end of this module.

Hint: The default InputFormat will provide the Mapper with (key, value) pairs where the key is the byte offset into the file, and the value is a line of text. To get the filename of the current input, use the following code:

FileSplit fileSplit = (FileSplit)reporter.getInputSplit();
String fileName = fileSplit.getPath().getName();

More Tips

Chaining Jobs

Not every problem can be solved with a MapReduce program, but fewer still are those which can be solved with a single MapReduce job. Many problems can be solved with MapReduce, by writing several MapReduce steps which run in series to accomplish a goal:

Map1 -> Reduce1 -> Map2 -> Reduce2 -> Map3…

You can easily chain jobs together in this fashion by writing multiple driver methods, one for each job. Call the first driver method, which uses JobClient.runJob() to run the job and wait for it to complete. When that job has completed, then call the next driver method, which creates a new JobConf object referring to different instances of Mapper and Reducer, etc. The first job in the chain should write its output to a path which is then used as the input path for the second job. This process can be repeated for as many jobs are necessary to arrive at a complete solution to the problem.

Many problems which at first seem impossible in MapReduce can be accomplished by dividing one job into two or more.

Hadoop provides another mechanism for managing batches of jobs with dependencies between jobs. Rather than submit a JobConf to the JobClient‘s runJob() or submitJob() methods, org.apache.hadoop.mapred.jobcontrol.Job objects can be created to represent each job; A Job takes a JobConf object as its constructor argument. Jobs can depend on one another through the use of the addDependingJob() method. The code:


says that Job x cannot start until y has successfully completed. Dependency information cannot be added to a job after it has already been started. Given a set of jobs, these can be passed to an instance of the JobControl class. JobControl can receive individual jobs via the addJob() method, or a collection of jobs via addJobs(). The JobControl object will spawn a thread in the client to launch the jobs. Individual jobs will be launched when their dependencies have all successfully completed and when the MapReduce system as a whole has resources to execute the jobs. The JobControl interface allows you to query it to retrieve the state of individual jobs, as well as the list of jobs waiting, ready, running, and finished. The job submission process does not begin until the run() method of the JobControl object is called.

Troubleshooting: Debugging MapReduce

When writing MapReduce programs, you will occasionally encounter bugs in your programs, infinite loops, etc. This section describes the features of MapReduce that will help you diagnose and solve these conditions.

Log Files: Hadoop keeps logs of important events during program execution. By default, these are stored in the logs/ subdirectory of the hadoop-version/ directory where you run Hadoop from. Log files are named hadoop-username-service-hostname.log. The most recent data is in the .log file; older logs have their date appended to them. The username in the log filename refers to the username under which Hadoop was started — this is not necessarily the same username you are using to run programs. The service name refers to which of the several Hadoop programs are writing the log; these can be jobtracker, namenode, datanode, secondarynamenode, or tasktracker. All of these are important for debugging a whole Hadoop installation. But for individual programs, the tasktracker logs will be the most relevant. Any exceptions thrown by your program will be recorded in the tasktracker logs.

The log directory will also have a subdirectory called userlogs. Here there is another subdirectory for every task run. Each task records its stdout and stderr to two files in this directory. Note that on a multi-node Hadoop cluster, these logs are not centrally aggregated — you should check each TaskNode’s logs/userlogs/ directory for their output.

Debugging in the distributed setting is complicated and requires logging into several machines to access log data. If possible, programs should be unit tested by running Hadoop locally. The default configuration deployed by Hadoop runs in “single instance” mode, where the entire MapReduce program is run in the same instance of Java as called JobClient.runJob(). Using a debugger like Eclipse, you can then set breakpoints inside the map() or reduce() methods to discover your bugs.

In Module 5, you will learn how to use additional features of MapReduce to distribute auxiliary code to nodes in the system. This can be used to enable debug scripts which run on machines when tasks fail.

Listing and Killing Jobs:

It is possible to submit jobs to a Hadoop cluster which malfunction and send themselves into infinite loops or other problematic states. In this case, you will want to manually kill the job you have started.

The following command, run in the Hadoop installation directory on a Hadoop cluster, will list all the current jobs:

$ bin/hadoop job -list

This will produce output that looks something like:

1 jobs currently running
JobId   State   StartTime       UserName
job_200808111901_0001   1       1218506470390   aaron

You can use this job id to kill the job; the command is:

$ bin/hadoop job -kill jobid

Substitute the “job_2008...” from the -list command for jobid.

Additional Language Support

Hadoop itself is written in Java; it thus accepts Java code natively for Mappers and Reducers. Hadoop also comes with two adapter layers which allow code written in other languages to be used in MapReduce programs.


Pipes is a library which allows C++ source code to be used for Mapper and Reducer code. Applications which require high numerical performance may see better throughput if written in C++ and used through Pipes. This library is supported on 32-bit Linux installations.

The include files and static libraries are present in the c++/Linux-i386-32/ directory under your Hadoop installation. Your application should include include/hadoop/Pipes.hh and TemplateFactory.hh and link against lib/libhadooppies.a; with gcc, include the arguments -L${HADOOP_HOME}/c++/Linux-i386-32/lib -lhadooppipes to do the latter.

Both key and value inputs to pipes programs are provided as STL strings (std::string). A program must still define an instance of Mapper and Reducer; these names have not changed. (They, like all other classes defined in Pipes, are in the HadoopPipes namespace.) Unlike the classes of the same names in Hadoop itself, the map() and reduce() functions take in a single argument which is a reference to an object of type MapContext and ReduceContext respectively. The most important methods contained in each of these context objects are:

const std::string& getInputKey();
const std::string& getInputValue();
void emit(const std::string& key, const std::string& value);

The ReduceContext class also contains an additional method to advance the value iterator:

bool nextValue();

Defining a Pipes Program: A program to use with Pipes is defined by writing classes extending Mapper and Reducer. (And optionally, Partitioner; see Module 5.) Hadoop must then be informed which classes to use to run the job.

An instance of your C++ program will be started by the Pipes framework in main() on each machine. This should do any (hopefully brief) configuration required for your task. It should then define a Factory to create Mapper and Reducer instances as necessary, and then run the job by calling the runTask() method. The simplest way to define a factory is with the following code:

using namespace HadoopPipes;

void main() {
  // classes are indicated to the factory via templates
  // TODO: Substitute your own class names in below.
  TemplateFactory2<MyMapperClass, MyReducerClass> factory();

  // do any configuration you need to do here

  // start the task
  bool result = runTask(factory);

Running a Pipes Program: After a Pipes program has been written and compiled, it can be launched as a job with the following command: (Do this in your Hadoop home directory)

$ bin/hadoop pipes -input inputPath -output outputPath -program path/to/pipes/program/executable

This will deploy your Pipes program on all nodes and run the MapReduce job through it. By running bin/hadoop pipes with no options, you can see additional usage information which describes how to set additional configuration values as necessary.

The Pipes API contains additional functionality to allow you to read settings from the JobConf, override the Partitioner class, and use RecordReaders in a more direct fashion for higher performance. See the header files in c++/Linux-i386-32/include/hadoop for more information.

Hadoop Streaming

Whereas Pipes is an API that provides close coupling between C++ application code and Hadoop, Streaming is a generic API that allows programs written in virtually any language to be used as Hadoop Mapper and Reducer implementations.

The official Hadoop documentation contains a thorough introduction to Streaming, and briefer notes on the wiki. A brief overview is presented here.

Hadoop Streaming allows you to use arbitrary programs for the Mapper and Reducer phases of a MapReduce job. Both Mappers and Reducers receive their input on stdin and emit output (key, value) pairs on stdout.

Input and output are always represented textually in Streaming. The input (key, value) pairs are written to stdin for a Mapper or Reducer, with a ‘tab’ character separating the key from the value. The Streaming programs should split the input on the first tab character on the line to recover the key and the value. Streaming programs write their output to stdout in the same format: key \t value \n.

The inputs to the reducer are sorted so that while each line contains only a single (key, value) pair, all the values for the same key are adjacent to one another.

Provided it can handle its input in the text format described above, any Linux program or tool can be used as the mapper or reducer in Streaming. You can also write your own scripts in bash, python, perl, or another language of your choice, provided that the necessary interpreter is present on all nodes in your cluster.

Running a Streaming Job: To run a job with Hadoop Streaming, use the following command:

$ bin/hadoop jar contrib/streaming/hadoop-version-streaming.jar

The command as shown, with no arguments, will print some usage information. An example of how to run real commands is given below:

$ bin/hadoop jar contrib/streaming-hadoop-0.18.0-streaming.jar -mapper \
    myMapProgram -reducer myReduceProgram -input /some/dfs/path \
    -output /some/other/dfs/path

This assumes that myMapProgram and myReduceProgram are present on all nodes in the system ahead of time. If this is not the case, but they are present on the node launching the job, then they can be “shipped” to the other nodes with the -file option:

$ bin/hadoop jar contrib/streaming-hadoop-0.18.0-streaming.jar -mapper \
    myMapProgram -reducer myReduceProgram -file \
    myMapProgram -file myReduceProgram -input some/dfs/path \
    -output some/other/dfs/path

Any other support files necessary to run your program can be shipped in this manner as well.


This module described the MapReduce execution platform at the heart of the Hadoop system. By using MapReduce, a high degree of parallelism can be achieved by applications. The MapReduce framework provides a high degree of fault tolerance for applications running on it by limiting the communication which can occur between nodes, and requiring applications to be written in a “dataflow-centric” manner.

Solution to Inverted Index Code

The following source code implements a solution to the inverted indexer problem posed at the checkpoint. The source code is structurally very similar to the source for Word Count; only a few lines really need to be modified.

import java.io.IOException;
import java.util.Iterator;
import java.util.StringTokenizer;

import org.apache.hadoop.fs.Path;
import org.apache.hadoop.io.LongWritable;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapred.FileInputFormat;
import org.apache.hadoop.mapred.FileOutputFormat;
import org.apache.hadoop.mapred.FileSplit;
import org.apache.hadoop.mapred.JobClient;
import org.apache.hadoop.mapred.JobConf;
import org.apache.hadoop.mapred.MapReduceBase;
import org.apache.hadoop.mapred.Mapper;
import org.apache.hadoop.mapred.OutputCollector;
import org.apache.hadoop.mapred.Reducer;
import org.apache.hadoop.mapred.Reporter;

public class LineIndexer {

  public static class LineIndexMapper extends MapReduceBase
      implements Mapper<LongWritable, Text, Text, Text> {

    private final static Text word = new Text();
    private final static Text location = new Text();

    public void map(LongWritable key, Text val,
        OutputCollector<Text, Text> output, Reporter reporter)
        throws IOException {

      FileSplit fileSplit = (FileSplit)reporter.getInputSplit();
      String fileName = fileSplit.getPath().getName();

      String line = val.toString();
      StringTokenizer itr = new StringTokenizer(line.toLowerCase());
      while (itr.hasMoreTokens()) {
        output.collect(word, location);

  public static class LineIndexReducer extends MapReduceBase
      implements Reducer<Text, Text, Text, Text> {

    public void reduce(Text key, Iterator<Text> values,
        OutputCollector<Text, Text> output, Reporter reporter)
        throws IOException {

      boolean first = true;
      StringBuilder toReturn = new StringBuilder();
      while (values.hasNext()){
        if (!first)
          toReturn.append(", ");

      output.collect(key, new Text(toReturn.toString()));

   * The actual main() method for our program; this is the
   * "driver" for the MapReduce job.
  public static void main(String[] args) {
    JobClient client = new JobClient();
    JobConf conf = new JobConf(LineIndexer.class);



    FileInputFormat.addInputPath(conf, new Path("input"));
    FileOutputFormat.setOutputPath(conf, new Path("output"));



    try {
    } catch (Exception e) {

Module 5: Advanced MapReduce Features


In Module 4 you learned the basics of programming with Hadoop MapReduce. That module explains how data moves through a general MapReduce architecture, and what particular methods and classes facilitate the use of the Hadoop for processing. In this module we will look more closely at how to override Hadoop’s functionality in various ways. These techniques allow you to customize Hadoop for application-specific purposes.

Goals for this Module:

  • Understand advanced Hadoop features
  • Be able to use Hadoop on Amazon EC2 and S3


  1. Introduction
  2. Goals for this Module
  3. Outline
  4. Custom Data Types
    1. Writable Types
    2. Custom Key Types
    3. Using Custom Types
    4. Faster Comparison Operations
    5. Final Writable Notes
  5. Input Formats
    1. Custom File Formats
    2. Alternate Data Sources
  6. Output Formats
  7. Partitioning Data
  8. Reporting Custom Metrics
  9. Distributing Auxiliary Job Data
  10. Distributing Debug Scripts
  11. Using Amazon Web Services
  12. References

Custom Data Types

Hadoop MapReduce uses typed data at all times when it interacts with user-provided Mappers and Reducers: data read from files into Mappers, emitted by mappers to reducers, and emitted by reducers into output files is all stored in Java objects.

Writable Types

Objects which can be marshaled to or from files and across the network must obey a particular interface, called Writable, which allows Hadoop to read and write the data in a serialized form for transmission. Hadoop provides several stock classes which implement Writable: Text (which stores String data), IntWritable, LongWritable, FloatWritable, BooleanWritable, and several others. The entire list is in the org.apache.hadoop.io package of the Hadoop source (see the API reference).

In addition to these types, you are free to define your own classes which implement Writable. You can organize a structure of virtually any layout to fit your data and be transmitted by Hadoop. As a motivating example, consider a mapper which emits key-value pairs where the key is the name of an object, and the value is its coordinates in some 3-dimensional space. The key is some string-based data, and the value is a structure of the form:

struct point3d {
  float x;
  float y;
  float z;

The key can be represented as a Text object, but what about the value? How do we build a Point3D class which Hadoop can transmit? The answer is to implement the Writable interface, which requires two methods:

public interface Writable {
  void readFields(DataInput in);
  void write(DataOutput out);

The first of these methods initializes all of the fields of the object based on data contained in the binary stream in. The latter writes all the information needed to reconstruct the object to the binary stream out. The DataInput and DataOutput classes (part of java.io) contain methods to serialize most basic types of data; the important contract between your readFields() and write() methods is that they read and write the data from and to the binary stream in the same order. The following code implements a Point3D class usable by Hadoop:

public class Point3D implements Writable {
  public float x;
  public float y;
  public float z;

  public Point3D(float x, float y, float z) {
    this.x = x;
    this.y = y;
    this.z = z;

  public Point3D() {
    this(0.0f, 0.0f, 0.0f);

  public void write(DataOutput out) throws IOException {

  public void readFields(DataInput in) throws IOException {
    x = in.readFloat();
    y = in.readFloat();
    z = in.readFloat();

  public String toString() {
    return Float.toString(x) + ", "
        + Float.toString(y) + ", "
        + Float.toString(z);

Listing 5.1: A Point class which implements Writable

Custom Key Types

As written, the Point3D type will work as a value type like we require for the mapper problem described above. But what if we want to emit Point3D objects as keys too? In Hadoop MapReduce, if (key, value) pairs sent to a single reduce task include multiple keys, the reducer will process the keys in sorted order. So key types must implement a stricter interface, WritableComparable. In addition to being Writable so they can be transmitted over the network, they also obey Java’s Comparable interface. The following code listing extends Point3D to meet this interface:

public class Point3D implements WritableComparable {
  public float x;
  public float y;
  public float z;

  public Point3D(float x, float y, float z) {
    this.x = x;
    this.y = y;
    this.z = z;

  public Point3D() {
    this(0.0f, 0.0f, 0.0f);

  public void write(DataOutput out) throws IOException {

  public void readFields(DataInput in) throws IOException {
    x = in.readFloat();
    y = in.readFloat();
    z = in.readFloat();

  public String toString() {
    return Float.toString(x) + ", "
        + Float.toString(y) + ", "
        + Float.toString(z);

  /** return the Euclidean distance from (0, 0, 0) */
  public float distanceFromOrigin() {
    return (float)Math.sqrt(x*x + y*y + z*z);

  public int compareTo(Point3D other) {
    float myDistance = distanceFromOrigin();
    float otherDistance = other.distanceFromOrigin();

    return Float.compare(myDistance, otherDistance);

  public boolean equals(Object o) {
    if (!(other instanceof Point3D)) {
      return false;

    Point3D other = (Point3D)o;
    return this.x == other.x && this.y == other.y
        && this.z == other.z;

  public int hashCode() {
    return Float.floatToIntBits(x)
         ^ Float.floatToIntBits(y)
         ^ Float.floatToIntBits(z);

Listing 5.2: A WritableComparable version of Point3D

It is important for key types to implement hashCode() as well; the section on Partitioners later in this module explains why. The methods hashCode() and equals() have been provided in this version of the class as well.

Using Custom Types

Now that you have created a custom data type, Hadoop must be told to use it. You can control the output key or value data type for a job by using the setOutputKeyClass() and setOutputValueClass() methods of the JobConf object that defines your job. By default, this will set the types expected as output from both the map and reduce phases. If your Mapper emits different types than the Reducer, you can set the types emitted by the mapper with the JobConf‘s setMapOutputKeyClass() and setMapOutputValueClass() methods. These implicitly set the input types expected by the Reducer. The types delivered as input to the Mapper are governed by the InputFormat used; see the next section of this module for more details.

Faster Comparison Operations

The default sorting process for keys will read instances of the key type in from a stream, parsing the byte stream with the readFields() method of the key class, and then call the compareTo() method of the key class on the two objects. For faster performance, it may be possible to decide on an ordering between two keys just by looking at the byte streams and without parsing all of the data contained therein. For example, consider comparing strings of text. If characters are read in sequentially, then a decision can be made on their ordering as soon as a character position is found where the two strings differ. Even if all of the bytes for the object must be read in, the object itself does not necessarily need to be instantiated around those bytes. To support this higher-speed sorting mechanism, you can extend the WritableComparator class with a comparator specific to your own data type. In particular, the method which should be overridden is

public int compare(byte[] b1, int s1, int l1, byte[] b2, int s2, int l2)

The default implementation is in the class org.apache.hadoop.io.WritableComparator. The relevant method has been reproduced here:

  public int compare(byte[] b1, int s1, int l1, byte[] b2, int s2, int l2) {
    try {
      buffer.reset(b1, s1, l1);                   // parse key1

      buffer.reset(b2, s2, l2);                   // parse key2

    } catch (IOException e) {
      throw new RuntimeException(e);

    return compare(key1, key2);                   // compare them

Its operation is exactly as described above; it performs the straightforward comparison of the two objects after they have been individually deserialized from their separate byte streams (the b variables), which each have their own start offset (s) and length (l) attributes. Both objects must be fully constructed and deserialized before comparison can occur. The Text class, on the other hand, allows incremental comparison via its own implementation of this method. The code from org.apache.hadoop.io.Text is shown here:

   /** A WritableComparator optimized for Text keys. */
  public static class Comparator extends WritableComparator {
    public Comparator() {

    public int compare(byte[] b1, int s1, int l1,
                       byte[] b2, int s2, int l2) {
      int n1 = WritableUtils.decodeVIntSize(b1[s1]);
      int n2 = WritableUtils.decodeVIntSize(b2[s2]);
      return compareBytes(b1, s1+n1, l1-n1, b2, s2+n2, l2-n2);

The Text object is serialized by first writing its length field to the byte stream, followed by the UTF-encoded string. The method decodeVIntSize determines the length of the integer describing the length of the byte stream. The comparator then skips these bytes, directly comparing the UTF-encoded bytes of the actual string-portion of the stream in the compareBytes() method. As soon as it finds a character in which the two streams differ, it returns a result without examining the rest of the strings.

Note that you do not need to manually specify this comparator’s use in your Hadoop programs. Hadoop automatically uses this special comparator implementation for Text data due to the following code being added to Text‘s static initialization:

  static {
    // register this comparator
    WritableComparator.define(Text.class, new Comparator());

Final Writable Notes

Defining custom writable types allows you to intelligently use Hadoop to manipulate higher-level data structures, without needing to use toString() to convert all your data types to text for sending over the network. If you will be using a type in a lot of MapReduce jobs, or you must process a very large volume of them (as is usually the case in Hadoop), defining your own data type classes will provide a significant performance benefit.

Exercise: Assume that we have a mapper which emits line segments as keys and values. A line segment is defined by its endpoints. For our purposes, line segments can be ordered by their lengths. Implement a LineSegment class which implements WritableComparable. Hint: make use of Point3D objects.

Input Formats

The InputFormat defines how to read data from a file into the Mapper instances. Hadoop comes with several implementations of InputFormat; some work with text files and describe different ways in which the text files can be interpreted. Others, like SequenceFileInputFormat, are purpose-built for reading particular binary file formats. These types are described in more detail in Module 4.

More powerfully, you can define your own InputFormat implementations to format the input to your programs however you want. For example, the default TextInputFormat reads lines of text files. The key it emits for each record is the byte offset of the line read (as a LongWritable), and the value is the contents of the line up to the terminating '\n' character (as a Text object). If you have multi-line records each separated by a $ character, you could write your own InputFormat that parses files into records split on this character instead.

Another important job of the InputFormat is to divide the input data sources (e.g., input files) into fragments that make up the inputs to individual map tasks. These fragments are called “splits” and are encapsulated in instances of the InputSplit interface. Most files, for example, are split up on the boundaries of the underlying blocks in HDFS, and are represented by instances of the FileInputSplit class. Other files may be unsplittable, depending on application-specific data. Dividing up other data sources (e.g., tables from a database) into splits would be performed in a different, application-specific fashion. When dividing the data into input splits, it is important that this process be quick and cheap. The data itself should not need to be accessed to perform this process (as it is all done by a single machine at the start of the MapReduce job).

The TextInputFormat divides files into splits strictly by byte offsets. It then reads individual lines of the files from the split in as record inputs to the Mapper. The RecordReader associated with TextInputFormat must be robust enough to handle the fact that the splits do not necessarily correspond neatly to line-ending boundaries. In fact, the RecordReader will read past the theoretical end of a split to the end of a line in one record. The reader associated with the next split in the file will scan for the first full line in the split to begin processing that fragment. All RecordReader implementations must use some similar logic to ensure that they do not miss records that span InputSplit boundaries.

Custom File Formats

In this section we will describe how to develop a custom InputFormat that reads files of a particular format.

Rather than implement InputFormat directly, it is usually best to subclass the FileInputFormat. This abstract class provides much of the basic handling necessary to manipulate files. If we want to parse the file in a particular way, then we must override the getRecordReader() method, which returns an instance of RecordReader: an object that can read from the input source. To motivate this discussion with concrete code, we will develop an InputFormat and RecordReader implementation which can read lists of objects and positions from files. We assume that we are reading text files where each line contains the name of an object and then its coordinates as a set of three comma-separated floating-point values. For instance, some sample data may look like the following:

ball, 3.5, 12.7, 9.0
car, 15, 23.76, 42.23
device, 0.0, 12.4, -67.1

We must read individual lines of the file, separate the key (Text) from the three floats, and then read those into a Point3D object as we developed earlier.

The ObjectPositionInputFormat class itself is very straightforward. Since it will be reading from files, all we need to do is define a factory method for RecordReader implementations:

public class ObjectPositionInputFormat extends
    FileInputFormat<Text, Point3D> {

  public RecordReader<Text, Point3D> getRecordReader(
      InputSplit input, JobConf job, Reporter reporter)
      throws IOException {

    return new ObjPosRecordReader(job, (FileSplit)input);

Listing 5.3: InputFormat for object-position files

Note that we define the types of the keys and values emitted by the InputFormat in its definition; these must match the types read in as input by the Mapper in its class definition.

The RecordReader implementation is where the actual file information is read and parsed. We will implement this by making use of the LineRecordReader class; this is the RecordReader implementation used by TextInputFormat to read individual lines from files and return them unparsed. We will wrap the LineRecordReader with our own implementation which converts the values to the expected types. By using LineRecordReader, we do not need to worry about what happens if a record spans an InputSplit boundary, since this underlying record reader already has logic to take care of this fact.

class ObjPosRecordReader implements RecordReader<Text, Point3D> {

  private LineRecordReader lineReader;
  private LongWritable lineKey;
  private Text lineValue;

  public ObjPosRecordReader(JobConf job, FileSplit split) throws IOException {
    lineReader = new LineRecordReader(job, split);

    lineKey = lineReader.createKey();
    lineValue = lineReader.createValue();

  public boolean next(Text key, Point3D value) throws IOException {
    // get the next line
    if (!lineReader.next(lineKey, lineValue)) {
      return false;

    // parse the lineValue which is in the format:
    // objName, x, y, z
    String [] pieces = lineValue.toString().split(",");
    if (pieces.length != 4) {
      throw new IOException("Invalid record received");

    // try to parse floating point components of value
    float fx, fy, fz;
    try {
      fx = Float.parseFloat(pieces[1].trim());
      fy = Float.parseFloat(pieces[2].trim());
      fz = Float.parseFloat(pieces[3].trim());
    } catch (NumberFormatException nfe) {
      throw new IOException("Error parsing floating point value in record");

    // now that we know we'll succeed, overwrite the output objects

    key.set(pieces[0].trim()); // objName is the output key.

    value.x = fx;
    value.y = fy;
    value.z = fz;

    return true;

  public Text createKey() {
    return new Text("");

  public Point3D createValue() {
    return new Point3D();

  public long getPos() throws IOException {
    return lineReader.getPos();

  public void close() throws IOException {

  public float getProgress() throws IOException {
    return lineReader.getProgress();

Listing 5.4: RecordReader for object-position files

You can control the InputFormat used by your MapReduce job with the JobConf.setInputFormat() method.

Exercise: Write an InputFormat and RecordReader that read strings of text separated by '$' characters instead of newlines.

Alternate Data Sources

An InputFormat describes both how to present the data to the Mapper and where the data originates from. Most implementations descend from FileInputFormat, which reads from files on the local machine or HDFS. If your data does not come from a source like this, you can write an InputFormat implementation that reads from an alternate source. For example, HBase (a distributed database system) provides a TableInputFormat that reads records from a database table. You could imagine a system where data is streamed to each machine over the network on a particular port; the InputFormat reads data from the port and parses it into individual records for mapping.

Output Formats

The InputFormat and RecordReader interfaces define how data is read into a MapReduce program. By analogy, the OutputFormat and RecordWriter interfaces dictate how to write the results of a job back to the underlying permanent storage. Several useful OutputFormat implementations are described in Module 4. The default format (TextOutputFormat) will write (key, value) pairs as strings to individual lines of an output file (using the toString() methods of the keys and values). The SequenceFileOutputFormat will keep the data in binary, so it can be later read quickly by the SequenceFileInputFormat. These classes make use of the write() and readFields() methods of the specific Writable classes used by your MapReduce pass.

You can define your own OutputFormat implementation that will write data to an underlying medium in the format that you control. If you want to write to output files on the local system or in HDFS, you should extend the FileOutputFormat abstract class. When you want to use a different output format, you can control this with the JobConf.setOutputFormat() method.

Why might we want to define our own OutputFormat? A custom OutputFormat allows you to exactly control what data is put into a file, and how it is laid out. Suppose another process you use has a custom input file format. Your MapReduce job is supposed to generate inputs compatible with this program. You may develop an OutputFormat implementation which will produce the correct type of file to work with this subsequent process in your tool chain. As an example of how to write an OutputFormat, we will walk through the implementation of a simple XML-based format developed for this tutorial, XmlOutputFormat. Given a set of (key, value) pairs from the Reducer, (e.g., (k1, v1), (k2, v2), etc…) this will generate a file laid out like so:



The code to generate these files is presented below:

import java.io.DataOutputStream;
import java.io.IOException;

import org.apache.hadoop.fs.FSDataOutputStream;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.io.NullWritable;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapred.FileOutputFormat;
import org.apache.hadoop.mapred.JobConf;
import org.apache.hadoop.mapred.RecordWriter;
import org.apache.hadoop.mapred.Reporter;
import org.apache.hadoop.util.Progressable;

public class XmlOutputFormat<K, V> extends FileOutputFormat {

  protected static class XmlRecordWriter<K, V> implements RecordWriter<K, V> {
    private static final String utf8 = "UTF-8";

    private DataOutputStream out;

    public XmlRecordWriter(DataOutputStream out) throws IOException {
      this.out = out;

     * Write the object to the byte stream, handling Text as a special case.
     * @param o
     *          the object to print
     * @throws IOException
     *           if the write throws, we pass it on
    private void writeObject(Object o) throws IOException {
      if (o instanceof Text) {
        Text to = (Text) o;
        out.write(to.getBytes(), 0, to.getLength());
      } else {

    private void writeKey(Object o, boolean closing) throws IOException {
      if (closing) {
      if (closing) {

    public synchronized void write(K key, V value) throws IOException {

      boolean nullKey = key == null || key instanceof NullWritable;
      boolean nullValue = value == null || value instanceof NullWritable;

      if (nullKey && nullValue) {

      Object keyObj = key;

      if (nullKey) {
        keyObj = "value";

      writeKey(keyObj, false);

      if (!nullValue) {

      writeKey(keyObj, true);

    public synchronized void close(Reporter reporter) throws IOException {
      try {
      } finally {
        // even if writeBytes() fails, make sure we close the stream

  public RecordWriter<K, V> getRecordWriter(FileSystem ignored, JobConf job,
      String name, Progressable progress) throws IOException {
    Path file = FileOutputFormat.getTaskOutputPath(job, name);
    FileSystem fs = file.getFileSystem(job);
    FSDataOutputStream fileOut = fs.create(file, progress);
    return new XmlRecordWriter<K, V>(fileOut);

The FileOutputFormat which XmlOutputFormat subclasses will handle most of the heavy lifting. The only method directly implemented in XmlOutputFormat is getRecordWriter(), which is a factory method for the RecordWriter object which will actually write the file. The inner class XmlRecordWriter is the implementation which generates files in the format shown above. The RecordWriter is initialized with an output stream connected to a file in the output file system. At the same time, the XML prologue is written into the output file. The particular output file system and filename associated with this output stream are determined based on the current job configuration. The XmlRecordWriter‘s write() method is then called each time a (key, value) pair is provided to the OutputCollector by the Reducer. When the Reducer finishes, the close() method of the XmlRecordWriter will write the XML epilogue and close the underlying stream.

Partitioning Data

“Partitioning” is the process of determining which reducer instance will receive which intermediate keys and values. Each mapper must determine for all of its output (key, value) pairs which reducer will receive them. It is necessary that for any key, regardless of which mapper instance generated it, the destination partition is the same. If the key "cat" is generated in two separate (key, value) pairs, they must both be reduced together. It is also important for performance reasons that the mappers be able to partition data independently — they should never need to exchange information with one another to determine the partition for a particular key.

Hadoop uses an interface called Partitioner to determine which partition a (key, value) pair will go to. A single partition refers to all (key, value) pairs which will be sent to a single reduce task. Hadoop MapReduce determines when the job starts how many partitions it will divide the data into. If twenty reduce tasks are to be run (controlled by the JobConf.setNumReduceTasks()) method), then twenty partitions must be filled.

The Partitioner defines one method which must be filled:

public interface Partitioner<K, V> extends JobConfigurable {
  int getPartition(K key, V value, int numPartitions);

The getPartition() method receives a key and a value and the number of partitions to split the data across; a number in the range [0, numPartitions) must be returned by this method, indicating which partition to send the key and value to. For any two keys k1 and k2, k1.equals(k2) implies getPartition(k1, *, n) == getPartition(k2, *, n).

The default Partitioner implementation is called HashPartitioner. It uses the hashCode() method of the key objects modulo the number of partitions total to determine which partition to send a given (key, value) pair to.

For most randomly-distributed data, this should result in all partitions being of roughly equal size. If the data in your data set is skewed in some way, however, this might not produce good results. For example, if you know that the key 0 will appear much more frequently than any other key, then you may want to send all the 0-keyed data to one partition, and distribute the other keys over all other partitions by their hashCode(). Also, if the hashCode() method for your data type does not provide uniformly-distributed values over its range, then data may not be sent to reducers evenly. Poor partitioning of data means that some reducers will have more data input than others, which usually means they’ll have more work to do than the other reducers. Thus the entire job will wait for one reducer to finish its extra-large share of the load, when it might have been possible to spread that across many different reducers.

If you are dissatisfied with the performance of HashPartitioner, you are free to write your own Partitioner implementation. It can be general-purpose, or tailored to the specific data types or values that you expect to use in your application. After implementing the Partitioner interface, use the JobConf.setPartitionerClass() method to tell Hadoop to use it for your job.

Reporting Custom Metrics

The Hadoop system records a set of metric counters for each job that it runs. For example, the number of input records mapped, the number of bytes it reads from or writes to HDFS, etc. To profile your applications, you may wish to record other values as well. For example, if the records sent into your mappers fall into two categories (call them “A” and “B”), you may wish to count the total number of A-records seen vs. the total number of B-records.

The Reporter object passed in to your Mapper and Reducer classes can be used to update counters. The same set of counter variables can be contributed to by all Mapper and Reducer instances across your cluster. The values are aggregated by the master node of the cluster, so they are “thread-safe” in this manner.

Counters are incremented through the Reporter.incrCounter() method. The names of the counters are defined as Java enum‘s. The following example demonstrates how to count the number of “A” vs. “B” records seen by the mapper:

public class MyMapper extends MapReduceBase implements
    Mapper<Text, Text, Text, Text> {

  static enum RecordCounters { TYPE_A, TYPE_B, TYPE_UNKNOWN };

  // actual definitions elided
  public boolean isTypeARecord(Text input) { ... }
  public boolean isTypeBRecord(Text input) { ... }

  public void map(Text key, Text val, OutputCollector<Text, Text> output,
      Reporter reporter) throws IOException {

    if (isTypeARecord(key)) {
      reporter.incrCounter(RecordCounters.TYPE_A, 1);
    } else if (isTypeBRecord(key)) {
      reporter.incrCounter(RecordCounters.TYPE_B, 1);
    } else {
      reporter.incrCounter(RecordCounters.TYPE_UNKNOWN, 1);

    // actually process the record here, call
    // output.collect( .. ), etc.

If you launch your job with JobClient.runJob(), the diagnostic information printed to stdout when the job completes will contain the values of all the counters. Both runJob() and submitJob() will return a RunningJob object that refers to the job in question. The RunningJob.getCounters() method will return a Counters object that contains the values of all the counters so that you can query them programmatically. The Counters.getCounter(Enum key) method returns the value of a particular counter.

Distributing Auxiliary Job Data

The bulk of the data that you process in a MapReduce job will probably be stored in large files spread across the HDFS. You can reliably store petabytes of information in HDFS and individual jobs can process several terabytes at a time. The HDFS access model, however, assumes that the data from a file should be read into a single mapper. The individual files stored in HDFS are very large and can possibly be broken into different chunks for processing in parallel.

Sometimes it is necessary for every Mapper to read a single file; for example, a distributed spell-check application would require every Mapper to read in a copy of the dictionary before processing documents. The dictionary will be small (only a few megabytes), but needs to be widely available so that all nodes can reach it.

Hadoop provides a mechanism specifically for this purpose, called the distributed cache. The distributed cache can contain small data files needed for initialization or libraries of code that may need to be accessed on all nodes in the cluster.

To use the distributed cache to disseminate files, create an instance of the DistributedCache class when setting up your job. Use the DistributedCache.addCacheFile() method to add names of files which should be sent to all nodes on the system. The file names are specified as URI objects; unless qualified otherwise, they assume that the file is present on the HDFS in the path indicated. You can copy local files to HDFS with the FileSystem.copyFromLocalFile() method.

When you want to retrieve files from the distributed cache (e.g., when the mapper is in its configure() step and wants to load config data like the dictionary mentioned above), use the DistributedCache.getLocalCacheFiles() method to retrieve the list of paths local to the current node for the cached files. These are copies of all cached files, placed in the local file system of each worker machine. (They will be in a subdirectory of mapred.local.dir.) Each of the paths returned by getLocalCacheFiles() can be accessed via regular Java file I/O mechanisms, such as java.io.FileInputStream.

As a cautionary note: If you use the local JobRunner in Hadoop (i.e., what happens if you call JobClient.runJob() in a program with no or an empty hadoop-conf.xml accessible), then no local data directory is created; the getLocalCacheFiles() call will return an empty set of results. Unit test code should take this into account.

Suppose that we were writing an inverted index builder. We do not want to include very common words such “the,” “a,” “and,” etc. These so-called stop words might all be listed in a file. All the mappers should read the stop word list when they are initialized, and then filter the index they generate against this list. We can disseminate a list of stop words to all the Mappers with the following code. The first listing will put the stop-words file into the distributed cache:

  public static final String LOCAL_STOPWORD_LIST =

  public static final String HDFS_STOPWORD_LIST = "/data/stop_words.txt";

  void cacheStopWordList(JobConf conf) throws IOException {
    FileSystem fs = FileSystem.get(conf);
    Path hdfsPath = new Path(HDFS_STOPWORD_LIST);

    // upload the file to hdfs. Overwrite any existing copy.
    fs.copyFromLocalFile(false, true, new Path(LOCAL_STOPWORD_LIST),

    DistributedCache.addCacheFile(hdfsPath.toUri(), conf);

This code copies the local stop_words.txt file into HDFS, and then tells the distributed cache to send the HDFS copy to all nodes in the system. The next listing actually uses the file in the mapper:

class IndexMapperExample implements Mapper {
  void configure(JobConf conf) {
    try {
      String stopwordCacheName = new Path(HDFS_STOPWORD_LIST).getName();
      Path [] cacheFiles = DistributedCache.getLocalCacheFiles(conf);
      if (null != cacheFiles && cacheFiles.length > 0) {
        for (Path cachePath : cacheFiles) {
          if (cachePath.getName().equals(stopwordCacheName)) {
    } catch (IOException ioe) {
      System.err.println("IOException reading from distributed cache");

  void loadStopWords(Path cachePath) throws IOException {
    // note use of regular java.io methods here - this is a local file now
    BufferedReader wordReader = new BufferedReader(
        new FileReader(cachePath.toString()));
    try {
      String line;
      this.stopWords = new HashSet<String>();
      while ((line = wordReader.readLine()) != null) {
    } finally {

  /* actual map() method, etc go here */

The code above belongs in the Mapper instance associated with the index generation process. We retrieve the list of files cached in the distributed cache. We then compare the basename of each file (using Path.getName()) with the one we expect for our stop word list. Once we find this file, we read the words, one per line, into a Set instance that we will consult during the mapping process.

The distributed cache has additional uses too. For instance, you can use the DistributedCache.addArchiveToClassPath() method to send a .jar file to all the nodes. It will be inserted into the classpath as well, so that classes in the archive can be accessed by all the nodes.

Distributing Debug Scripts

Hadoop will generate a large number of log files for a job, distributed across all the nodes that participated in the job’s execution. Often times only a subset of these logs will be of interest when debugging failing tasks. MapReduce can help with this by running a user-provided script when either a map or reduce task fails. These scripts are provided the names of files containing the stdout and stderr from the task, as well as the task’s Hadoop log and job.xml file (i.e., its complete JobConf in serialized form).

These scripts will be run on whichever node encounters failing tasks. You can use these scripts to perform automation to allow you to more easily inspect only the failing tasks: e.g., email the stdout/stderr to an administrator email address; upload the failed task’s log files to a common NFS-mounted “debug dump” directory, preserve local state modifications made by map tasks, etc.

Separate scripts can be provided for map and reduce task failure. They each receive as arguments, in order, the names of files containing the task’s stdout, stderr, syslog, and jobconf. Because they are run on all the task nodes, and not on the client machine where the job was submitted, these scripts must be sent to the nodes through the distributed cache listed above.

The following method will enable failed task scripts on a MapReduce job being prepared. It assumes that you have given it the names of two scripts (e.g., bash scripts) which do your debug actions with the log filenames provided (e.g., copy them to a shared NFS mount). In this script these are /home/aaron/src/map-fail and reduce-fail.

  private static final String FAILED_MAP_SCRIPT_NAME = "map-fail";
  private static final String FAILED_REDUCE_SCRIPT_NAME = "reduce-fail";

  private static final String HDFS_SCRIPT_DIR = "/debug";

  private static final String HDFS_FAILED_MAP_SCRIPT =
  private static final String HDFS_FAILED_REDUCE_SCRIPT =
  private static final String LOCAL_FAILED_MAP_SCRIPT  =
    "/home/aaron/src/" + FAILED_MAP_SCRIPT_NAME;
  private static final String LOCAL_FAILED_REDUCE_SCRIPT =
    "/home/aaron/src/" + FAILED_REDUCE_SCRIPT_NAME;

  public static void setupFailedTaskScript(JobConf conf) throws IOException {

    // create a directory on HDFS where we'll upload the fail scripts
    FileSystem fs = FileSystem.get(conf);
    Path debugDir = new Path(HDFS_SCRIPT_DIR);

    // who knows what's already in this directory; let's just clear it.
    if (fs.exists(debugDir)) {
      fs.delete(debugDir, true);

    // ...and then make sure it exists again

    // upload the local scripts into HDFS
    fs.copyFromLocalFile(new Path(LOCAL_FAILED_MAP_SCRIPT),
        new Path(HDFS_FAILED_MAP_SCRIPT));
    fs.copyFromLocalFile(new Path(LOCAL_FAILED_REDUCE_SCRIPT),

    conf.setMapDebugScript("./" + FAILED_MAP_SCRIPT_NAME);
    conf.setReduceDebugScript("./" + FAILED_REDUCE_SCRIPT_NAME);

    URI fsUri = fs.getUri();

    String mapUriStr = fsUri.toString() + HDFS_FAILED_MAP_SCRIPT
        + "#" + FAILED_MAP_SCRIPT_NAME;
    URI mapUri = null;
    try {
      mapUri = new URI(mapUriStr);
    } catch (URISyntaxException use) {
      throw new IOException(use);

    DistributedCache.addCacheFile(mapUri, conf);

    String reduceUriStr = fsUri.toString() + HDFS_FAILED_REDUCE_SCRIPT
    URI reduceUri = null;
    try {
      reduceUri = new URI(reduceUriStr);
    } catch (URISyntaxException use) {
      throw new IOException(use);

    DistributedCache.addCacheFile(reduceUri, conf);

How does this all work? The scripts are, presumably, initially hosted on the client machine that is submitting the job. The client is responsible for injecting them into HDFS and enabling them in the distributed cache. It first creates the HDFS_SCRIPT_DIR and then uploads the local script files into this directory.

It must then set the commands for the TaskTracker to execute to run the scripts. This is accomplished by the lines:

    conf.setMapDebugScript("./" + FAILED_MAP_SCRIPT_NAME);
    conf.setReduceDebugScript("./" + FAILED_REDUCE_SCRIPT_NAME);

The distributed cache copies the files to the mapred.local.dir on each task node. The TaskTracker will then execute the scripts if necessary. But the TaskTracker does not run with its working directory set to mapred.local.dir. Fortunately, the distributed cache can be told to create symlinks in the working directory for files in the distributed cache. The third line of the snippit above enables this functionality. Now ./FAILED_MAP_SCRIPT_NAME will point to the copy of FAILED_MAP_SCRIPT_NAME in the local cache directory, and the script can be run.

But before that can happen, we must add the files themselves to the distributed cache. (As of now they are only in HDFS.) Ordinarily, we could just call DistributedCache.addCacheFile(new Path("hdfs_path_to_some_file").toUri()) on a filename and that would be sufficient. But since we need to create symlinks, we must provide the distributed cache with information as to how the symlink should be created–what filename it should take in the working directory. This is provided as the URI “anchor” part following the “#” in the URI. A subtlety of Hadoop’s Path class is that if you put a ‘#’ in the path string, it will URL-encode it and treat it as part of the filename. Therefore, we use some extra code to construct our URIs manually to ensure that the ‘#’ remains unescaped.

Using Amazon Web Services

Hadoop’s power comes from its ability to perform work on a large number of machines simultaneously. What if you want to experiment with Hadoop, but do not have many machines? While operations on a two or four-node cluster are functionally equivalent to those on a 40 or 100-node cluster, processing larger volumes of data will require a larger number of nodes.

Amazon provides machines for rent on demand through their Elastic Compute Cloud (a.k.a. EC2) service. EC2 is part of a broader set of services collectively called the Amazon Web Services, or AWS. EC2 allows you request a set of nodes (“instances” in their parlance) for as long as you need them. You pay by the instance*hour, plus costs for bandwidth. You can use EC2 instances to run a Hadoop cluster. Hadoop comes with a set of scripts which will provision EC2 instances.

The first step in this process is visit the EC2 web site (link above) and click “Sign Up For This Web Service”. You will need to create an account and provide billing information. Then follow the instructions in the Getting started guide to set up your account and configure your system to run the AWS tools.

Once you have done so, follow the instructions in the Hadoop wiki specific to running Hadoop on Amazon EC2. While more details are available in the above document, the shortest steps to provisioning a cluster are:

  • Edit src/contrib/ec2/bin/hadoop-ec2-env.sh to contain your Amazon account information and parameters about the desired cluster size.
  • Execute src/contrib/ec2/bin/hadoop-ec2 launch-cluster.

After the cluster has been started, you can log in to the head node over ssh with the bin/hadoop-ec2 login script, and perform your MapReduce computation. When you are done, log out and type bin/hadoop-ec2 terminate-cluster to release the EC2 instances. The contents of the virtual hard drives on the instances will disappear, so be sure to copy off any important data with scp or another tool first!

A very thorough introduction to configuring Hadoop on EC2 and running a test job is provided in the Amazon Web Services Developer Wiki site.


API Reference – Current Hadoop documentation

Amazon Web Services EC2 Getting Started Guide – Amazon, Inc.

Hadoop on Amazon EC2 – Official Hadoop wiki page

Running Hadoop MapReduce on Amazon EC2 and Amazon S3 – Tom White, Amazon Web Services Developer Connection, July 2007


Module 6: Related Topics


Hadoop by itself allows you to store and process very large volumes of data. However, building a large-scale distributed system can require functionality not provided by this base. Several other tools and systems have been created to fill the gaps and deliver a more full-featured set of distributed systems engineering tools.

Goals for this Module:

  • Understand how distributed consensus systems can be used to bootstrap larger distributed systems.
  • Understand how to write queries in the Pig log-processing language


  1. Introduction
  2. Goals for this Module
  3. Outline
  4. ZooKeeper
    1. Motivation
    2. Data Storage in ZooKeeper
    3. ZooKeeper Applications
    4. Distributed Consensus
  5. Pig
    1. Motivation
    2. Pig Latin
    3. Pig Latin Data Types
    4. Loading Data Into Pig
    5. Pig Latin Operators
    6. Setting Up Pig
  6. References



Suppose you are building a large-scale distributed system. Several different services need to be brought online and must discover one another. It is not guaranteed that each service will have a fixed master IP address. For example, it may be the case that you start the same service on 100 nodes, and they elect a master from among whichever of the 100 boots first. Each of these disparate services must communicate with each other. How do all the nodes of each service find the master IP address of each other service? How do all the nodes in a single service agree on which one of them becomes the master?

ZooKeeper is a service designed to handle all of these problems. ZooKeeper will allow you to store small amounts of information in a central location. It provides coordinated access to this information. Most importantly, it provides high-availability: the ZooKeeper service is intended to run on a set of several machines, which prevents the loss of individual nodes from bringing down the cluster. But these nodes communicate information in a careful way, ensuring that all nodes in the ZooKeeper cluster provide the same consistent answer for a query, regardless of which ZooKeeper server you contact.

Data Storage in ZooKeeper

Several ZooKeeper daemons can be started on different machines. Clients can connect to any daemon in the cluster; the clients will always see the same view of the ZooKeeper world regardless of which daemon they connect to. User data is stored in objects with a hierarchical addressing system similar to that used by a conventional file system. It has a root object named /, and additional nodes can be extended off of this in a tree-like fashion. Each node of the tree can both hold data (i.e., act like a file) and have child nodes (i.e., act like a directory). The amount of data that can be stored in an object is small: there is a hard cap of 1 MB. The reason for the cap is so that the entire data store can be stored in the RAM of the ZooKeeper machines. This allows requests to be dispatched with high throughput. Changes are written to disk to provide permanence, but read requests are entirely handled by the data cached in memory. This is usually not a major limitation; the data stored at a node is intended to be used as a pointer. For example, ZooKeeper may know about the filename in another conventional file system, which contains the authoritative configuration file for a distributed system. The distributed system components first contact ZooKeeper to get the definitive filename, and then fetch that file for the configuration.

ZooKeeper Applications

ZooKeeper can be used for a variety of distributed coordination tasks. In addition to leader election, system bootstrapping, and various types of locks (mutual exclusion, reader/writer, etc), other synchronization primitives such as barriers, producer/consumer queues, priority queues, and multi-phase commit operations can be encoded in ZooKeeper. The ZooKeeper tutorial and recipes pages describe how to implement these algorithms. ZooKeeper itself is implemented in Java, but provides APIs for both Java- and C-based programs.

ZooKeeper can also be used as a central message board for an application. Individual nodes of a distributed system can store their current operational status in ZooKeeper for easy central reporting. The ZooKeeper service can also be used to form sub-groups of nodes or other hierarchical arrangements within a distributed system.

As mentioned, data stored in ZooKeeper is accessed by manipulating the nodes in the data hierarchy. This is done in a manner similar to file system access. But ZooKeeper does not implement the POSIX file system API. On the other hand, it also adds a set of other primitives not ordinarily found in a file system. Nodes can be opened with a number of special flags. One such flag is “ephemeral,” meaning that the node disappears when the client who opened it disconnects. Another such flag is “sequence,” which means that ZooKeeper will append a sequential id number to the node name you are trying to create. These id numbers are handed out in order, and the same id number is not reused. ZooKeeper does not provide exclusive locks on nodes directly, but a lock can be created by careful use of the ephemeral and sequence flags. The ZooKeeper recipes wiki page describes how to implement global locks using these flags. It also describes protocols for implementing shared (reader-writer) and revocable locks.

Distributed Consensus

A reasonable question is how the ZooKeeper service can function across multiple nodes and remain synchronized. If distributed synchronization is why your services must use ZooKeeper, how does ZooKeeper itself bootstrap this capability?

ZooKeeper implements a distributed consensus protocol. ZooKeeper internally uses a leader election protocol such as Paxos to determine which node in the ZooKeeper service is the master. While clients connect to any node in the ZooKeeper service, these additional nodes will forward agreed-upon facts back to clients. Updates to the shared state require the intervention of the master. All updates to the shared state are ordered with timestamps. These timestamped updates are then disseminated to the nodes in the ZooKeeper service. When a majority of nodes acknowledge an update, it is said to be held by a quorum of the nodes. Any fact that a quorum has agreed upon may be returned to clients. Conversely, any updates that have not reached a quorum will not be returned to the clients. The timestamps are used to order the updates to elements of the data store. If multiple updates are made to the state of an individual node, the newest update is used.

The use of a quorum ensures that the service always returns consistent answers. Because a vote is effectively held before returning a response, any nodes which hold stale data will be outvoted by the nodes with more current information. This also makes ZooKeeper resilient to failure. Up to 49% of the ZooKeeper service nodes can shut down or become desynchronized before ZooKeeper loses its ability to authoritatively answer responses. So if 11 nodes run ZooKeeper, up to 5 of these may disconnect without incident. After more than half the nodes fail, ZooKeeper will refuse service until the machines are restored.

If the node of the ZooKeeper cluster which was elected leader fails, then a new leader election will be held and the cluster will continue to function.

The reason for electing a leader in such a system is to ensure that timestamps assigned to updates are only issued by a single authority. ZooKeeper is designed to reduce or eliminate possible race conditions in distributed applications.

One consequence of ZooKeeper’s design is that it is intended to serve many more read requests than writes. A ZooKeeper cluster can handle hundreds or thousands of clients, issuing many tens of thousands of requests per second–if the majority of these requests (90–99%) are reads. Only a small fraction should be updates.

ZooKeeper Example

The following code excerpt shows how to use ZooKeeper to implement a “barrier.” A barrier separates a process into two logical halves. Multiple machines running in coordination with one another will all perform the first half of the process. No machine can begin the second half of the process until everyone has completed the first half. The barrier sits between these processes. As nodes reach the barrier, they all wait until everyone has reached the barrier. Then all nodes are released to begin the second half. A distributed barrier implementation written for ZooKeeper follows:

Watcher watcher = new Watcher() {
  public void process(WatchEvent event) {}

ZooKeeper zk = new ZooKeeper(hosts, 3000, watcher);

Object notifyObject = new Object();
String root;
int size;

Barrier(ZooKeeper zk, String name, int size) throws KeeperException, InterruptedException {
  this.zk = zk;
  this.root = name;
  this.size = size;
  // Make sure the  barrier node exists
  try {
    zk.create(root, new byte[0], Ids.OPEN_ACL_UNSAFE, 0);
  } catch (NodeExistsException e) {}

/** work with everyone **/

 * Join barrier
 * @return
 * @throws KeeperException
 * @throws InterruptedException */
boolean enter() throws KeeperException, InterruptedException {
  zk.create(root + "/" + name, new byte[0], Ids.OPEN_ACL_UNSAFE, CreateFlags.EPHEMERAL);
  while (true) {
    synchronized (notifyObject) {
      ArrayList<String> list = zk.getChildren(root, new Watcher() {
        public void process(Event e) { notifyObject.notifyAll(); }

      if (list.size() < size)  {
      } else {
        return true;

 * Wait until all reach barrier
 * @return
 * @throws KeeperException
 * @throws InterruptedException */
boolean leave() throws KeeperException, InterruptedException {
  zk.delete(root + "/" + name, 0);
  while (true) {
    synchronized (notifyObject) {
      ArrayList<String> list = zk.getChildren(root, new Watcher() {
        public void process(Event e) { notifyObject.notifyAll(); }

      if (list.size() > 0) {
      } else {
        return true;

Listing 6.1: ZooKeeper Barrier Example



Pig is a platform for analyzing large data sets. Pig’s language, Pig Latin, lets you specify a sequence of data transformations such as merging data sets, filtering them, and applying functions to records or groups of records. Users can create their own functions to do special-purpose processing.

Pig Latin programs execute in a distributed fashion on a cluster. Our current implementation compiles Pig Latin programs into Map/Reduce jobs, and executes them using Hadoop on Kryptonite.

Thur purpose of Pig is to answer queries over semi-structured data such as log files. Large volumes of data are in mostly-organized formats such as log files, which define a set of standard fields for each entry. While the MapReduce programming model on top of Hadoop provides a convenient mechanism for delivering a large volume of log-structured information to an analysis program, writing analyses of mostly-structured information involves writing a large amount of tedious processing code.

Pig is a high-level language for writing queries over this sort of data. A query planner then compiles queries written in this language (called “Pig Latin”) into maps and reduces which are then executed on a Hadoop cluster.

Anything which could be written in Pig can also be implemented as straight Java-based Hadoop MapReduce. But while individual programmers could develop their own suite of data analysis functions such as join, order by, etc., this requires individual programmers to develop their own (non-standard) libraries, and test them. Pig provides a tested and supported suite of the most common data-aggregation functions. It also allows programmers to provide their own application-specific code for purposes of loading and saving data, as well as for performing more complicated record-by-record evaluations.

Pig Latin

The programming language used to write Pig queries is called Pig Latin.

The MapReduce programming model can be thought of as composed of three distinct phases:

  1. Process input records
  2. Form groups of related records
  3. Process groups into outputs

In MapReduce, the first two of these steps are handled by the mapper, and the third step is handled by the reducer. Pig Latin exposes explicit primitives that perform actions from each phase. These primitives can be composed and reordered. Furthermore, it includes additional primitives to handle operations such as filtering and joining data sets.

Pig Latin Data Types

Values in Pig Latin can be expressed by four basic data types:

  • An atom is any atomic value (e.g., "fish")
  • A tuple is a record of multiple values with fixed arity. e.g., ("dog", "sparky").
  • A data bag is a collection of an arbitrary number of values. e.g., {("dog", "sparky"), ("fish", "goldie")}. Data bags support a scan operation for iterating through their contents.
  • A data map is a collection with a lookup function translating keys to values. e.g., ["age" : 25]

All data types are fully nestable; bags may contain tuples, and maps may contain bags or other maps, etc. This differs from a traditional database model, where data must be normalized into lists of atoms. By allowing data types to be composed in this manner, Pig queries line up better to the conceptual model of the data held by the programmer. Data types may also be heterogeneous. For example, the fields of a tuple may each have different types; some may be atoms, others may be more tuples, etc. The values in a bag may hold different types, as may the values in data maps. These can vary from one record to the next in the bag. Data map keys must be atoms, for efficiency reasons.

Loading Data Into Pig

The first step in using Pig is to load data into a program. Pig provides a LOAD statement for this purpose. Its format is: result = LOAD 'filename' USING fn() AS (field1, field2, ...).

This statement returns a bag of values of all the data contained in the named file. Each record in the bag is a tuple, with the fields named by field1, field2, etc. The fn() is a user-provided function that reads in the data. Pig supports user-provided Java code throughout to handle the application-specific bits of parsing. Pig Latin itself is the “glue” that then holds these application-specific functions together, routing records and other data between them.

An example data loading command is:

queries = LOAD 'query_log.txt'
          USING myLoad()
          AS (userId, queryString, timestamp)

The user-defined functions to load data (e.g., myLoad()) do not need to be provided. A default function for loading data exists, which will parse tab-delimited records. If the programmer did not specify field names in the AS clause, they would be addressed by positional parameters: $0, $1, and so forth.

The default loader is called PigStorage(). This loader can read files containing character-delimited tuple records. These tuples must contain only atomic values; e.g., cat, turtle, fish. Other loaders are listed in the PigBuiltins page of the Pig wiki. PigStorage() takes as an argument the character to use to delimit fields. For example, to load a table of three tab-delimited fields, the following statement can be used:

data = LOAD 'tab_delim_data.txt' USING PigStorage('\t') AS (user, time, query)

A different argument could be passed to PigStorage() to read comma- or space-delimited fields.

Pig Latin Operators

Pig Latin provides a number of operators which filter, join, or otherwise organize data.

FOREACH: The FOREACH command operates on each element of a data bag. This is useful, for instance, for processing each input record in a bag returned by a LOAD statement.

FOREACH bagname GENERATE expression, expression...

This statement iterates over the contents of a bag. It applies the expressions on the right of the GENERATE keyword to the data provided by the current record emitted from the bag. The expressions may be, for example, the names of fields. So to extract the names of all users who accessed the site (based on the query_log.txt example shown above), we could write a query like:

FOREACH queries GENERATE userId;

In the FOREACH statement, each element of the bag is considered independently. There are no expressions which reference multiple elements being extracted from the bag’s iterator at a time; this allows the statement to be processed in parallel using Hadoop MapReduce.

Expressions emitted by the GENERATE element are not limited to the names of fields; they can be fields (by name like userId or by position like $0), constants, algebraic operations, map lookups, conditional expressions, or FLATTEN expressions, described below.

Finally, these expressions may also call user-provided functions that are written in Java. These user-provided functions have access to the entire current record through a Pig library; in this way, Pig can be used as the heavy-lifting component to automate record-by-record mapping using an application-specific Java function to perform tricky parsing or evaluation logic. Pig also provides several of the most commonly-needed functions, such as COUNT, AVG, MIN, MAX, and SUM.

FLATTEN is an expression which will eliminate a level of nesting. Given a tuple which contains a bag, FLATTEN will emit several tuples each of which contains one record from the bag. For example, if we had a bag of records containing a person’s name and a list of types of pets they own:

(Alice, { turtle, goldfish, cat })
(Bob, { dog, cat })

A FLATTEN command would eliminate the inner bags like so:

(Alice, turtle)
(Alice, goldfish)
(Alice, cat)
(Bob, dog)
(Bob, cat)

FILTER statements iterate over a bag and return a new bag containing all elements which pass a conditional expression, e.g.:

adults = FILTER people BY age > 21;

The COGROUP and JOIN operations perform similar functions: they unite related data elements from multiple data sets. The difference is that JOIN acts like the SQL JOIN statement, creating a flat set of output records containing the joined cross-product of the input records. The COGROUP operator, on the other hand, groups the elements by their common field and returns a set of records each containing two separate bags. The first bag is the records of the first data set with the common field, and the second bag is the records of the second data set containing the common field.

To illustrate the difference, suppose we had the flattened data set mapping people to their pets, and another flattened data set mapping people to their friends. We could create a “pets of friends” data set out of these like the following. Here are the input data sets:

pets: (owner, pet)
(Alice, turtle)
(Alice, goldfish)
(Alice, cat)
(Bob, dog)
(Bob, cat)

friends: (friend1, friend2)
(Cindy, Alice)
(Mark, Alice)
(Paul, Bob)

Here is what is returned by COGROUP:

COGROUP pets BY owner, friends BY friend2; returns:

( Alice, {(Alice, turtle), (Alice, goldfish), (Alice, cat)},
         {(Cindy, Alice), (Mark, Alice)} )
( Bob, {(Bob, dog), (Bob, cat)}, {(Paul, Bob)} )

Contrasted with the more familiar, non-hierarchical JOIN operator:

JOIN pets BY owner, friends BY friend2; returns:

(Alice, turtle, Cindy)
(Alice, turtle, Mark)
(Alice, goldfish, Cindy)
(Alice, goldfish, Mark)
(Alice, cat, Cindy)
(Alice, cat, Mark)
(Bob, dog, Paul)
(Bob, cat, Paul)

In general, COGROUP command supports grouping on as many data sets as are desired. Three or more data sets can be joined in this fashion. It is also possible to group up elements of only a single data set; this is supported through an alternate keyword, GROUP.

A GROUP ... BY statement will organize a bag of records into bags of related items based on the field identified as their common key field. e.g., the pets bag from the previous example could be grouped up with:

GROUP pets BY owner; returns:

( Alice, {(Alice, turtle), (Alice, goldfish), (Alice, cat)} )
( Bob, {(Bob, dog), (Bob, cat)} )

In this way, GROUP and FLATTEN are effectively inverses of one another.

More complicated statements can be realized as well: operations which expect a data set as input do not need to use an explicitly-named data set; they can use one generated “inline” with another FILTER, GROUP or other statement.

When the final data set has been created by a Pig Latin script, the output can be saved to a file with the STORE command, which follows the form:

STORE data set INTO 'filename' USING function()

The provided function specifies how to serialize the data to the file; if it is omitted, then a default serializer will write plain-text tab-delimited files.

A number of additional operators exist for the purposes of removing duplicate records, sorting records, etc. This paper explains the additional operators and expression syntaxes in greater detail.

Setting Up Pig

Pig is an Apache incubator project; it has not made official packaged releases, but the source code can be retrieved from their subversion repository.

The Pig Incubator home page contains instructions on retrieving the Pig sources and compiling them.


Module 7: Managing a Hadoop Cluster


Hadoop can be deployed on a variety of scales. The requirements at each of these will be different. Hadoop has a large number of tunable parameters that can be used to influence its operation. Furthermore, there are a number of other technologies which can be deployed with Hadoop for additional capabilities. This module describes how to configure clusters to meet varying needs in terms of size, processing power, and reliability and availability.

Goals for this Module:

  • Understand differences in requirements for different sizes of Hadoop clusters
  • Learn how to configure Hadoop for a variety of deployment scopes


  1. Introduction
  2. Goals for this Module
  3. Outline
  4. Basic Setup
    1. Java Requirements
    2. Operating System
    3. Downloading and Installing Hadoop
  5. Important Directories
  6. Selecting Machines
  7. Cluster Configurations
    1. Small Clusters: 2-10 Nodes
    2. Medium Clusters: 10-40 Nodes
    3. Large Clusters: Multiple Racks
  8. Performance Monitoring
    1. Ganglia
    2. Nagios
  9. Additional Tips
  10. References & Resources

Basic Setup

This section discusses the general platform requirements for Hadoop.

Java Requirements

Hadoop is a Java-based system. Recent versions of Hadoop require Sun Java 1.6.

Compiling Java programs to run on Hadoop can be done with any number of commonly-used Java compilers. Sun’s compiler is fine, as is ecj, the Eclipse Compiler for Java. A bug in gcj, the GNU Compiler for Java, causes incompatibility between generated classes and Hadoop; it should not be used.

Operating System

As Hadoop is written in Java, it is mostly portable between different operating systems. Developers can and do run Hadoop under Windows. The various scripts used to manage Hadoop clusters are written in a UNIX shell scripting language that assumes sh- or bash-like behavior. Thus running Hadoop under Windows requires cygwin to be installed. The Hadoop documentation stresses that a Windows/cygwin installation is for development only. The vast majority of server deployments today are on Linux. (Other POSIX-style operating systems such as BSD may also work. Some Hadoop users have reported successfully running the system on Solaris.) The instructions on this page assume a command syntax and system design similar to Linux, but can be readily adapted to other systems.

Downloading and Installing Hadoop

Hadoop is available for download from the project homepage at http://hadoop.apache.org/core/releases.html. Here you will find several versions of Hadoop available.

The versioning strategy used is major.minor.revision. Increments to the major version number represent large differences in operation or interface and possibly significant incompatible changes. At the time of this writing (September 2008), there have been no major upgrades; all Hadoop versions have their major version set to 0. The minor version represents a large set of feature improvements and enhancements. Hadoop instances with different minor versions may use different versions of the HDFS file formats and protocols, requiring a DFS upgrade to migrate from one to the next. Revisions are used to provide bug fixes. Within a minor version, the most recent revision contains the most stable patches.

Within the releases page, two or three versions of Hadoop will be readily available, corresponding to the highest revision number in the most recent two or three minor version increments. The stable version is the highest revision number in the second most recent minor version. Production clusters should use this version. The most recent minor version may include improved performance or new features, but may also introduce regressions that will be fixed in ensuing revisions.

At the time of this writing, 0.18.0 is the most recent version, with 0.17.2 being the “stable” release. These example instructions assume that version 0.18.0 is being used; the directions will not change significantly for any other version, except by substituting the new version number where appropriate.

To install Hadoop, first download and install prerequisite software. This includes Java 6 or higher. Distributed operation requires ssh and sshd. Windows users must install and configure cygwin as well. Then download a Hadoop version using a web browser, wget, or curl, and then unzip the package:

gunzip hadoop-0.18.0.tar.gz
tar vxf hadoop-0.18.0.tar

Within the hadoop-0.18.0/ directory which results, there will be several subdirectories. The most interesting of these are bin/, where scripts to run the cluster are located, and conf/ where the cluster’s configuration is stored.

Enter the conf/ directory and modify hadoop-env.sh. The JAVA_HOME variable must be set to the base directory of your Java installation. It is recommended that you install Java in the same location on all machines in the cluster, so this file can be replicated to each machine without modification.

The hadoop-site.xml file must also be modified to contain a number of configuration settings. The sections below address the settings which should be included here.

If you are interested in setting up a development installation, running Hadoop on a single machine, the Hadoop documentation includes getting started instructions which will configure Hadoop for standalone or “pseudo-distributed” operation.

Standalone installations run all of Hadoop and your application inside a single Java process. The distributed file system is not used; file are read from and written to the local file system. Such a setup can be helpful for debugging Hadoop applications.

Pseudo-distributed operation refers to the use of several separate processes representing the different daemons (NameNode, DataNode, JobTracker, TaskTracker) and a separate task process to perform a Hadoop job, but with all processes running on a single machine. A pseudo-distributed instance will have a functioning NameNode/DataNode managing a “DFS” of sorts. Files in HDFS are in a separate namespace from the local file system, and are stored as block objects in a Hadoop-managed directory. However, it is not truly distributed, as no processing or data storage is performed on remote notes. A pseudo-distributed instance can be extended into a fully distributed cluster by adding more machines to function as Task/DataNodes, but more configuration settings are usually required to deploy a Hadoop cluster for multiple users.

The rest of this document deals with configuring Hadoop clusters of multiple nodes, intended for use by one or more developers.

After the conf/hadoop-site.xml is configured according to one of the models in the getting started, the sections below, or your own settings, two more files must be written.

The conf/masters file contains the hostname of the SecondaryNameNode. This should be changed from “localhost” to the fully-qualified domain name of the node to run the SecondaryNameNode service. It does not need to contain the hostname of the JobTracker/NameNode machine; that service is instantiated on whichever node is used to run bin/start-all.sh, regardless of the masters file. The conf/slaves file should contain the hostname of every machine in the cluster which should start TaskTracker and DataNode daemons. One hostname should be written per line in each of these files, e.g.:


The master node does not usually also function as a slave node, except in installations across only 1 or 2 machines.

If the nodes on your cluster do not support passwordless ssh, you should configure this now:

$ ssh-keygen -t dsa -P '' -f ~/.ssh/id_dsa
$ cat ~/.ssh/id_dsa.pub >> ~/.ssh/authorized_keys

This will enable passwordless ssh login to the local machine. (You can verify that this works by executing ssh localhost.) The ~/.ssh/id_dsa.pub and authorized_keys files should be replicated on all machines in the cluster.

At this point, the configuration must be replicated across all nodes in the cluster. Small clusters may use rsync or copy the configuration directory to each node. Larger clusters should use a configuration management system such as bcfg2, smartfrog, or puppet. NFS should be avoided as much as is possible, as it is a scalability bottleneck. DataNodes should never share block storage or other high-bandwidth responsibilities over NFS, and should avoid sharing configuration information over NFS if possible.

Various directories should be created on each node. The NameNode requires the NameNode metadata directory:

$ mkdir -p /home/hadoop/dfs/name

And every node needs the Hadoop tmp directory and DataNode directory created. Rather than logging in to each node and performing the steps multiple times manually, the file bin/slaves.sh allows a command to be executed on all nodes in the slaves file. For example, we can create these directories by executing the following commands on the NameNode:

$ mkdir -p /tmp/hadoop  # make the NameNode's tmp dir
$ ${HADOOP_HOME}/bin/slaves.sh "mkdir -p /tmp/hadoop"
$ ${HADOOP_HOME}/bin/slaves.sh "mkdir -p /home/hadoop/dfs/data"

The environment variables $HADOOP_CONF_DIR and $HADOOP_SLAVES are used by the bin/slaves.sh script to find the slave machines list. The provided command is then executed over ssh. If you need particular ssh options, the contents of the $HADOOP_SSH_OPTS variable are passed to ssh as arguments.

We then format HDFS by executing the following command on the NameNode:

$ bin/hadoop namenode -format

And finally, start the cluster:

$ bin/start-all.sh

Now it is time to load in data and start processing it with Hadoop! Good luck!

The remainder of this document discusses various trade-offs in cluster configurations for different sizes, and reviews the settings which may be placed in the hadoop-site.xml file.

Important Directories

One of the basic tasks involved in setting up a Hadoop cluster is determining where the several various Hadoop-related directories will be located. Where they go is up to you; in some cases, the default locations are inadvisable and should be changed. This section identifies these directories.

Directory Description Default location Suggested location
HADOOP_LOG_DIR Output location for log files from daemons ${HADOOP_HOME}/logs /var/log/hadoop
hadoop.tmp.dir A base for other temporary directories /tmp/hadoop-${user.name} /tmp/hadoop
dfs.name.dir Where the NameNode metadata should be stored ${hadoop.tmp.dir}/dfs/name /home/hadoop/dfs/name
dfs.data.dir Where DataNodes store their blocks ${hadoop.tmp.dir}/dfs/data /home/hadoop/dfs/data
mapred.system.dir The in-HDFS path to shared MapReduce system files ${hadoop.tmp.dir}/mapred/system /hadoop/mapred/system


This table is not exhaustive; several other directories are listed in conf/hadoop-defaults.xml. The remaining directories, however, are initialized by default to reside under hadoop.tmp.dir, and are unlikely to be a concern.

It is critically important in a real cluster that dfs.name.dir and dfs.data.dir be moved out from hadoop.tmp.dir. A real cluster should never consider these directories temporary, as they are where all persistent HDFS data resides. Production clusters should have two paths listed for dfs.name.dir which are on two different physical file systems, to ensure that cluster metadata is preserved in the event of hardware failure.

A multi-user configuration should also definitely adjust mapred.system.dir. Hadoop’s default installation is designed to work for standalone operation, which does not use HDFS. Thus it conflates HDFS and local file system paths. When enabling HDFS, however, MapReduce will store shared information about jobs in mapred.system.dir on the DFS. If this path includes the current username (as the default hadoop.tmp.dir does), this will prevent proper operation. The current username on the submitting node will be the username who actually submits the job, e.g., “alex.” All other nodes will have the current username set to the username used to launch Hadoop itself (e.g., “hadoop”). If these do not match, the TaskTrackers will be unable to find the job information and run the MapReduce job.

For this reason, it is also advisable to remove ${user.name} from the general hadoop.tmp.dir.

While most of the directories listed above (all the ones with names in “foo.bar.baz” form) can be relocated via the conf/hadoop-site.xml file, the HADOOP_LOG_DIR directory is specified in conf/hadoop-env.sh as an environment variable. Relocating this directory requires editing this script.

Selecting Machines

Before diving into the details of configuring nodes, we include a brief word on choosing hardware for a cluster. While the processing demands of different organizations will dictate a different machine configuration for optimum efficiency, there are are commonalities associated with most Hadoop-based tasks.

Hadoop is designed to take advantage of whatever hardware is available. Modest “beige box” PCs can be used to run small Hadoop setups for experimentation and debugging. Providing greater computational resources will, to a point, result in increased performance by your Hadoop cluster. Many existing Hadoop deployments include Xeon processors in the 1.8-2.0GHz range. Hadoop jobs written in Java can consume between 1 and 2 GB of RAM per core. If you use HadoopStreaming to write your jobs in a scripting language such as Python, more memory may be advisable. Due to the I/O-bound nature of Hadoop, adding higher-clocked CPUs may not be the most efficient use of resources, unless the intent is to run HadoopStreaming. Big data clusters, of course, can use as many large and fast hard drives as are available. However, too many disks in a single machine will result in many disks not being used in parallel. It is better to have three machines with 4 hard disks each than one machine with 12 drives. The former configuration will be able to write to more drives in parallel and will provide greater throughput. Finally, gigabit Ethernet connections between machines will greatly improve performance over a cluster connected via a slower network interface.

It should be noted that the lower limit on minimum requirements for running Hadoop is well below the specifications for modern desktop or server class machines. However, multiple pages on the Hadoop wiki suggest similar specifications to those posted here for high-performance cluster design. (See [1][2].)

Cluster Configurations

This section provides cluster configuration advice and specific settings for clusters of varying sizes. These sizes were picked to demonstrate basic categories of clusters; your own installation may be a hybrid of different aspects of these profiles. Here we suggest various properties which should be included in the conf/hadoop-site.xml file to most effectively use a cluster of a given size, as well as other system configuration elements. The next section describes how to finish the installation after implementing the configurations described here. You should read through each of these configurations in order, as configuration suggestions for larger deployments are based on the preceding ones.

Small Clusters: 2-10 Nodes

Setting up a small cluster for development purposes is a very straightforward task. When using two nodes, one node will act as both NameNode/JobTracker and a DataNode/TaskTracker; the other node is only a DataNode/TaskTracker. Clusters of three or more machines typically use a dedicated NameNode/JobTracker, and all other nodes are workers.

A relatively minimalist configuration in conf/hadoop-site.xml will suffice for this installation:


Clusters closer to the 8-10 node range may want to set dfs.replication to 3. Values higher than 3 are usually not necessary. Individual files which are heavily utilized by a large number of nodes may have their particular replication factor manually adjusted upward independent of the cluster default.

Medium Clusters: 10-40 Nodes

This category is for clusters that occupy the majority of a single rack. Additional considerations for high availability and reliability come into play at this level.

The single point of failure in a Hadoop cluster is the NameNode. While the loss of any other machine (intermittently or permanently) does not result in data loss, NameNode loss results in cluster unavailability. The permanent loss of NameNode data would render the cluster’s HDFS inoperable.

Therefore, another step should be taken in this configuration to back up the NameNode metadata. One machine in the cluster should be designated as the NameNode’s backup. This machine does not run the normal Hadoop daemons (i.e., the DataNode and TaskTracker). Instead, it exposes a directory via NFS which is only mounted on the NameNode (e.g., /mnt/namenode-backup/). The cluster’s hadoop-site.xml file should then instruct the NameNode to write to this directory as well:


The NameNode will write its metadata to each directory in the comma-separated list of dfs.name.dir. If /mnt/namenode-backup is NFS-mounted from the backup machine, this will ensure that a redundant copy of HDFS metadata is available. The backup node should serve /mnt/namenode-backup from /home/hadoop/dfs/name on its own drive. This way, if the NameNode hardware completely dies, the backup machine can be brought up as the NameNode with no reconfiguration of the backup machine’s software. To switch the NameNode and backup nodes, the backup machine should have its IP address changed to the original NameNode’s IP address, and the server daemons should be started on that machine. The IP address must be changed to allow the DataNodes to recognize it as the “original” NameNode for HDFS. (Individual DataNodes will cache the DNS entry associated with the NameNode, so just changing the hostname is insufficient; the name reassignment must be performed at the IP address level.)

The backup machine still has Hadoop installed and configured on it in the same way as every other node in the cluster, but it is not listed in the slaves file, so normal daemons are not started there.

One function that the backup machine can be used for is to serve as the SecondaryNameNode. Note that this is not a failover NameNode process. The SecondaryNameNode process connects to the NameNode and takes periodic snapshots of its metadata (though not in real time). The NameNode metadata consists of a snapshot of the file system called the fsimage and a series of deltas to this snapshot called the editlog. With these two files, the current state of the system can be determined exactly. The SecondaryNameNode merges the fsimage and editlog into a new fsimage file that is a more compact representation of the file system state. Because this process can be memory intensive, running it on the backup machine (instead of on the NameNode itself) can be advantageous.

To configure the SecondaryNameNode daemon to run on the backup machine instead of on the master machine, edit the conf/masters file so that it contains the name of the backup machine. The bin/start-dfs.sh and bin/start-mapred.sh (and by extension, bin/start-all.sh) scripts will actually always start the master daemons (NameNode and JobTracker) on the local machine. The slaves file is used for starting DataNodes and TaskTrackers. The masters file is used for starting the SecondaryNameNode. This filename is used despite the fact that the master node may not be listed in the file itself.

A cluster of this size may also require nodes to be periodically decommissioned. As noted in Module 2, several machines cannot be turned off simultaneously, or data loss may occur. Nodes must be decommissioned on a schedule that permits replication of blocks being decommissioned. To prepare for this eventuality in advance, an excludes file should be added to the conf/hadoop-site.xml:


This property should provide the full path to the excludes file (the actual location of the file is up to you). You should then create an empty file with this name:

$ touch /home/hadoop/excludes

While the dfs.hosts.exclude property allows the definition of a list of machines which are explicitly barred from connecting to the NameNode (and similarly, mapred.hosts.exclude for the JobTracker), a large cluster may want to explicitly manage a list of machines which are approved to connect to a given JobTracker or NameNode.

The dfs.hosts and mapred.hosts properties allow an administrator to supply a file containing an approved list of hostnames. If a machine is not in this list, it will be denied access to the cluster. This can be used to enforce policies regarding which teams of developers have access to which MapReduce sub-clusters. These are configured in exactly the same way as the excludes file.

Of course, at this scale and above, 3 replicas of each block are advisable; the hadoop-site.xml file should contain:


By default, HDFS does not preserve any free space on the DataNodes; the DataNode service will continue to accept blocks until all free space on the disk is exhausted, which may cause problems. The following setting will require each DataNode to reserve at least 1 GB of space on the drive free before it writes more blocks, which helps preserve system stability:


Another parameter to watch is the heap size associated with each task. Hadoop caps the heap of each task process at 200 MB, which is too small for most data processing tasks. This cap is set as a parameter passed to the child Java process. It is common to override this with a higher cap by specifying:


This will provide each child task with 512 MB of heap. It is not unreasonable in some cases to specify -Xmx1024m instead. In the interest of providing only what is actually required, it may be better to leave this set to 512 MB by default, and allowing applications to manually configure for a full GB of RAM/task themselves.

Using multiple drives per machine

While small clusters often have only one hard drive per machine, more high-performance configurations may include two or more disks per node. Slight configuration changes are required to make Hadoop take advantage of additional disks.

DataNodes can be configured to write blocks out to multiple disks via the dfs.data.dir property. It can take on a comma-separated list of directories. Each block is written to one of these directories. E.g., assuming that there are four disks, mounted on /d1, /d2, /d3, and /d4, the following (or something like it) should be in the configuration for each DataNode:


MapReduce performance can also be improved by distributing the temporary data generated by MapReduce tasks across multiple disks on each machine:


Finally, if there are multiple drives available in the NameNode, they can be used to provide additional redundant copies of the NameNode metadata in the event of the failure of one drive. Unlike the above two properties, where one drive out of many is selected to write a piece of data, the NameNode writes to each comma-separated path in dfs.name.dir. If too many drives are listed here it may adversely affect the performance of the NameNode, as the probability of blocking on one or more I/O operations increases with the number of devices involved, but it is imperative that the sole copy of the metadata does not reside on a single drive.

Large Clusters: Multiple Racks

Configuring multiple racks of machines for Hadoop requires further advance planning. The possibility of rack failure now exists, and operational racks should be able to continue even if entire other racks are disabled. Naive setups may result in large cross-rack data transfers which adversely affect performance. Furthermore, in a large cluster, the amount of metadata under the care of the NameNode increases. This section proposes configuring several properties to help Hadoop operate at very large scale, but the numbers used in this section are just guidelines. There is no single magic number which works for all deployments, and individual tuning will be necessary. These will, however, provide a starting point and alert you to settings which will be important.

The NameNode is responsible for managing metadata associated with each block in the HDFS. As the amount of information in the rack scales into the 10’s or 100’s of TB, this can grow to be quite sizable. The NameNode machine needs to keep the blockmap in RAM to work efficiently. Therefore, at large scale, this machine will require more RAM than other machines in the cluster. The amount of metadata can also be dropped almost in half by doubling the block size:


This changes the block size from 64MB (the default) to 128MB, which decreases pressure on the NameNode’s memory. On the other hand, this potentially decreases the amount of parallelism that can be achieved, as the number of blocks per file decreases. This means fewer hosts may have sections of a file to offer to MapReduce tasks without contending for disk access. The larger the individual files involved (or the more files involved in the average MapReduce job), the less of an issue this is.

In the medium configuration, the NameNode wrote HDFS metadata through to another machine on the rack via NFS. It also used that same machine to checkpoint the NameNode metadata and compact it in the SecondaryNameNode process. Using this same setup will result in the cluster being dependent on a single rack’s continued operation. The NFS-mounted write-through backup should be placed in a different rack from the NameNode, to ensure that the metadata for the file system survives the failure of an individual rack. For the same reason, the SecondaryNameNode should be instantiated on a separate rack as well.

With multiple racks of servers, RPC timeouts may become more frequent. The NameNode takes a continual census of DataNodes and their health via heartbeat messages sent every few seconds. A similar timeout mechanism exists on the MapReduce side with the JobTracker. With many racks of machines, they may force one another to timeout because the master node is not handling them fast enough. The following options increase the number of threads on the master machine dedicated to handling RPC’s from slave nodes:


These settings were used in clusters of several hundred nodes. They should be scaled up accordingly with larger deployments.

The following settings provide additional starting points for optimization. These are based on the reported configurations of actual clusters from 250 to 2000 nodes.

Property Range Description
io.file.buffer.size 32768-131072 Read/write buffer size used in SequenceFiles (should be in multiples of the hardware page size)
io.sort.factor 50-200 Number of streams to merge concurrently when sorting files during shuffling
io.sort.mb 50-200 Amount of memory to use while sorting data
mapred.reduce.parallel.copies 20-50 Number of concurrent connections a reducer should use when fetching its input from mappers
tasktracker.http.threads 40-50 Number of threads each TaskTracker uses to provide intermediate map output to reducers
mapred.tasktracker.map.tasks.maximum 1/2 * (cores/node) to 2 * (cores/node) Number of map tasks to deploy on each machine.
mapred.tasktracker.reduce.tasks.maximum 1/2 * (cores/node) to 2 * (cores/node) Number of reduce tasks to deploy on each machine.


Rack awareness

In a multi-rack configuration, it is important to ensure that replicas of blocks are placed on multiple racks to minimize the possibility of data loss. Thus, a rack-aware placement policy should be used. A basic rack awareness script is provided in Module 2. The guidelines there suggest how to set up a basic rack awareness policy; due to the heterogeneity of network topologies, a definitive general-purpose solution cannot be provided here.

This tutorial targets Hadoop version 0.18.0. While most of the interfaces described will work on other, older versions of Hadoop, rack-awareness underwent a major overhaul in version 0.17. Thus, the following does not apply to version 0.16 and before.

One major consequence of the upgrade is that while rack-aware block replica placement has existed in Hadoop for some time, rack-aware task placement has only been added in version 0.17. If Hadoop MapReduce cannot place a task on the same node as the block of data which the task is scheduled to process, then it picks an arbitrary different node on which to schedule the task. Starting with 0.17.0, tasks will be placed (when possible) on the same rack as at least one replica of an input data block for a job, which should further minimize the amount of inter-rack data transfers required to perform a job.

Hadoop includes an interface called DNSToSwitchMapping which allows arbitrary Java code to be used to map servers onto a rack topology. The configuration key topology.node.switch.mapping.impl can be used to specify a class which meets this interface. More straightforward than writing a Java class for this purpose, however, is to use the default mapper, which executes a user-specified script (or other command) on each node of the cluster, which returns the rack id for that node. These rack ids are then aggregated and sent back to the NameNode.

Note that the rack mapping script used by this system is incompatible with the 0.16 method of using dfs.network.script. Whereas dfs.network.script runs on each DataNode, a new script specified by topology.script.file.name is run by the master node only. To set the rack mapping script, specify the key topology.script.file.name in conf/hadoop-site.xml.

Cluster contention

If you are configuring a large number of machines, it is likely that you have a large number of users who wish to submit jobs to execute on it. Hadoop’s job scheduling algorithm is based on a simple FIFO scheduler. Using this in a large deployment without external controls or policies agreed upon by all users can lead to lots of contention for the JobTracker, causing short jobs to be delayed by other long-running tasks and frustrating users.

An advanced technique to combat this problem is to configure a single HDFS cluster which spans all available machines, and configure several separate MapReduce clusters with their own JobTrackers and pools of TaskTrackers. All MapReduce clusters are configured to use the same DFS and the same NameNode; but separate groups of machines have a different machine acting as JobTracker (i.e., subclusters have different settings for mapred.job.tracker). Breaking machines up into several smaller clusters, each of which contains 20-40 TaskTrackers, provides users with lower contention for the system. Users may be assigned to different clusters by policy, or they can use the JobTracker status web pages (a web page exposed on port 50030 of each JobTracker) to determine which is underutilized.

Multiple strategies exist for this assignment process. It is considered best practice to stripe the TaskTrackers associated with each JobTracker across all racks. This maximizes the availability of each cluster (as they are all resistant to individual rack failure), and works with the HDFS replica placement policy to ensure that each MapReduce cluster can find rack-local replicas of all files used in any MapReduce jobs.

Performance Monitoring

Multiple tools exist to monitor large clusters for performance and troubleshooting. This section briefly highlights two such tools.


Ganglia is a performance monitoring framework for distributed systems. Ganglia provides a distributed service which collects metrics on individual machines and forwards them to an aggregator which can report back to an administrator on the global state of a cluster.

Ganglia is designed to be integrated into other applications to collect statistics about their operation. Hadoop includes a performance monitoring framework which can use Ganglia as its backend. Instructions are available on the Hadoop wiki as to how to enable Ganglia metrics in Hadoop. Instructions are also included below.

After installing and configuring Ganglia on your cluster, to direct Hadoop to output its metric reports to Ganglia, create a file named hadoop-metrics.properties in the $HADOOP_HOME/conf directory. The file should have the following contents:



This assumes that gmond is running on each machine in the cluster. Instructions on the Hadoop wiki note that (in the experience of the wiki article author) this may result in all nodes reporting their results as “localhost” instead of with their individual hostnames. If this problem affects your cluster, an alternate configuration is proposed, in which all Hadoop instances speak directly with gmetad:



Where @GMETAD@ is the hostname of the server on which the gmetad service is running. If deploying Ganglia and Hadoop on a very large number of machines, the impact of this configuration (vs. the standard Ganglia configuration where individual services talk to gmond on localhost) should be evaluated.


While Ganglia will monitor Hadoop-specific metrics, general information about the health of the cluster should be monitored with an additional tool.

Nagios is a machine and service monitoring system designed for large clusters. Nagios will provide useful diagnostic information for tuning your cluster, including network, disk, and CPU utilization across machines.

Additional Tips

The following are a few additional pieces of small advice:

  • Create a separate user named “hadoop” to run your instances; this will separate the Hadoop processes from any users on the system. Do not run Hadoop as root.
  • If Hadoop is installed in /home/hadoop/hadoop-0.18.0, link /home/hadoop/hadoop to /home/hadoop/hadoop-0.18.0. When upgrading to a newer version in the future, the link can be moved to make this process easier on other scripts that depend on the hadoop/bin directory.

Creative Commons License
“Hadoop Tutorial from Yahoo!” by Yahoo! Inc. is licensed under a Creative Commons Attribution 3.0 Unported License.