Give me a Buzz: mail2boopal@gmail.com
Apache Hadoop Basics
“Big Data”
A “big” shift is occurring. Today, the enterprise collects more data than ever before, from a wide variety of sources and in a wide variety of formats. Along with traditional transactional and analytics data stores, we now collect additional data across social media activity, web server log files, financial transactions and sensor data from equipment in the field. Deriving meaning from all this data using conventional database technology and analytics tools would be impractical. A new set of technologies has enabled this shift.
Now an extremely popular term, “big data” technology seeks to transform all this raw data into meaningful and actionable insights for the enterprise. In fact, big data is about more than just the “bigness” of the data. Its key characteristics, coined by industry analysts are the “Three V’s,” which include volume (size) as well as velocity (speed) and variety (type). While these terms are effective descriptors, big data also needs to provide value and we often find that as the intersection of transactions, interactions and observations.
Each of these data source introduces its own set of complexities to data processing; combined they can easily push a given application or data set beyond the pale of traditional data processing tools, methods or systems. In these cases, a new approach is required. Big data represents just that.
At its core then, big data is the collection, management and manipulation of data using a new generation of technologies. Driven by the growing importance of data to the enterprise, big data techniques have become essential to enterprise competitiveness.
Apache Hadoop: Platform for Big Data
Apache Hadoop is a platform that offers an efficient and effective method for storing and processing massive amounts of data. Unlike traditional offerings, Hadoop was designed and built from the ground up to address the requirements and challenges of big data. Hadoop is powerful in its ability to allow businesses to stop worrying about building big-data-capable infrastructure and to focus on what really matters: extracting business value from the data.
Apache Hadoop use cases are many, and show up in many industries, including: risk, fraud and portfolio analysis in financial services; behavior analysis and personalization in retail; social network, relationship and sentiment analysis for marketing; drug interaction modeling and genome data processing in healthcare and life sciences… to name a few.
The first step in harnessing the power of Hadoop for your business is to understand the basics: what Hadoop is, where it comes from, how it can be applied to your business processes, and how to get started using it.
The Early Days of Hadoop at Yahoo!
Hadoop has its roots at Yahoo!, whose Internet search engine business required the continuous processing of large amounts of Web page data. In 2005 Eric Baldeschwieler (aka “E14” and Hortonworks CTO) challenged Owen O’Malley (Hortonworks co-founder) and several others to solve a really hard problem: store and process the data on the internet in a simple, scalable and economically feasible way. They looked at traditional storage approaches but quickly realized they just weren’t going to work for the type of data (much of it unstructured) and the sheer quantity Yahoo! would have to deal with.
The team designed and prototype a new framework for Yahoo! Search. This framework, called Dreadnaught, was implemented in the C++ programming language and modeled after research published by Google describing its Google File System and a distributed processing algorithm called MapReduce.
At the same time, an open-source search engine project called Nutch, lead by Doug Cutting and Mike Carafella, had implemented similar functions using Java. By the end of 2005, Doug and Mike had the Java version working on small cluster of 20 nodes. Impressed by their progress, Yahoo! Hired Cutting in January 2006, and merged the two teams together, choosing the Java version of MapReduce because it had some of the search functions already built out. With an expanded team including Owen O’Malley and Arun Murthy, the code base of this distributed processing system tripled by mid-year. Soon, a 200-node research cluster was up and running and the first real users were getting started. Apache Hadoop had liftoff.
The Power of Open Source
Apache Hadoop is an open source project governed by the Apache Software Foundation (ASF). As part of the ASF, development of will remain open and transparent for all future users and more importantly freely available.
Eric and team realized that with a community of like-minded individuals, Hadoop would innovate far faster. At the same time, they’d enable other organizations to realize some of the same benefits that they were starting to see from their early efforts. When organizations such as Facebook, LinkedIn, eBay, Powerset, Quantcast and others began picking up Hadoop and innovating in areas beyond the initial focus, it reinforced the fact that the choice of community driven open source was the right one.
A case in point being when a small startup (Powerset) started working on a project to support tables on HDFS inspired by Google’s BigTable paper; that effort turned into what’s now Apache HBase! Further… Facebook started an effort to build a SQL layer on top of MapReduce, which became Apache Hive!
It has proven time and again when it comes to platform technologies like Hadoop that community-driven open source will always outpace the innovation of a single group of people or single company.
Commercial Adoption
Apache Hadoop became a foundational technology at Yahoo, underlying a wide range of business-critical applications. Companies in nearly every vertical started to adopt Hadoop. By 2010, the community had grown to thousands of users and broad enterprise momentum had been established.
After many years architecting and operating the Hadoop infrastructure at Yahoo! and contributing heavily to the open source community, E14 and 20+ Hadoop architects and engineers spun out of Yahoo! to form Hortonworks in 2011. Having seen what it could do for Yahoo, Facebook, eBay, LinkedIn and others, their singular objective is to focus on making Apache Hadoop into a platform that is easy to use and consume by the broader market of enterprise customers and partners.
Understanding Hadoop and Related Services
At its core, Apache Hadoop is a framework for scalable and reliable distributed data storage and processing. It allows for the processing of large data sets across clusters of computers using a simple programming model. It is designed to scale up from single servers to thousands of machines, aggregating the local computation and storage from each server. Rather than relying on expensive hardware, the Hadoop software detects and handles any failures that occur, allowing it to achieve high availability on top of inexpensive commodity computers, each individually prone to failure.
At the core of Apache Hadoop are the Hadoop Distributed File System, or HDFS, and Hadoop MapReduce, which provides a framework for distributed processing.
HDFS
The Hadoop Distributed File System is a scalable and reliable distributed storage system that aggregates the storage of every node in a Hadoop cluster into a single global file system. HDFS stores individual files in large blocks, allowing it to efficiently store very large or numerous files across multiple machines and access individual chunks of data in parallel, without needing to read the entire file into a single computer’s memory. Reliability is achieved by replicating the data across multiple hosts, with each block of data being stored, by default, on three separate computers. If an individual node fails, the data remains available and an additional copy of any blocks it holds may be made on new machines to protect against future failures.
This approach allows HDFS to dependably store massive amounts of data. For instance, in late 2012, the Apache Hadoop clusters at Yahoo had grown to hold over 350 petabytes (PB) of data across 40,000+ servers. Once data has been loaded into HDFS, we can begin to process it with MapReduce.
MapReduce
MapReduce is the programming model that allows Hadoop to efficiently process large amounts of data. MapReduce breaks large data processing problems into multiple steps, namely a set of Maps and Reduces, that can each be worked on at the same time (in parallel) on multiple computers.
MapReduce is designed to work with of HDFS. Apache Hadoop automatically optimizes the execution of MapReduce programs so that a given Map or Reduce step is run on the HDFS node that contains locally the blocks of data required to complete the step. Explaining the MapReduce algorithm in a few words can be difficult, but we provide an example in Appendix A for the curious or technically inclined. For the rest, suffice it to say that MapReduce has proven itself in its ability to allow data processing problems that once required many hours to complete on very expensive computers to be written as programs that run in minutes on a handful of rather inexpensive machines. And, while MapReduce can require a shift in thinking on the part of developers, many problems not traditionally solved using the method are easily expressed as MapReduce programs.
Hadoop 2.0: YARN
As previously noted, Hadoop was initially adopted by many of the large web properties and was designed to meet their needs for large web-scale batch type processing. As clusters grew and adoption expanded, so did the number of ways that users wanted to interact with the data stored in Hadoop. As with any successful open-source project, the broader ecosystem of Hadoop users responded by contributing additional capabilities to the Hadoop community, with some of the most popular examples being Apache Hive for SQL-based querying, Apache Pig for scripted data processing and Apache HBase as a NoSQL database.
These additional open source projects opened the door for a much richer set of applications to be built on top of Hadoop – but they didn’t really address the design limitations inherent in Hadoop; specifically, that it was designed as a single application system with MapReduce at the core (i.e. batch-oriented data processing). Today, enterprise applications want to interact with Hadoop in a host of different ways: batch, interactive, analyzing data streams as they arrive, and more. And most importantly, they need to be able to do this all simultaneously without any single application or query consuming all of the resources of the cluster to do so. Enter YARN.
YARN is a key piece of Hadoop version 2 which is currently under development in the community. It provides a resource management framework for any processing engine, including MapReduce. It allows new processing frameworks to use the power of the distributed scale of Haodop. It transforms Hadoop from a single application to a multi application data system. It is the future of Hadoop.
The Hadoop Project Ecosystem
As Apache Hadoop has matured, a number of important tools have been built to support it. These tools may be categorized into Hadoop Data Services and Hadoop Operational Services based on the functionality they offer.
Hadoop Data Services Hadoop
Data Services are tools that allow users to more easily manipulate and process data. They include the following:
• Apache Hive Apache Hive is data warehouse infrastructure built on top of Hadoop for providing data summarization, ad-hoc query, and the analysis of large datasets. It provides a mechanism for imparting structure onto the data in Hadoop and for querying that data using a SQL-like language called HiveQL (HQL). Hive eases integration between Hadoop and various business intelligence and visualization tools.
• Apache Pig Apache Pig allows you to write complex map reduce transformations using a simple scripting language called Pig Latin. Pig Latin defines a set of transformations on a data set such as aggregate, join and sort. Pig translates the Pig Latin into Hadoop MapReduce so that it can be executed within HDFS.
• Apache HCatalog HCatalog is a table and storage management layer for Hadoop that enables users with different data processing tools – including Hive, Pig and MapReduce – to more easily read and write data. HCatalog’s table abstraction presents users with a relational view of data in HDFS and ensures that users need not worry about where or in what format their data is stored.
• Apache HBase Apache HBase is a non-relational database that runs on top of HDFS. It provides fault-tolerant storage and quick access to large quantities of sparse data. It also adds transactional capabilities to Hadoop, allowing users to conduct updates, inserts and deletes.
• Apache Sqoop Apache Sqoop is a tool designed for efficiently transferring bulk data between Apache Hadoop and structured datastores such as relational databases. Sqoop imports data from external sources either directly into HDFS or into systems like Hive and HBase. Sqoop can also be used to extract data from Hadoop and export it to external repositories such as relational databases and data warehouses.
• Apache Flume Apache Flume is a service for efficiently collecting, aggregating, and moving large amounts of log data. Flume allows log data from many different sources, such as web servers, to be easily stored in a centralized HDFS repository.
Hadoop Operational Services
Distributed computing at scale can present significant operational challenges. Several projects have emerged to aid in the operations and management of a Hadoop cluster. These include:
• Apache Ambari Apache Ambari provides an intuitive set of tools to monitor, manage and efficiently provision an Apache Hadoop cluster. Ambari simplifies the operation and hides the complexity of Hadoop, making Hadoop work as a single, cohesive data platform.
• Apache Oozie Apache Oozie is a Java web application used to schedule Apache Hadoop jobs. Oozie combines multiple jobs sequentially into one logical unit of work. It supports Hadoop jobs for MapReduce, Pig, Hive, and Sqoop.
• Apache ZooKeeper Apache Zookeeper provides operational services for a Hadoop cluster, including a distributed configuration service, a synchronization service and a naming registry. These services allow Hadoop’s distributed processes to more easily coordinate with one another.
Hadoop Distributions
Apache Hadoop is open source. Any interested party can visit the Hadoop web site to download the project’s source code or binaries directly from the ASF. Downloading directly, however, presents a number of challenges to enterprises that want to get started quickly. Modern Hadoop deployments require not only MapReduce and HDFS, but many—often all—of the supporting projects already discussed, along with a variety of other open source software. Assuming the enterprise even knows which software is required, this diversity adds complexity to the process of obtaining, installing and maintaining an Apache Hadoop deployment. Each of the aforementioned projects is developed independently, with its own release cycle and versioning scheme. Not all versions of all projects are API compatible, so care must be taken to choose versions of each project that are known to work together. Achieving production-level stability for a given set of projects presents even more complexity. To ensure stability, an enterprise must attempt to determine which versions of which projects have been tested together, by whom, and under what conditions—information that is not often readily available.
How then can an enterprise ensure that its first steps with Hadoop go smoothly, that the software they download is internally compatible and completely stable, and that support will be available when needed? The answer to these questions is a Hadoop Distribution, which provides an integrated, pre-packaged bundle of software that includes all required Hadoop components, related projects, and supporting software. Enterprise-focused Hadoop distributions integrate pre-certified and pre-tested versions of Apache Hadoop and related projects and make them available in a single easily installed and managed download.
The Future of Hadoop
Hadoop has “crossed the chasm” from a framework for early adopters, developers and technology enthusiasts to a strategic data platform embraced by innovative CTOs and CIOs across mainstream enterprises. These people, who want to improve the performance of their companies and unlock new business opportunities, realize that including Apache Hadoop as a deeply integrated supplement to their current data architecture offers the fastest path to reaching their goals while maximizing their existing investments.
Appendix A – MapReduce Example
As you might guess from the name MapReduce, the first of these steps is the Map; the next is the Reduce. To understand how Map and Reduce work, consider the following example, based on an extended explanation by Steve Krenzel1.
Let’s say a social media site wants to calculate every member's common friends once a day and store those results. As is typical for inputs to and outputs from MapReduce tasks, friends are stored as a key-value pair expressed in the form Member->[List of Friends]:
A -> B C D
B -> A C D E
C -> A B D E …
The first step in calculating the members’ common friends using MapReduce is the Map, in which each of the friend lists above gets mapped to a new key-value pair. In this case, the mappers sort and group the input so that a new key is output for every combination of two members, whose value is simply a list of the first member’s friends followed by a list of the second member’s friends. In our example, (A B) forms a new key whose value is a list of A’s friends (BCD) and B’s friends (ACDE):
(A B) -> (B C D) (A C D E)
…
All these intermediate key-value pairs are then sent to the Reduce step, which simply outputs for each key any friends in both members’ friend lists; in this case resulting in:
(A B) : (C D)
…
which says that members A and B have C and D as common friends. Without MapReduce, this process would require a slow, serial process of comparing A’s friends with B’s friends, A friends with C’s friends, A’s friends with D’s friends, etc., for every one of A’s friends. And so on, for every member of the social media site. Easy when you’re just starting out and have five members, but the traditional approach quickly becomes untenable when membership gets into the millions.
With MapReduce and HDFS, the map and reduce tasks are distributed across the Hadoop cluster and many tasks run simultaneously across the cluster. Because of the data locality property of Hadoop, the map tasks each run on a node that contains the blocks of input data (lines in the Friends List file) that they will operate on.
0 comments: