воскресенье, 21 сентября 2014 г.

What is BigData?

I found amazing definition of BigData in the book: BigData Analytics. Here is:

Big data is, admittedly, an overhyped buzzword used by software and hardware companies to boost their sales. Behind the hype, however, there is a real and extremely important technology trend with impressive business potential. Although big data is often associated with social media, we will show that it is about much more than that. Before we venture into definitions, however, let’s have a look at some facts about big data.
Back in 2001, Doug Laney from Meta Group (an IT research company acquired by Gartner in 2005) wrote
a research paper in which he stated that e-commerce had exploded data management along three dimensions: volumes, velocity, and variety. These are called the three Vs of big data and, as you would expect, a number of vendors have added more Vs to their own definitions.

Volume is the first thought that comes with big data: the big part. Some experts consider Petabytes the starting point of big data. As we generate more and more data, we are sure this starting point will keep growing. However, volume in itself is not a perfect criterion of big data, as we feel that the other two Vs have a more direct impact.
Velocity refers to the speed at which the data is being generated or the frequency with which it is delivered. Think of the stream of data coming from the sensors in the highways in the Los Angeles area, or the video cameras in some airports that scan and process faces in a crowd. There is also the click stream data of popular e-commerce web sites.
Variety is about all the different data and file types that are available. Just think about the music files in the iTunes store (about 28 million songs and over 30 billion downloads), or the movies in Netflix (over 75,000), the articles in
the New York Times web site (more than 13 million starting in 1851), tweets (over 500 million every day), foursquare check-ins with geolocation data (over five million every day), and then you have all the different log files produced by any system that has a computer embedded. When you combine these three Vs, you will start to get a more complete picture of what big data is all about.

Another characteristic usually associated with big data is that the data is unstructured. We are of the opinion that there is no such thing as unstructured data. We think the confusion stems from a common belief that if data cannot conform to a predefined format, model, or schema, then it is considered unstructured.
An e-mail message is typically used as an example of unstructured data; whereas the body of the e-mail could be considered unstructured, it is part of a well-defined structure that follows the specifications of RFC-2822, and contains a set of fields that include From, To, Subject, and Date. This is the same for Twitter messages, in which the body of the message, or tweet, can be considered unstructured as well as part of a well-defined structure.
In general, free text can be considered unstructured, because, as we mentioned earlier, it does not necessarily conform to a predefined model. Depending on what is to be done with the text, there are many techniques to process it, most of which do not require predefined formats.
Relational databases impose the need for predefined data models with clearly defined fields that live in tables, which can have relations between them. We call this Early Structure Binding, in which you have to know in advance what questions are to be asked of the data, so that you can design the schema or structure and then work with the data to answer them.
As big data tends to be associated with social media feeds that are seen as text-heavy, it is easy to understand why people associate the term unstructured with big data. From our perspective, multistructured is probably a more accurate description, as big data can contain a variety of formats (the third V of the three Vs).
It would be unfair to insist that big data is limited to so-called unstructured data. Structured data can also be considered big data, especially the data that languishes in secondary storage hoping to make it some day to the data warehouse to be analyzed and expose all the golden nuggets it contains. The main reason this kind of data is usually ignored is because of its sheer volume, which typically exceeds the capacity of data warehouses based on relational databases.
At this point, we can introduce the definition that Gartner, an Information Technology (IT) consultancy, proposed in 2012: “Big data are high volume, high velocity, and/or high variety information assets that require new forms of processing to enable enhanced decision making, insight discovery and processes optimization.” We like this definition, because it focuses not only on the actual data but also on the way that big data is processed. Later in this book, we will get into more detail on this.
We also like to categorize big data, as we feel that this enhances understanding. From our perspective, big
data can be broken down into two broad categories: human-generated digital footprints and machine data. As our interactions on the Internet keep growing, our digital footprint keeps increasing. Even though we interact on a daily basis with digital systems, most people do not realize how much information even trivial clicks or interactions leave behind. We must confess that before we started to read Internet statistics, the only large numbers we were familiar with were the McDonald’s slogan “Billions and Billions Served” and the occasional exposure to U.S. politicians talking about budgets or deficits in the order of trillions. Just to give you an idea, we present a few Internet statistics that show the size of our digital footprint. We are well aware that they are obsolete as we write them, but here they are anyway:
  • By February 2013, Facebook had more than one billion users, of which 618 million were active on a daily basis. They shared 2.5 billion items and “liked” other 2.7 billion every day, generating more than 500 terabytes of new data on a daily basis.
  • In March 2013, LinkedIn, which is a business-oriented social networking site, had more than 200 million members, growing at the rate of two new members every second, which generated 5.7 billion professionally oriented searches in 2012.
  • Photos are a hot subject, as most people have a mobile phone that includes a camera. The numbers are mind-boggling. Instagram users upload 40 million photos a day, like 8,500 of them every second, and create about 1,000 comments per second. On Facebook, photos are uploaded at the rate of 300 million per day, which is about seven petabytes worth of data a month. By January 2013, Facebook was storing 240 billion photos.
  • Twitter has 500 million users, growing at the rate of 150,000 every day, with over 200 million of the users being active. In October 2012, they had 500 million tweets a day.
  • Foursquare celebrated three billion check-ins in January 2013, with about five million check-ins a day from over 25 million users that had created 30 million tips.
  • On the blog front, WordPress, a popular blogging platform reported in March 2013 almost
    40 million new posts and 42 million comments per month, with more than 388 million people viewing more than 3.6 billion pages per month. Tumblr, another popular blogging platform, also reported, in March 2013, a total of almost 100 million blogs that contain more than
    44 billion posts. A typical day at Tumblr at the time had 74 million blog posts.

  • Pandora, a personalized Internet radio, reported that in 2012 their users listened to 13 billion hours of music, that is, about 13,700 years worth of music.
  • In similar fashion, Netflix announced their users had viewed one billion hours of videos in July 2012, which translated to about 30 percent of the Internet traffic in the United States. As if that is not enough, in March 2013, YouTube reported more than four billion hours watched per month and 72 hours of video uploaded every minute.
  • In March 2013, there were almost 145 million Internet domains, of which about 108 million used the famous “.com” top level domain. This is a very active space; on March 21, there were 167,698 new and 128,866 deleted domains, for a net growth of 38,832 new domains.
  • In the more mundane e-mail world, Bob Al-Greene at Mashable reported in November 2012 that there are over 144 billion e-mail messages sent every day, with about 61 percent of them from businesses. The lead e-mail provider is Gmail, with 425 million active users.
Reviewing these statistics, there is no doubt that the human-generated digital footprint is huge. You can quickly identify the three Vs; to give you an idea of how big data can have an impact on the economy, we share the announcement Yelp, a user-based review site, made in January 2013, when they had 100 million unique visitors and over one million reviews: “A survey of business owners on Yelp reported that, on average, customers across all categories surveyed spend $101.59 in their first visit. That’s everything from hiring a roofer to buying a new mattress and even your morning cup of joe. If each of those 100 million unique visitors spent $100 at a local business in January, Yelp would have influenced over $10 billion in local commerce.”
We will not bore you by sharing statistics based on every minute or every second of the day in the life of the Internet. However, a couple of examples of big data in action that you might relate with can consolidate the notion; the recommendations you get when you are visiting the Amazon web site or considering a movie in Netflix, are based on big data analytics the same way that Walmart uses it to identify customer preferences on a regional basis and stock their stores accordingly. By now you must have a pretty good idea of the amount of data our digital footprint creates and the impact that it has in the economy and society in general. Social media is just one component of big data.
The second category of big data is machine data. There is a very large number of firewalls, load balancers, routers, switches, and computers that support our digital footprint. All of these systems generate log files, ranging from security and audit log files to web site log files that describe what a visitor has done, including the infamous abandoned shopping carts.
It is almost impossible to find out how many servers are needed to support our digital footprint, as all companies are extremely secretive on the subject. Many experts have tried to calculate this number for the most visible companies, such as Google, Facebook, and Amazon, based on power usage, which (according to a Power Usage Effectiveness indicator that some of these companies are willing to share) can provide some insight as to the number of servers they have in their data centers. Based on this, James Hamilton in a blog post of August 2012 published server estimates conjecturing that Facebook had 180,900 servers and Google had over one million servers. Other experts state that Amazon had about 500 million servers in March 2012. In September 2012, the New York Times ran a provocative article that claimed that there are tens of thousands of data centers in the United States, which consume roughly 2 percent of all electricity used in the country, of which 90 percent or more goes to waste, as the servers are not really being used.
We can only guess that the number of active servers around the world is in the millions. When you add to this all the other typical data center infrastructure components, such as firewalls, load balancers, routers, switches, and many others, which also generate log files, you can see that there is a lot of machine data generated in the form of log files by the infrastructure that supports our digital footprint.
What is interesting is that not long ago most of these log files that contain machine data were largely ignored. These log files are a gold mine of useful data, as they contain important insights for IT and the business because they are a definitive record of customer activity and behavior as well as product and service usage. This gives companies end-to-end transaction visibility, which can be used to improve customer service and ensure system security, and also helps to meet compliance mandates. What’s more, the log files help you find problems that have occurred and can assist you in predicting when similar problems can happen in the future. 
In addition to the machine data that we have described so far, there are also sensors that capture data on a real-time basis. Most industrial equipment has built-in sensors that produce a large amount of data. For example, a blade in a gas turbine used to generate electricity creates 520 Gigabytes a day, and there are 20 blades in one
of these turbines. An airplane on a transatlantic flight produces several Terabytes of data, which can be used to streamline maintenance operations, improve safety, and (most important to an airline’s bottom line) decrease fuel consumption.

Another interesting example comes from the Nissan Leaf, an all-electric car. It has a system called CARWINGS, which not only offers the traditional telematics service and a smartphone app to control all aspects of the car but wirelessly transmits vehicle statistics to a central server. Each Leaf owner can track their driving efficiency and compare their energy economy with that of other Leaf drivers. We don’t know the details of the information that Nissan is collecting from the Leaf models and what they do with it, but we can definitely see the three Vs in action in this example.
In general, sensor-based data falls into the industrial big data category, although lately the “Internet of Things” has become a more popular term to describe a hyperconnected world of things with sensors, where there are over 300 million connected devices that range from electrical meters to vending machines. We will not be covering
this category of big data in this book, but the methodology and techniques described here can easily be applied to industrial big data analytics.

Alternate Data Processing Techniques
Big data is not only about the data, it is also about alternative data processing techniques that can better handle the three Vs as they increase their values. The traditional relational database is well known for the following characteristics:
  • Transactional support for the ACID properties:
    • Atomicity: Where all changes are done as if they are a single operation.
    • Consistency: At the end of any transaction, the system is in a valid state.
    • Isolation: The actions to create the results appear to have been done sequentially, one at a time.
    • Durability: All the changes made to the system are permanent.
  • The response times are usually in the subsecond range, while handling thousands of
    interactive users.
  • The data size is in the order of Terabytes.
  • Typically uses the SQL-92 standard as the main programming language.
In general, relational databases cannot handle the three Vs well. Because of this, many different approaches have been created to tackle the inherent problems that the three Vs present. These approaches sacrifice one or more of the ACID properties, and sometimes all of them, in exchange for ways to handle scalability for big volumes, velocity, or variety. Some of these alternate approaches will also forgo fast response times or the ability to handle a high number of simultaneous users in favor of addressing one or more of the three Vs.
Some people group these alternate data processing approaches under the name NoSQL and categorize them according to the way they store the data, such as key-value stores and document stores, where the definition of a document varies according to the product. Depending on who you talk to, there may be more categories. 

The open source Hadoop software framework is probably the one that has the biggest name recognition in the big data world, but it is by no means alone. As a framework it includes a number of components designed to solve the issues associated with distributed data storage, retrieval and analysis of big data. It does this by offering two basic functionalities designed to work on a cluster of commodity servers:
  • A distributed file system called HDFS that not only stores data but also replicates it so that it is always available.
  • A distributed processing system for parallelizable problems called MapReduce, which is a two-step approach. In the first step or Map, a problem is broken down into many small ones and sent to servers for processing. In the second step or Reduce, the results of the Map step are combined to create the final results of the original problem.
Some of the other components of Hadoop, generally referred to as the Hadoop ecosystem, include Hive, which
is a higher level of abstraction of the basic functionalities offered by Hadoop. Hive is a data warehouse system in which the user can specify instructions using the SQL-92 standard and these get converted to MapReduce tasks. Pig is another high-level abstraction of Hadoop that has a similar functionality to Hive, but it uses a programming language called Pig Latin, which is more oriented to data flows.
HBase is another component of the Hadoop ecosystem, which implements Google’s Bigtable data store. Bigtable is a distributed, persistent multidimensional sorted map. Elements in the map are an uninterpreted array of bytes, which are indexed by a row key, a column key, and a timestamp.
There are other components in the Hadoop ecosystem, but we will not delve into them. We must tell you that in addition to the official Apache project, Hadoop solutions are offered by companies such as Cloudera and Hortonworks, which offer open source implementations with commercial add-ons mainly focused on cluster management. MapR is a company that offers a commercial implementation of Hadoop, for which it claims higher performance.
Other popular products in the big data world include:
  • Cassandra, an Apache open source project, is a key-value store that offers linear scalability and fault tolerance on commodity hardware.
  • DynamoDB, an Amazon Web Services offering, is very similar to Cassandra.
  • MongoDB, an open source project, is a document database that provides high performance,
    fault tolerance, and easy scalability.
  • CouchDB, another open source document database that is distributed and fault tolerant.
    In addition to these products, there are many companies offering their own solutions that deal in different ways with the three Vs.