10 Things you must know about Charts, Graphs & Visualization

"In God we trust, all others must bring data" said W. Edwards Deming. But we all know it is easier said than understood when data is actually presented in front of you. Often you need a statistician to interpret it. And they are rare! What you need is a straight and simple way to represent data that brain can interpret relatively easily. Here comes the need for Charts, Graphs and other visualization.

I started exploring the various Charts & Graphs libraries for generating insight for my hobby project www.fanffair.com. Goal was to identify the top posts based on statistical analysis of Like, Share and Comments data associated each post and share the top posts on fanffair Facebook Page. That in turn led to second hobby project of mine www.statspanda.com. In this blog post, I am going to share my notes, learning & observations as I sifted through various charting libraries ranging from Google Charts, morris.js, raphael.js, xCharts, nvd3, Flot charts, dygraphs, Rickshaw, Highcharts et al to D3.js.

All of the above are the charting libraries that either you download or use as hosted-library and invoke the JavaScript methods to create a chart or graph in your web application, albeit some of them are pretty straight and simple to use while libraries like D3.js requires some amount of programming skills and knowledge of HTML, Javascript etc. But unlike Charting libraries, D3.js gives all the power and flexibility for creating exotic visualizations. You can as well pick some of the visualization created and curated by Mike Bostok.

If you want to avoid either of the options, then you can go for online services that take your data and generate the charts and graphs online. There are many such services, some of them are really big. Many of them are often categorized as data analytics companies as they deal a lot in the data layer. I have provided a list of such services/companies in later part of this post. In fact, statspanda.com will qualify for this category.

There is a fourth category where companies sell ready-to-use Dashborad softwares. They have pre-built charts and graphs but not limited to charting and graphs. One needs to customize, implement and host them on their own.

To start with, I will cover various chart types and some basics around their usage. I will also briefly touch upon the technologies used behind these charting libraries. Then I will move on to talk very briefly about some of the libraries I have explored closely. This will be followed by a list of companies who sell Charts and Graphs as service or as Dashboard software.  

Popular Chart Types (#1)

Line Chart
Spline Chart
Bar Chart
Pie Chart
Doughnut Chart
Area Chart
Bubble Chart
Scatter Plot
Bullet Chart
Gauge Chart
Combination Chart
Creative Visualization

Sample Charts generated using StatsPanda.com



Factors Behind Making a Choice (#2)

There are host of factors that may influence your choice of a particular type of chart or visualization. Here are few factors that I could compile 

Size of Data points - this is probably the most important factor behind making a choice. For example, a Pie Chart may not work very well if your data point is more than say 30 but Line Chart may still be good choice. Similarly, an Area Chart or Line chart may represent a very large set of data. However, remember number of data points also indicates whether it's a low level or high level data. Millions of raw data feeds may translate into couple of data points when aggregated by those factors.

Types of Data Points - Next, check if you are dealing with only one type of data or there multiple different types of data that need to be plotted. Often combination charts are used if data types are different or different set of data need to represented. Example - using a combination of Bar Chart & Line chart for representing revenue and profit numbers of one company viz-a-viz using multiple Line Charts for representing revenue numbers of different companies.

Relationship of Data Points - If multiple Data Types are being presented in a Combination Chart, then if the different types data are related or not also influence the choice of a Chart type.

Complexity Of Relationship - How the data sets are related matters in the selection of a chart. Complex relationships like many-to-many relations or multiple-relations between two data points are not easy represent in regular Charts & Graphs and calls for creative visualization.

Static vs Animation vs Flow Representation - Most of the Charts and Visualizations really capture a snapshot or static state of data. In certain cases you may need to show a transition from one snapshot of data to the other (i.e. animated) or capture the states that the data pass through (i.e. data flow). A good example is Sankey Chart.   

A general rule of thumb as per a paper from IBM

• Pie chart: 3-10
  
• Bar chart: fewer than 50
  
• Line chart: fewer than 500
  
• Bubble plot: fewer than 500
  
• Scatter plot: fewer than 10,000
• Creative Visualization comes handy for data points > 10,000  

Rendering Engines (#3)

SVG - Scalable Vector Graphics (SVG) is an XML-based vector image format for two-dimensional graphics with support for interactivity and animation. The SVG specification is an open standard developed by the World Wide Web Consortium (W3C) since 1999. Widely supported. Traditionally, MS-Internet Explorer has been a laggard but starting with IE 8 SVG is supported in IE. Some of the popular libraries like morris.js, D3.js and others built on top of D3.js use SVG.

Canvas (HTML5) - Introduced as part of HTML 5 spec. example: flot, chartjs, jqPlot and other Javascript based visualisation libraries.

VML - Vector Markup Language (VML) is an XML-based file format for two-dimensional vector graphics. Developed and promoted by Microsoft.

Chart Performance (#4)

Check out Charts performance at http://jsperf.com/charts-comparison-d3-js-kendo-highcharts-echart-flot-gr 


HTML5 Based Libraries (#5)

chartjs.org - Offers six chart types. HTML5 Canvas based, responsive. One of the best options for the chart types it supports.

canvasjs.com - Pretty good & rich collection of Charts. Free for non-commercial use.


D3.js & D3.js Based Libraries (#6)

d3.js is the best and gives highest amount of flexibility but that comes at the cost of added complexity in terms of programming skills while developing a chart or graph. If you want to avoid that, please refer to some of the libraries and that pretty straight to use.

NVD3 d3.js based charts library, has good options. Some of them are unique when compared to other generally available charts e.g. Scatter / Bubble Chart, Stacked Area Chart. Reads data from CSV and other text formats.

C3.js - Another D3.js based charting library. Pretty comprehensive set of charts, simple and easy to use, provides good documentation and available under The MIT License (MIT). C3.js has a pretty active Google Group, easy to get support.

Rickshaw built on top of d3.js. Very neat library for creating time series graphs, line chart, bar chart etc. It is free, open source and available under MIT License

http://d3pie.org - nice pie charts, offers good options. Available under The MIT License (MIT)

xCharts - D3.js based library. Uses HTML, CSS, SVG. Default charts have polished look but very limited options. It was developed by https://www.tenxer.com/ and has been made free with no strings attached.

dimplejs - it's crazy, it's very good. Reads data from CSV and other text formats. It's an open source project by http://align-alytics.com/

dc.js - It's amazing. Just look at this - http://dc-js.github.io/dc.js/ . It's available under Apache License, Version 2.0 (the "License")
Few cool examples - http://dc-js.github.io/dc.js/examples/cust.html

http://d3plus.org Built on top of d3.js, simple to use. It has limited set of examples/apis but some of them are pretty good e.g. Geo Map, Tree Map, Simple Network.

https://github.com/mbostock/d3/wiki/Gallery - d3 example library in one page.

http://bl.ocks.org/mbostock - Amazing Charts and Visualization examples by Mike Bostock, the creator of D3.js and many other libraries used for rich visualization.

http://bost.ocks.org/mike/ - Another repository by Mike Bostock.



Pure JavaScript (#7)

Flot Charts - Flot Charts are pure Javascript library. 

Google Charts - It has very rich set of charts & graphs, simple to use. You may like look at AngularJs Google Chart Tools directive as well.

jqPlot - Pure Javascript charting library.  Offers good set of options. Look and feel is not very polished. 

morris.js - raphael.js based library. Simple to use. Cool fluid look and feel. Limited options, most of the common scenarios are covered. Uses HTML, CSS, SVG

dygraphs - has good options for time series graphs but look and feel is not very polished. However, dygraphs can handle huge datasets running into millions of data points. It takes .txt and .csv file as inputs.


Mini Charting Libraries (#8)

Peity is a simple jquery.com plugin that converts an element's content into a simple mini pie line or bar chart.

jQuery Sparkline generates sparklines (small inline charts) directly in the browser using data supplied either inline in the HTML, or via javascript.

Commercial Offerings (#9)

This section covers a list of Charts and Graphs libraries. 

HighCharts - Offers very rich set of options. Probably one of the best libraries. Available free for non-commercial use, under Creative Commons License

Canvas JS - HTML5 JavaScript Charting Library with a simple API and 10x better performance compared SVG/Flash based charts. Charts are responsive & can run across devices including iPhone, Android, Desktops, etc. Offers good set of Charts and Graphs.

ZingChart - Javascript Charting library. Offers a rich set of options. Visually appealing and good for large data sets. 

JSChart - Javascript Charts. Limited options.

Ember Charts - Open source. Offers limited popular options of Charts, Tables etc.


Charts & Graphs as Service (#10)

statspanda.com : Offers creation and hosting of Charts, Graphs & Dashboards as Service. Purely REST Api driven, also provides a REST API Console.

tableau.com : Grand Daddy of Visualisation. Offers Desktop, Cloud based service as well as Back-end data integration. Has huge array of very creative Charts Graphs and Dashboards. 

datawrapper.de : Creation of embeddable Charts, Maps. Primarily used by publishing, news and media companies. 

chartio.com : Provides data integration layer with various sources like CSV, Amazon RedShift or Stripe etc. It then convert the data into intuitive Visualisation such as Charts, Graphs etc. 

chartblocks.com : Basic chart building tool. It reads the data from spreadsheet and gives tools to customise the cha. The charts can be shared in popular social media or can be embedded as iFrame. 

infogr.am : Create charts and infographics and publish them easily, primarily used by the news and media companies. 

jaspersoft.com : Create Charts, Graphs, Dashboard and embed or publish. Good integration with Amazon AWS (RDS, Redshift) . Owned by TIBCO. Focuses on Apps and On-prem Applications for embedding the Charts, Graphs and Dashboards created/hosted on their platform.

Amazon QuickSight : Works with AWS dataset to create visualisation. Offers good set of Graphs, Tables and other Visualisation. Amazon QuickSight uses a new, Super-fast, Parallel, In-memory Calculation Engine (“SPICE”) to perform advanced calculations and render visualizations rapidly. 

domo.com : Offers good set of connectors. Dashboards are designed for specific roles and for different industry verticals. Rich set of chart, graphs and other visualization are offered. Provides good integration with Amazon. 

gooddata.com : Primary positioning as Business Intelligence platform for real time analytics. Offers good set of charts, graphs and dashboard options. Has good set of Connectors.

keen.io : Keen IO is an API platform that lets developers collect and study custom events at a massive scale and converts them to visualisation. It's designed in a way so that users can embed the visualisation like Charts, Dashboards in their apps, websites. All popular Charts and Graphs are available for the Dashboards. 

chartbeat.com : Focused at Advertising and Publishing industries. Provides insights using Charts, Graphs and other visualisation.

vida.io : Create attractive visualisation, embed and use them, can create Dashboard. 

plot.ly : Plotly is the data visualization and collaboration platform for engineers and data scientists. Provides integration with Python, Excel, MATLAB & R. Offering available in three different formats - Cloud, On Prem, and Desktop Tool. 

zoomdata.com : Focused at Visualisation for Big Data, can handle large volume of data. Provides integration with Hadoop, NoSQL, ElasticSearch, Solr and Spark.

datahero.com : Provides connectors to a large array of Cloud Based data sources like Box, Dropbox, Stripe, Hubspot, MailChimp, Google Drive, Google Analytics, MixPanel etc. You can then convert the data to a Chart, Graph and other visualisation. You can create one dashboard composed of multiple Charts, Graphs from different Datasources. Have focused offerings for various industry verticals.

getdataseed.com : Provides data exploration tool sets that can import data from Excel Spreadsheet and from fetch Dropbox, Google Drive or any public url. Can convert the data into Charts, Graphs, Maps and other Visualisation. Provides REST Api integration. 

anaplan.com : Excel Spreadsheet on Cloud with ability to create crisp Charts, Graphs, Dashboards and other visualisation on the fly. Excel Spreadsheet like data is editable. Has focused offerings around Finance, Sales, Operations and HR.

collabion.com : Creates Charts, Graphs and Dashboards from SharePoint Data. Need to Downloaded and installed, need to Server Licenses.

www.domo.com : Offers good options of Charts, Graphs and other visualisation. Provides wide range of connectors for various datasources and Apps ranging from Excel, Google sheets, Box, Facebook, Marketo, Salesforce etc. Solutions are tailored for various roles, industry verticals and operations.

www.klipfolio.com : Offers creation of company Dashboards composed of cool Charts, Graphs and other visualisation. Provides a wide range of connectors to all major datasources and various options for data infusion. 

Microsoft Power BI : Available as Desktop Application and mobile Apps for iOS, Android and Windows Phones. Wide range of Charts, Graphs can be glued to a Dashboard quickly, supports drag and drop features. Provides wide range of Datasource connectors. Supports natural language query inside a Dashboard.

RJMetrics.com : Offers creation of Charts, Graphs and Dashboards on Cloud. Has good Data Integration with popular Datasources. RJMetrics Pipeline transfers the to Redshift. Good for large volume of Data - from less than 5 million rows to upside of 500 million rows sync up. 

http://kilometer.io : SaaS Analytics Tool

https://chartmogul.com : Subscription Analytics

gramener.com They are into creative visualization. Provides visualization of the data from various business operations like Sales, Marketing, HR etc. 

Finally, I still feel the blog looks like a work in progress. In fact, I would spend more time taking a deeper look into all the companies listed under #10. However, I hope you found the post useful even in current shape and form.

How to Add Charts, Graphs and Visualization to a Blog Post

Over last few months, I have written few blog posts where I used pretty sophisticated Charts, Graphs and Visualization. They make the blog post lot more meaningful and readable, readers love those charts, tables & displays. In this post, I will share the service I use for those visuals and few quick steps on how to create a chart and add that to your blog.

I use the services of http://www.StatsPanda.com, it's in public beta and it's free for "Individual" users.

StatsPanda.com Home Page

Once you login using your Facebook account, it redirects to registration page and during registration it gives two options - "Individual" and "Enterprise". Choose "Individual".

Go to API Console. I will suggest that you spend some time exploring the listed Visualization APIs. You may have to spend some time trying out the APIs in order to understand the JSON Inputs structures. It's not super complicated, although it may take some time. Initially you can use the example input data provided along with API Documentation.

API Console to create Charts, Graphs etc

Once decided, go ahead and create the Chart of your choice and that would give you a unique url and iFrame Code for your chart. For this blog I created a Stacked Area Chart with the example data that API Console provided. In the above screenshot, you can see a greyed out box right below the chart. The content of that box is copied below for your reference. You can straight away use that in your blog.

<iframe src="http://www.statspanda.com/charts/ui/nvd3/stacked-area-chart?unique_key=dc27b99e933da8969218fb3d8e06aeafdc61e170dad7c5aed1b5d41381b2d8" style="border: #E8E8E8 1px solid; height: 300px; width: 100%;"></iframe>

Here is an example how the Visualisation really appears in a blog once you include the iFrame code snippet to your blog.



The chart is dynamic and not a jpg pic, data is being served from the Website. In fact, if you want you can "Edit" the input JSON Data and have a different chart of same type.

Hopefully you find it useful and able to use the charts and graphs in your blogs.

Top 10 Machine Learning Libraries and Services for Java Developers

In recent times, Machine Learning has emerged as one of the most talked about topics in the field of information science and technology. Although the subject has probably intrigued the researchers and academicians for decades, it's only now every Software Engineer is trying to get a hang of it. It's changing the face of computing for ever and in a way, accelerating the move from Software Eating The World to Software Eating Software. 

In this blog, I will cover (rather list out) some of the top Machine Learning libraries and services available for a Java developer and try to highlight some of the salient points associated with each of those libraries. Please note some of the most powerful machine learning libraries are in Python and they are not covered in this post at all.

Before I go to the list of libraries, here is a short list of algorithms or broad categories of problems that most of the Machine Learning libraries would cover either fully or partially -

• Classification
• Regression
• Clustering
• Ranking

Here are few examples of application:

• Outlier Detection
• Recommendation
• Natural Language Processing
• Neural Networks

1. Apache SPARK MLlib 

• Top ML Library and growing fast
• Part of Apache Spark
• Infinite Scalability
• On Premise - https://spark.apache.org/mllib
• As Service - https://databricks.com/product/databricks-cloud

2. Deeplearning4j - http://deeplearning4j.org

• One of the top ML Libraries in Java
• Integrates with Hadoop, Spark
• Use Cases
• Face/image recognition
• Voice search
• Speech-to-text (transcription)
• Spam filtering (anomaly detection)
• E-commerce fraud detection
• Regression 

3. Apache Mahout - http://mahout.apache.org

• Runs on Hadoop Cluster, so infinite scalability
• Good for recommendations

5. Google Prediction APIs (as service) - https://cloud.google.com/prediction

Types of the problems where you may see few ready examples 

• Classification
• Regression 

Google Prediction provides two types of APIs - one that leverages the hosted models and the rest where you have to train the model with you sufficient and then expect the APIs to predict.


5. IBM Watson + AlchemyAPI (as service) - http://www.ibm.com/smarterplanet/us/en/ibmwatson


You can try Alchemy APIs at http://www.alchemyapi.com/products/demo . One of the newest and a very significant acquisition by IBM is Alchemy API. Here are the two primary offerings from Alchemy API -

• AlchemyLanguage - Text Analytics and Natural Language Processing.
• AlchemyVision - Leverages deep learning for photo and image processing.

Easy to get registered and get started.


6. MS Machine Learning (as service) - http://azure.microsoft.com/en-in/services/machine-learning

Microsoft has done big deal around coming up with an intuitive UI where users can create a model, train it and run the analytics - all in drag and drops. For a new user it would take sometime to get accustomed to the various UI controls and how Application works. Developers can create her own model and sell it in Azure Marketplace. 

Outlook account works seamlessly. 

7.  Amazon AWS Machine Learning - https://aws.amazon.com/machine-learning 

• Provides visualisation tools to create ML Models
• Simple API support for the models generated this way
• Highly Scalable, can generate billions of predictions in day

8. Weka - http://www.cs.waikato.ac.nz/ml/weka

• Provides a graphical user interface, command line interface and Java API
• One of the most popular Java machine learning library
• Available under GPL License 

9. Mallet - http://mallet.cs.umass.edu

• Statistical natural language processing, document classification, clustering, topic modeling and information extraction.

10. H2O - http://0xdata.com

• In-memory data engine
• Designed for running various types of types of statistical computations (including Deep Learning)
• Works with Hadoop Distributed File System

Others deserving a mention but could not make it to the list of top 10

JSAT - https://code.google.com/p/java-statistical-analysis-tool

• Library for quickly getting started with Machine Learning problems
• Available under GPL 3 but author is open for discussion
• List of supported algorithms is impressive
• One man project, done in his free time. Creator is Edward Raff @EdwardRaffML

LensKit - http://lenskit.org

• Focused on building recommender system, primarily for research based projects
• Good for trying out. For scale and for production env, one can move to Apache Mahout.
• Actively developed, managed.

oryx - https://github.com/cloudera/oryx

• Built on top of Mahout
• Supports streaming instead of batch jobs, making it realtime
• Still in early stage. 

Java-ML - http://java-ml.sourceforge.net

• Provides a collection of algorithms
• No new release since 2012

Hopefully you will find this short post useful. Leave your feedback and comments.

Connect to me on twitter: @satya_paul

Did TCS under report employee strength in 2014?

In the wake of recent news on layoffs at TCS, I started looking into TCS Annual reports. As I sifted through the Annual Reports of TCS from last 10 years, something did not seem to be quite right around the employee strength reported in Annual Report in 2014.

Usually, for an IT Services Company the revenue growth is directly proportional to the growth in employee strength, it seemed odd that the Y-O-Y growth rate of Operating Profit was the highest in 2014 whereas the employee growth was the lowest in the history of TCS. So, I started looking into the details of the data based on TCS Annual Reports since 2005.

All the Charts are reproduced courtesy of www.StatsPanda.com

TCS Y-O-Y Key Growth Data (Revenue, Op Margin, Productivity, Avg Salary, Employee Growth)

 

In 2014, Operating Profit Growth(39.43%) and Productivity Gain(19.57%) were the highest while Employee Growth (8.62%) was the lowest in the history of TCS. That surely raises some eyebrows.
So, I decided to regenerate the same chart by tweaking the Employee Strength from 300,000 (as reported in AR 2014) to 320,000. Now, take a look at the chart below. The numbers look much more realistic - Productivity Gain(12.1%) & Employee Growth (15.86%).

TCS Y-O-Y Data (with Employee Strength as 320,000 as on Mar'2014)

Next, I looked into the Average Salary Increase data and the productivity gain data and tried to see the impact if we change the Employee Strength from 300 K to 320 K at the end of Mar'2014. Check out the generated charts and the analysis below.

TCS Avg Productivity vs Avg Salary when Employee Strength is 300K as on Mar'2014

As per the chart above, TCS Employees received an average salary increase of 14.35% in 2014 vis-a-vis 11.81% in 2013. But little bit of poking with the TCS employees revealed the Average Salary Increase was better in 2012 - 2013 than in 2013 - 2014. Then why do we see a different trend here? Now, take a look at the same chart but regenerated with employee strength as 320 K. You can see the Average Salary increase matches with the data heard from TCS Employees and consistent with the past trend - Average Salary increase changed from 14.35% to 7.21%.  

 
Similar observations can be made with respect to Productivity Gain as well. 19.57% Productivity gain in one financial year is near impossible unless there is fundamental shift in Business Model and Revenue Pattern. With correction of employee strength from 300 K to 320K, even that's get corrected from 19.57% to 12.1%, a much more grounded and realistic number.

TCS Avg Productivity vs Avg Salary when Employee Strength is 320K as on Mar'2014

Last take a look at the Utilization Ratio in the chart below. It has remained range bound over the years and no significant upside can be seen in 2014 w.r.t 2013. So, when there is no significant change in Utilization Ratio & there is a dip in Employee Growth Y-O-Y, how one would explain such a steep growth in Revenue as well Operating Profit level and gain in Productivity gain in 2014?

TCS Key Metrics along with Utilization Ratio

Based on the data and analysis above, it appears TCS did under report their total employee strength at the end on Mar 2014. This observation actually seems to be consistent with the news around layoffs as that's the only way to get rid of off-the-book-employees. So, why did TCS do this?

To show (at least on paper) -

1. Superior growth in productivity number.
2. Higher Average Salary Growth. 

These two are important factors when it comes to valuation. Above par growth in Productivity numbers does indicate superior & improving asset quality, better strategy and execution and that commands higher premium. Higher Average Salary Growth gets reflected in lower attrition as well as increase in ability to attract talent. These factors help de-risk current and future revenue stream. Overall, these numbers are very critical and any change in their values will have impact on the valuation as well. The under reporting does mean that TCS was planning to lay off people in subsequent period. However, this seems to be a reactive step and does indicate the discomfort of TCS Management with the increase in employee strength in higher salary band.

This problem is not going to disappear any time soon. You can expect to see similar waves of employees-in-higher-salary-band continue to keep bothering TCS Management for next few years (basically the hires from 2006 and earlier). The organisation will go through a period of turbulence and may force the management to re-engineer the business model leading to a period of uncertainty for TCS for next few years.

As for the ongoing lay offs this year, there seems to be a sense of urgency and an accelerated pace of reducing the employees at higher salary band. In fact, it's getting uglier now. This could be related to the fact that TCS under reported the employee strength at the end of Mar 2014 and as a result they have to do a catch up job now. So, while superior productivity numbers & better average salary increase made the investors happy & the valuation sky rocketed, it also created a challenge for TCS to handle now. It brought in unnecessary instability to the organisation and increased the risk. TCS needs to handle the employee layoffs more gracefully and better manage the investors' expectations to make the growth story sustainable. Hopefully they report more grounded numbers but closer to reality.

The valuation at current level does seem to be very high and I will not be surprised if we see a correction of 20% in next one to two years time (Rs 1900 - 2100 range).

This article is speculative in nature and a work of subjective interpretation of data available in public domain. I am not a TCS employee and never I was. I don't have any professional relationship with TCS whatsoever and have no share holding of TCS.

You can follow me at @satya_paul 

Distributed Datastores - let's take a look under the hood

This one is continuation from my last post where I had looked into various alternative options to traditional RDBMS Databases. In this post I will cover some of the basics and go over the factors that influence the choice of a datastore in general and NoSQL Databases in particular. I will also cover the trade offs associated with a choice of a datastore.

Fundamentally, a Database is a specialized software system that allows you to write/store ( i.e. create, update, delete), read, and even do some amount of processing of the data e.g. executing the aggregate functions.

In a world dominated by RDBMSs, Databases are expected to be ACID compliant, in fact, a must have & an important measure of Quality. This is the case with all the RDBMS and they have been doing that job fairly well for many decades. So, what changed recently? To the core, there are really few handful needs that became very important -

• Increased Complexity of relationship
• Need for Flexible Data Structure
• High Availability
• Scalability (typically referred as Web Scale) 

Increased Complexity of relationship between entities is handled well by Graph Databases. Typical applications include recommendations, social network etc.

Document Databases do exceedingly well when it comes to supporting Flexible Data Structures. Column Family Databases also provide some amount of flexibility, each row can have a different set of attributes. However, in this post, without getting into further details on those factors, I will shift the focus on last two points and explore how various parameters really influence the choice.

So, how does anyone achieve High Availability (HA) for any system? By building redundancy into the system and databases are no exception, they create replicated failover nodes. Failover nodes are exact replicas of the master node and remains passive unless required. Usually, Databases ensure HA but the challenge of ensuring HA is different when it comes to distributed, partitioned databases. Second, it is one thing ensuring HA against a node or machine failure and it's entirely different thing when it comes to ensuring that at no point DB should be unavailable should there be a Network, Machine, Power or any other failure e.g. Data Center goes down. Typically it is achieved by putting the replicas across different Data Centers spread over different geographies and those replicas are not offline. This is also known as Geographically Distributed High Availability (GDHA). Thus Network Partition tolerance becomes critical. Not all databases support GDHA. Note GDHA is more than Disaster Recovery (DR) where in the replicated nodes remain offline and used only when any disaster hits the master node. Usually the focus of DR Systems is not limited to Databases, they kind of keep the entire stack ready.

Other big issue is really about Scalability. How much a database (read RDBMS) can grow? It can grow as much as the largest machine will allow it to grow. But what if you hit that ceiling too? Obvious answer would be to put a second machine. That's correct, but then can the Database still meet the important quality measure called ACID or can be made highly Available when some of  the operations (i.e. reading, writing or processing of the Data) are happening in distributed systems? A simple answer is NO and that's the time you start looking into trade off matrix. You take a second look into the operations as discrete activities and take a call on what is critical for your business and what you can give up.

Before we go any further lets put the definition of ACID for reference:

• Atomic: Atomicity refers to the ability of the database to guarantee that either all of the tasks of a transaction are performed or none of them are. 
• Consistent: The consistency property ensures that the database remains in a consistent state before the start of the transaction and after the transaction is over (whether successful or not).
• Isolated: Isolation refers to the requirement that other operations/transactions cannot access or see the data in an intermediate state during a transaction.
• Durable: Durability refers to the guarantee that once the user has been notified of success, the transaction will persist, and not be undone.

Databases achieve these by effective handling of Concurrency i.e. how many person can act or modify the state of the data. Here are the various Concurrency handling mechanisms/options:

• Lock or Exclusive Lock or Pessimistic Lock. Some databases allow only one user to modify a record, row or document at a time. Preventive.
• MVCC (multi-version concurrency control) or Optimistic Lock is a mechanism that guarantees consistent reading. It allows multiple users to modify a record with multiple conflicting versions without acquiring an exclusive lock. However, it puts a check when it comes to committing the changes into the database. At that point it allows a successful commit only for the first user to attempt. 

Locks ensure changes are either committed or rolled back in case of a successful transaction and it rolls back everything in case of transaction failure.

Next, lets take a look at Replication i.e. copying the datastore to a different node. High Availability is achieved by replicating a database node. Replication comes in two forms:

• Master-slave replication makes one node the authoritative copy that handles writes while slaves synchronize with the master and may handle reads.
• Peer-to-peer/Master-Master replication allows writes to any node; the nodes coordinate to synchronize their copies of the data.

Master-slave replication reduces the chance of update conflicts but peer-to-peer replication avoids loading all writes onto a single point of failure. Other important factor to consider is when data is written on a node, it takes time before it is reflected on all the nodes. You can do it either synchronously or asynchronously for a particular transaction. Your choice will determine whether your database supports Consistency or Eventual Consistency. In case of Peer-to-Peer replication, the same record can be modified by two different transactions on two different nodes. How a databases handles these scenarios is also influenced by the choice of Consistency viz-a-viz Eventual Consistency. There are specific databases that excel in one usecase over the other. I do plan to cover that in my next blog.

While Database replication primarily helps to handle failover and ensures Higher Availability, it also helps Scalability. Master-Slave replication works well for Read Scalability while write operations can take place only on the Master node and Slaves then syncs up with the Master either synchronously or asynchronously. Peer-to-Peer or Master-Master replication helps achieve both Read and Write Scalability as both read and write operations can take place on all the replicas. Here, all the replicas will have the same copy of Database. This is traditionally known as Scaling Up or Vertical Scaling where a Database system can scale as much as a node can grow. This should work well for most of the systems. However, for infinite scale or web scale, one needs to go for Scale out or Horizontal Scaling where data is Partitioned or Sharded across multiple nodes. This allows Databases to grow infinitely just by adding new hardware (usually commodity hardware). Note each partitioned node will have different set of data and may have its own replicas for high availability, each partitioned node is actually a database in it's own capacity.

Scale Up vs Scale Out

Now lets look at the trade off matrix I mentioned earlier in this post. This trade off matrix is known as CAP Theorem. This is also known as Brewer's theorem. It states that it is impossible for a distributed computer system to simultaneously provide all three of the following guarantees:

• Consistency (all nodes see the same data at the same time). Note this consistency is different than what it is in ACID.
• Availability (a guarantee that every request receives a response about whether it was successful or failed)
• Partition tolerance (the system continues to operate despite arbitrary message loss or failure of part of the system)

A distributed system can achieve only two of them at a time.

Here is a nice summary of how different Datastores complies with CAP Theorem from a presentation by Aleksandar Bradic

Finally, here is a matrix, I prepared, to capture various parameters that one would consider while analyzing a Distributed DB System.

Not all values are filled. I will continue to work on this and update it further. 

Connect to me on twitter @satya_paul

Check out my storyboard on www.fanffair.com - http://www.fanffair.com/storyboard/satyajitp2011

RDBMS - one size fits all, but not anymore

Relational Databases (RDBMS) have been the de facto standard for storing information for fairly significant period of time, almost since the dawn of the Information Technology Industry. For software developers, traditionally, the obvious choice has been RDBMS and most of us really never gave any serious thought for storing data in any other format. But that's not the case anymore, now the developers are spoilt for choice. In last decade, many new offerings came out of the Labs of Google, Amazon, Facebook, Apache foundation & many niche DB Vendors. They are now mature with proven track records and challenging the dominance of Relation Databases. Yes, I am referring the new generation of Databases i.e. NoSQL Databases (also known as Schemaless Databases), In-memory Data stores and what they call New SQL Databases.

In this post, I will introduce the broad categories of NoSQL databases & various offerings under each category, a brief mention of various drivers behind this movement, cover few use cases where NoSQL databases will score better than the regular RDBMS databases and finally what you give up when you gain so much. What I don't intend to cover here is a detailed technical explanation of key concepts like ACID, BASE, CAP Theorem or details on sharding, partitioning, replication, map-reduce and other associated technological concepts. I do have a plan to cover them in a separate blog post.

Broad Classification of NoSQL Databases

NoSql or Schemaless Databases can broadly be categorized as following:

A key-value store: it's a storage system that stores values indexed by a key. Typically, you can query only by the key and values are opaque and can not be used for querying. This allows very fast read and write operations (a simple disk access) and this model is seen as a kind of non volatile cache (i.e. well suited if you need fast accesses by key to long-lived data). This category of databases are largely inspired by a paper by Amazon based on DynomoDB.

 A document-oriented database extends the previous model and values are stored in a structured format (a document, hence the name) that the database can understand. For example, a document could be a blog post and the comments and the tags stored in a denormalized way. Since the data are transparent, the store can do more work (like indexing fields of the document) and you're not limited to query only by key. IBM's Lotus Notes Database has largely influenced this category of databases.  

A graph database is a database that uses graph structures with nodes, edges, and properties to represent and store data. A graph database is any storage system that provides index-free adjacency. This means that every element contains a direct pointer to its adjacent elements and no index lookups are necessary.

Source: Presentation by Emil Eifrem, CEO of neo4j

Wide-Column Stores are derived from Google's BigTable. BigTable is a compressed, high performance, and proprietary data storage system built on Google File System. BigTable maps two arbitrary string values (row key and column key) and timestamp (hence three dimensional mapping) into an associated arbitrary byte array. It is not a relational database and can be better defined as a sparse, distributed multi-dimensional sorted map. BigTable is designed to scale into the petabyte range across hundreds or thousands of machines. Every row can have it's set of columns.

Paper by Kai Orend

A New SQL Database - Simply put it's Scalable RDBMS.

Note above definitions are summed up from various sources.

Examples of NoSQL Databases 

Here are few examples from each of them, I tried to pick the popular ones from each category.

Key-Value Stores: Redis, Riak, DynamoDB, Voldemort
Wide-Column Stores: BigTable (Internal to Google, partially available via AppStore), Cassandra, HBase
Document Stores: MongoDB, Couchbase, CouchDB, SimpleDB
Graph Databases: Neo4j, InfiniteGraph, FlockDB

Although the following lists are not part of NoSQL family but they are still very relevant and significant, hence mentioning few examples -

In-Memory Stores (Key-Value): memcached, Ehcache

New SQL Databases: VoltDB, Amazon RDS, NuoDB, MySQL Cluster, Clustrix

And for old time's sake -

Relational Databases: Maria DB, MySQL, Microsoft-SQL, IBM DB2, Oracle.

As you see the list is pretty long and at times, it may be slightly confusing as well. In the lists above, the products are grouped under a specific category based on what they claim or their dominant characteristics. But there could be instances where some of the offerings may overlap into two different categories.

Here is nice infographics grouping different offerings under different categories published by Matthew Aslett in his blog.

Usecases and Drivers behind the adoption of NoSQL Databases

Here are few drivers and usecases that drove the adoption of NoSQL databases.

Scalability: With the advent of consumer Internet and massive penetration on mobile devices, scalability challenges increased exponentially. Answer to this challenge was to go for Scale-out architecture (horizontal scaling) over Scale-up (vertical scaling). Traditional RDBMS offerings either failed to adopt the scale-out architecture or they became too complex to manage at that scale. NoSQL databases, particularly Key-Value Pair Databases and Wide Column Stores or Column family Databases came to the rescue. Some of the shining examples would include Google's BigTable and others in it's family (e.g. HBase, Cassandra etc), Amzaon's DynamoDB or other key-value stores like Redis, Riak and finally in memory caches like Memcached etc.

Here is a slide from Emil Eifrem's presentation that shows how various types of DBs are stacked up when it comes to scaling etc.

Schemaless Databases: Scalability was one major reason why companies experiencing Web Scale traffic started looking beyond RDMBS, but a large section of developers started adopting NoSQL Databases for their flexibility. Document databases with their ability to store as well as query JSON/BSON data became popular among the developers. It would be a nightmare to store data like Blogposts, Comments, nested comments etc using a traditional RDBMS and more so as those structures keep on changing. Lot of time, particularly in case of configuration data, each record may have different attributes. Using regular RDBMS databases managing such use cases become very messy - typically for a single such record there will be as many name value pairs entries as the number of attributes or one large table with all possible columns and each record will make use of a fraction of those columns. It becomes quite messy in terms of reading, writing and maintainability point of view. However, Document Databases and Wide Column Databases will work just perfect in these usecases. Unlike traditional RDBMS Databases, Schemaless databases keep related information together and avoid joins.

Economics: Third major reason lies in the economics around capacity building. Unlike large enterprises where they can forecast the future capacity requirements as well as have deep pocket to finance the large servers upfront, start-ups really don't know whether they would be successful enough to invest in large servers coupled with the fact that capital is far more scarce resource in the star-up world. So, upfront investment on large servers becomes a very difficult proposition & a huge entry barrier. They really needed a way to add capacity as they move on (& as far as they move on) and without impacting existing services. The architectures proposed by Amazon's Dynamo DB and Google's BigTable came to the relief. Scale-out architecture took over Scale-up architecture enabling companies to add capacity as they need them & allowing them to grow infinitely. This reduced the need for upfront capital requirement and lowered the entry barrier.

Graph Database: Need for Graph databases is distinctly different from others and represents the real world problem as they appear. In real world everything in interconnected either directly or through others. Graph represents all those connected entities as "nodes" and the connections as "relationship". It can help answer questions like "How do I reach New York?" or "How am I connected to Matt?" or "What movie Julie may like to watch?". Traditional databases fail to scale as the relationships become deeper. Large Social Network sites like Facebook, Linkedin, Twitter have their own graph database implementation and there are few commercially available Graph databases for everyone's use. However, Graph databases are not as scalable as other families of NoSQL databases.

One of the points I did not highlight here is "availability", that's really not a point of differentiation anymore rather a point of parity and most of the RDBMS vendors as well as NoSQL databases provide high availability through replication and fail over.

Conclusion

As you start benefiting from some of the NoSQL features, you must be aware of the areas where you have to give up. Here is short list of areas where you may have to do compromises:

• Accepting Eventual Consistency over transaction level consistency (ACID)
• Increased Complexity: Organizing data in a way that all related information remain under one sharded node. 
• Absence of SQL Query (for some of them).
• Absence of Joins or Aggregate Functions.
• Slowness of Two Phase Commits (2PC) when data spread over multiple nodes.
• Any mass update.

So,  it really comes down to the point that no single solution will address every problem we have and neither they are meant for that. But the good point is we have options.

To put the things in perspective, take a look at the projected revenue growth for various categories as per a study done by the 451 group

Connect to me on twitter @satya_paul
Check out my storyboard on www.fanffair.com - http://www.fanffair.com/storyboard/satyajitp2011