Category Archives: Amazon Web Services

Preparing for the Storm

storm-02

In a previous post, I described Big Data as containing certain criteria: Volume, Velocity and Structure (probably worth revising to Variety as that is now more commonly used). While Hadoop is currently the primary choice for analytics on Big Data, it is not necessarily designed to handle  velocity. That is where Storm comes in.

Hadoop is designed to operate in batch mode: its strength is in consolidating a lot of data, applying operations on the data in a massively parallel way and reducing the results to a file. Since this is a batch job, it has a defined beginning and a defined end. But what about data that is continuous and needs to be processed in real time? Given the popularity of Hadoop, several open source projects have risen to meet this challenge. Most notably, HBase is the project that is designed to work with Hadoop and provide for storing large quantities of sparse data leveraging the Hadoop Distributed File System (HDFS).

Storm Use Cases

But what about real-time analysis of a Twitter stream? This use case is a perfect fit for Storm. A recent project that was developed by Twitter, it is similar to Hadoop but tuned for dealing with high volume and velocity. Storm is a project that is intended to handle a real-time stream of information with no defined end. It can handle analyzing streams of data forever. if you are familiar with Yahoo’s S4, Storm is similar.  It handles three broad use cases (but not exclusively these):

  • Stream processing
  • Real time computation
  • Distributed remote procedure calls (RPC)

Storm is relatively new but has been adopted by several companies, most notably Twitter (obviously), Groupon and Alibaba. You’ll notice that even though these companies are in different industries, they all have to deal with large amounts of streaming data.

Components

Storm integrates with queuing systems (like Kestrel or JMS) and then writes results to a database. It is designed to be massively scalable, processing very high throughputs of messages with very low latency. It is also fault-tolerant so if any workers or nodes in a cluster die, then are automatically restarted. This gives it the ability to guarantee data processing – since it can track lineage, it will know if data isn’t processed and can make sure that it does. Since Storm has no concept of storage, you will need to store the results in another system (like a database or another application).

Storm also has its own components which bear some explaining. Although there are several components in a Storm cluster (including Zookeeper which is used for coordination and maintain state), there are three which are referenced frequently:

  • Topologies – this is equivalent to a Map Reduce job in Hadoop except that it doesn’t really end (unless you kill it of course)
  • Spouts – this is the primitive that takes source data and emits it into streams for processing
  • Bolts – this is the primitive that takes the data and does the actual processing (like filtering, joining, functions, etc.). This is similar to the part of the Map Reduce WordCount example that is summing the count of words.

Since Storm is intended to run in a highly distributed fashion – it is also a natural fit for the cloud. In fact, there is a project that makes it easy to deploy a Storm cluster on AWS. As an open source project, it can run on any Linux machine and works well with servers that can be horizontally scaled. If you want to learn more technical details regarding this project, check out the excellent wiki on Github, where you can also download the latest version.

Big Confusion about Big Data

Not only do people often confuse what exactly the term “Big Data” means, but the dizzying array of products that are out there that solve for Big Data problems add to the confusion. So what’s the difference between Hadoop, Cassandra, EMR, Big Query or Riak?

First, it’s important to define what Big Data is. First of all, I refer to Big Data to mean the data itself – although it is often used interchangeably with the solutions (such as Hadoop). I believe that data should satisfy 3 criteria before being considered “Big Data”:

  • Volume – the amount of data has to be large, in petabytes not just gigabytes
  • Velocity – the data has to be frequent, daily or even real-time
  • Structure – the data is typically but not always unstructured (like videos, tweets, chats)

elephant_rgb_sq

Hadoop

To deal with this type of data, Big Data solutions have been developed. One of the most well known is Hadoop. It was first developed by Doug Cutting after reading how Google implemented their distributed file system and Map Reduce functionality. As such, it is not a database but a framework that implements the Hadoop Distributed File System (HDFS) and Map Reduce. There are several distributions of this open-source product – Cloudera, Hortonworks and MapR being the most well-known. Amazon Elastic Map Reduce (EMR) is a cloud-based platform-as-a-service implementation of Hadoop. Instead of installing the distribution in your data center, you configure the jobs using Amazon’s platform.

Why would people use Hadoop? Typically they are pulling large amounts of unstructured data and need to be able to run sorting and simple calculations on the data. For example:

  • Counting words – this is the standard Map Reduce example
  • High-volume analysis – gathering and analyzing large scale ad network data
  • Recommendation engines – analyzing browsing and purchasing patterns to recommend a product
  • Social graphs – Determining relationships between individuals

One of the weaknesses of Hadoop is its job oriented nature. Map Reduce is designed to be a batch process so there is a significant penalty is waiting for the job to start up and complete. It is typically not a strong candidate for real-time analysis. It also has the challenge of having a master node that is a single point of failure. Recently, Cloudera has released their latest distribution that implements a failover master node but this is unique to their distribution. Additionally, developers typically use Java (although other languages are supported) and not SQL to create Map Reduce jobs.

Riak_product_logo

Cassandra / Riak / Dynamo

Amazon wrote a paper on how to implement a key value store and Cassandra and Riak are implementations of that key value store construct. Amazon also has their own implementation called Dynamo DB.

While there are differences among the implementations, customers will want to use these products because they want fast, predictable performance and built-in fault tolerance. This is for applications that require very low latency and highly available. Typical use cases would be:

  • Session Storage
  • User Data Storage
  • Scalable, low-latency storage for mobile apps
  • Critical data storage for medical data (a great example is how the Danish government uses Riak for their medical prescription program)
  • Building block for a custom distributed system

These solutions are typically also clustered in a ring formation with no particular node marked as a master and thus so single point of failure. Customer will typically not use these solutions if they require complex ad-hoc querying or heavy analytics. As with Hadoop, building queries here require using a programming language like Java or Erlang and not SQL. There is also a project called HBase which provides similar columnar data store functionality for data on Hadoop so that is another option that is available. This project was modeled after Google Big Table and not Dynamo.

mondodb

MongoDB

This is a project that has become very popular and is managed by 10Gen and also implemented in a SaaS model by MongoLabs. MongoDB has become very popular because it is designed with developers in mind and scales the gap between a relational database (like MySQL) and key-value store (like Riak). MongoDB is excellent if you need dynamic queries and want to maintain similar functionality to a relational database that is more friendly to object-oriented programming languages. 10Gen emphasizes that MongoDB is designed to easily scale out and also accelerate development (for example, if conducting agile development). Typical use cases would be:

  • Companies storing data in flat files
  • Using an XML store for data too complex to model in a relational database
  • Deployments to public or private clouds
  • Electronic health records

As with it’s other NoSQL brethren, MongoDB does not use SQL. Instead you will use a programming language (like Javascript) to interact with the database. It is also not good with applications that require multi-object transactions which MongoDB currently does not support.

dremel

Dremel

No, this is not a power tool but a technology implemented by Google in a recent research paper. In order to understand this product, you need to be familiar with Hive and Pig since they are often compared. As noted above, Hadoop has two main challenges – that the jobs have a start-up lag and that Java (or some other programming language) is required to write and implement MapReduce. Hive and Pig are open-source projects that sit on top of Hadoop and allow for users to implement SQL (in the case of Hive) or a scripting language (in the case of Pig by writing Pig Latin) to query data. Note that since this is a layer above Hadoop, MapReduce code is still being written and executed. It just translates SQL or Pig Latin into MapReduce code. This does solve for the case of being able to write queries against Hadoop but does not solve the latency problem. This is what Dremel attempts to do.

Dremel is designed to allow users to write almost real-time, interfactive and ad-hoc queries against a large data set. Because it uses its own query execution engine and doesn’t rely on MapReduce, this attempts to solve for the Hadoop latency problem. However, in its current implementation, it is purely an analysis tool since it is only read-only and organizes data in a columnar format. While there is an open-source project for this product, it is also currently commercially available as Google Big Query.

Where to go with Big Data

As you can see, there are a lot of products out there that all solve for Big Data problems in the different ways. It’s important to understand your use case and see which product is a best fit – it’s unlikely that you can find one that solves for the universe of Big Data problems. In addition, it is likely that you will still need to use a traditional relational database if you need to do reporting or interface with an enterprise application. Out of all the projects, Hadoop has the largest ecosystem with other supporting projects that broadens the functionality of the core product. As mentioned earlier, Hive is used to write SQL queries; there is also Sqoop that integrates relational data and HBase that provides low-latency capabilities. Fortunately, the majority of these projects are open-source so you only need to learn how to use them and find a cost-effective platform to implement it – like the cloud, a natural fit for distributed databases and NoSQL projects.

retention_dashboard_rptsrvr

Customer retention metrics

Last night (Tue July 19th), I was fortunate to be able to speak to the SVForum Business Intelligence special interest group (SIG).

After introducing the audience to DASHbay, I took them through an implementation we did using our Quick Analysis practice, which leverages open source software (especially BIRT and postgresql), cloud computing (on AWS), and rapid, iterative development.

The implementation itself was a dashboard, built with BIRT in less than a week and showing metrics for account acquisition and retention. The metrics help any business track not just how well they are acquiring customers, but how well they are keeping them.

Account retention dashboard
Our customer was able to get at the metrics via a URL to a server running in the cloud, set up just for them. It’s a great way to leverage cloud computing: no IT procurement costs or delays, and you only pay for it while you need it.

We talked about DASHbay’s Report Server product, which among other features, allows us to capture any useful piece of the report, and include it in any web page. It also provides permissioning and authentication, taxonomy for organizing reports, and more.

I got an excellent reception from the audience, and was pleased with the reaction and discussions afterwards. Thanks to all who attended!

If you didn’t get a chance to be there, please get in touch so we can talk to you more about our Report Server for BIRT, our Quick Analysis Service, or many custom BI and Data Analytics services. Customer Retention is one very useful application which we can provide, but our tools and techniques are applicable to most common business analysis problems.

Terry

Why DASHbay didn’t crash when Amazon-East did

By now, we’ve all heard about the Great Cloudburst of 2011. On April 21, Amazon’s Virginia-based data center experienced a huge reduction in service, triggered by what the company called “a networking event” and subsequent “re-mirroring of EBS volumes”.

I’ll leave examinations of the cause and response to other websites, and discuss the impact to DASHbay.

DASHbay builds and supports data-centric applications, focusing on open source software solutions, and often using cloud deployments. Amazon is our most frequently-used cloud data services provider.

At the time of the crash, several of our customers had mission-critical DASHbay-deployed applications running in the cloud. How did those customers fare, and therefore, how did DASHbay fare, since our customers’ problems are our problems?

I’m pleased to report that none of our customers were severely impacted by the outage.

Why not?

Here are some case studies of apps we built for clients, and the mitigation strategies that saved our bacon during Amazon’s failure.

The first is a real-time, high-availability mobile analytics collection application we deployed for Nielsen Mobile. Because this app’s continuous availability is mission critical, it was designed to not be dependent on any one AWS region. It failed over seamlessly to Amazon-West, and data-gathering continued normally. According to Brian Edgar, Group Program Manager at The Nielsen Company’s Telecom Practice: “While the outage at Amazon East was certainly bad news for Amazon and many of its clients, it was a great example of why the technology choices DASHbay recommended for us were ideal for this application. Our application was architected for geographic redundancy and fully leveraged the cloud model with dynamic DNS routing and load balancing using servers in multiple zones and regions. Our mission-critical, highly-available application experienced no outage at all. The Amazon regional failure proves we did the right thing.”

Another is a data acquisition app built for our client Credit.com. Unstructured data is gathered and marshaled into transaction reports. Data loss can directly impact Credit.com’s ability to monitor its own financial performance. This app was deployed only in the Amazon-East region, and was not available for over 24 hours. However, we anticipated the possibility of an outage, and had offshore staff in Nagpur, India, trained to perform manual workarounds for as long as necessary. These manual processes kicked in, and kept the data flowing. According to Credit.com’s CEO Ian Cohen, “We’ve been working with Dashbay for the last year and were really pleased with the measures they put in place to provide redundancies for our data acquisition applications. They positioned an offshore failsafe that allowed us to operate without interruption.”

What’s the message here? I think it’s this: data centers can fail! Design operational processes and real-time architectures with fail-over in mind. We used a variety of approaches, from human-intensive procedures that were nevertheless ready to go, to automated failover.

Which risk-mitigation strategies are right for a particular app? That depends on factors such as the volume of data and the tolerable latency of gathering and moving that data. We’re committed to thinking those factors through with our clients, and designing applications and processes with failover in mind.

One more important thing: let’s all keep the mindset of learning from mistakes, and if necessary changing architectures and backup procedures to keep our businesses running.

Terry Joyce, DASHbay founder