Seven Things to Consider When Choosing the Best Application Development Framework for Your Business

Once enterprises decide on proceeding with developing applications on Hadoop, they have to standardize on a specific application development framework. With this, enterprises can promote the reuse of code, set development best practices, comply with regulations, enable consistent quality, etc.

I've seen all kinds of projects across different of industries. I've seen smart people go with the newest technology, even if it's unproven, and have paid the price when they find it's not enterprise ready. I've also seen projects fail because they overemphasized the niche capabilities of the platform instead of the basics, such as ease of integration to existing systems.

Because of these experiences, I'm often asked to collaborate with enterprises to help them select their application development framework. As a result, this article will explore what you should consider when evaluating platforms for your current and future needs.

Three Things to Consider About Framework Architectures
The technology landscape is a moving target - and not stopping anytime soon
The technology landscape in Hadoop is fast evolving - new compute fabrics are continually introduced, some of which (in the future) may be a fit to run your application. As you evolve your Big Data applications and analytics, it's unlikely that you will have a single execution engine. For example, it's generally a bad idea to code directly on fabric-specific APIs (such as ones exposed by Spark) because it leads to building stove-pipe solutions and a loss in your ability to migrate applications to a newer and a better computational fabric in the future: Storm, Tez, Spark, Flink, etc.

Some compute fabrics will be better suited for analytics applications where others are better for high-volume data processing or complex logarithmic algorithms. Vendor lock-in to proprietary technologies rarely provide the long-term advantages touted by these proprietary technologies. Your application framework should prevent you from getting locked into a particular data storage format, Hadoop distributions, compute fabrics, programming languages and hardware.

Portability is increasingly more important
Ideally your application development platform provides an abstraction layer from the underlying compute framework and integration APIs to support a variety of languages. This is critical to providing autonomy to alter your architecture in the future as the ecosystem evolves.

For example, you should be able to write and test your application on Map Reduce and, with minor changes, run it on Tez, Flink, etc. Similarly, if you want to push your final data product to Elasticsearch instead of HBase, you should not have to change any application code. Equally, you should be able to move to a different distribution portability just as easily.

Portability also means extending the life of your legacy application code (on Hadoop) by enabling them to run within newer compute fabrics. You have invested a lot of time and resources building applications, which in a couple years will be legacy application code. Bottom line, your application architecture should support the integration of existing applications but in legacy frameworks without requiring you to rewrite them so you can leverage your past investment.

Ease of integration to existing infrastructure
In first-generation Hadoop applications, integration costs account for 60 percent to 90 percent of the total development costs for an implementation. It's still surprising how many people fail to address, in the early stages of projects, minimizing cost and risk of migrating raw data to Hadoop or pushing curated data products, developed on Hadoop, to the final destination IM systems. This oversight often results in project failure or the need for expensive rework as the project tries to scale. This happens because very often engineers make technical decisions based on what is required to be completed in a prototype, and not what is required in the final delivery.

Having a framework that gracefully supports your applications as they mature provides architects with the peace of mind that they do not need to validate all the possible functional needs upfront. What is required is a framework that let's you do simple things simply, but also scales to meet your more complex future needs.

Two Things to Consider About Application Development in These Frameworks
Reuse of existing developer skills and speed of development
Your development framework should expose multiple development interfaces not only in SQL but in Java, Scala, Cascalog, etc., in the same underlying framework so developers are free to use the interface that best matches their needs and skills.

In addition to development freedom, you need to consider where development is done. Developing directly on the cluster is inefficient. Even for running a job with a small data set, there is measurable latency involved in scheduling the job. In addition, debugging jobs running directly on the cluster becomes more difficult. What is required is to allow developers to perform iteration testing of their applications on a local development environment and, when the application is stable, deploy it on the cluster to run as a MapReduce application. Local development capabilities and the ability to deploy on different Hadoop compute engines without code changes will save days of development.

Active enforcement of DevOps best practices
Software engineering best practices promote the reuse of code, reduce development costs, promote consistency in quality and make the code maintainable. That is why we have well-defined philosophies on how to do continuous integration, test-driven development and defensive programming when creating software products. These software practices provide the guidelines that keep your applications healthy.

Alas, all these best practices are discarded when you use a framework that does not expose a programmatic interface. No matter how great a SQL or a GUI-based tool is, it is not possible to apply such practices in these frameworks. As a result, it is inevitable that either the code becomes unmaintainable, or developers refrain from leveraging the full power of the underlying platform.

Debugging in GUI and SQL-based frameworks in notoriously difficult - even more when you have implemented Java code residing in isolated UDF containers. Imagine stepping through the code in an application partly written in a 3GL language and part in Java - many times it's not possible. It may seem like a small thing, but it matters if it reduces the odds that your application will fail in mission-critical scenarios.

Two Things to Consider About Frameworks TCO and Enterprise Worthiness
Don't be fooled by claims of enterprise readiness - proof is required
Many of the popular Hadoop projects are straight out of university projects run by graduate students. In other cases, enterprises are now considering running Hadoop-related technologies previously proven only in technology companies employing tens of thousands of advanced computer scientists. New application frameworks should be evaluated as to whether they can fit within the operating model of your enterprise.

To begin, can your technical delivery organization quickly pick up the technology required to implement your applications on the framework? Do they have to learn a new programming language like Scala (good for people brushing up their resumes, but hardly a practical choice when a ten-thousand strong organization needs to be trained). How easy is it to learn the new technology? It is easy to be fooled by the guise that UI-based tools are always easy. In one training session for a GUI-based ETL tool, I once saw instructions that involved clicking on 12 different dialog boxes (try remembering the steps!) in what could have been done in a single API call - remember, simple things should be done simply without compromising your ability to do real-world things later.

Next, one of the biggest technical fallacies in the world of Big Data is that the "Big Data" technologies come tuned for scale. Scale can be the number of nodes in your cluster, the size of the data, the number of fields in your feed or the size of the file. However, even popular frameworks and tools fail for common use cases. Your application should not require a rewrite if the size of the cluster or the data changes. The real test of a framework is not just in the attractiveness of its APIs, but in its ability to demonstrate scale, a capability that takes years to develop.

Your aim in selecting the right framework to develop applications is to ensure that the applications can be successfully deployed in production. A production-grade application not only runs with consistent behavior, but also supports the operational requirements that demonstrate compliance. Can your framework support version management, lineage, detection of outlier behavior, future capacity requirements, service-level agreement (SLA) management and performance bottleneck analysis?

Start-to-finish operational readiness
When choosing an application development framework, be sure to select one that helps you develop applications through the entire lifecycle of data - ensure you can take your data from start to a finished data product on one framework, enabling better engineering, governance and quicker time-to-market. With GUI-only tools, developing applications demanding more than just simple join and sort operations requires assembling modules outside the framework, making developing production-grade applications more challenging.

Monitoring how an application is performing is also critical to delivering a responsive Hadoop infrastructure. Ensure your chosen framework enables complete application execution monitoring in real-time, which shows the current operating state so that you can easily identify when an application fails or is encountering performance issues preventing it from executing. For many using GUI-based ETL tools, this type of insight is impossible to achieve or requires building and maintaining separate monitoring applications.

Selecting a framework is a decision that will be continually validated by all the developers developing applications, operators supporting them, regulators validating compliance and executives justifying ROIs on their programs. There are places to be a visionary, but don't be the first person jumping in to vouch for a new technology. Similarly, do not be among the hundreds acting on a hyped technology that still has results to deliver.