Data Lakes: Safe Way to Swim in Big Data?

It has been more than five years since James Dixon of Pentaho coined the term “data lake.” His original post suggests, “If you think of a data mart as a store of bottled water – cleansed and packaged and structured for easy consumption – the data lake is a large body of water in a more natural state.” The analogy is a simple one, but in my experience talking with many end users there is still mystery surrounding the concept. In this post I’d like to clarify what a data lake is, review the reasons an organization might consider using one and the challenges they present, and outline some developments in software tools that support data lakes.

Data lakes offer a way to deal with big data. A data lake combines massive storage capabilities for any type of data in any format as well as processing power to transform and analyze the data. Often data lakes are implemented using Hadoop technology. Raw, detailed data from various sources is loaded into a single consolidated repository to enable analyses that look across any data available to the user. To understand why data lakes have become popular it’s helpful to contrast this approach with the enterprise data warehouse (EDW). In some ways an EDW is similar to a data lake. Both act as a centralized repository for information from across an organization. However, the data loaded into an EDW is generally summarized, structured data. EDWs are typically based on relational database technologies, which are designed to deal with structured information. And while advances have been made in the scalability of relational databases, they are generally not as scalable as Hadoop. Because these technologies are not as scalable, it is not practical to store all the raw data that come in to the organization. Hence there is a need for summarization. In contrast, a data lake contains the most granular data generated across the organization. The data may be structured information, such as sales transaction data, or unstructured information, such as email exchanged in customer service interactions.

Hadoop is often used with data lakes because it can store and manage large volumes of both structured and unstructured data for subsequent analytic processing. The advent of Hadoop made it feasible and more affordable to store much larger volumes of information, and organizations began collecting and storing the raw detail from various systems throughout the organization. Hadoop has also become a repository for unstructured information such as social media and semistructured data such as log files. In fact, our benchmark research shows that social media data is the second-most important source of external information used in big data analytics.

In addition to handling larger volumes and more varieties of information, data lakes enable faster access to information as it is generated. Since data is gathered in its raw form, no preprocessing is needed. Therefore, information can be added to the data lake as soon as it is generated and collected. This approach has caused some controversy with many industry analysts and even vendors to raise concerns about data lakes turning into data swamps. In general, the concerns about data lakes becoming data swamps center around the lack of governance of the data in a data lake, an appropriate topic here. These collections of data should be governed like any other set of information assets within an organization. The challenge was that most of the governance tools and technologies had been developed for relational databases and EDWs. In essence, the big data technologies used for data lakes had gotten ahead of themselves, without incorporating all the features needed to support enterprise deployments.

Another, perhaps more minor controversy centers around terminology. I raise this issue so that, regardless of the terminology a vendor chooses, you can recognize data lakes and be aware of the challenges. Cloudera uses the term Enterprise Data Hub to represent essentially the same concept as a data lake. Hortonworks embraces the data lake terminology as evidenced in this post. IBM acknowledges the value of data lakes as well as its challenges in this post, but Jim Kobielus, IBM’s Big Data Evangelist, questioned the terminology in a more recent post on LinkedIn, and the term “data lake” is not featured prominently on IBM’s website.

Despite the controversy and challenges, data lakes are continuing to grow in popularity. They provide important capabilities for data science. First, they contain the detailed data necessary to perform predictive analytics. Second, they allow efficient access to unstructured data such as social media or other text from customer interactions. For business this information can create a more complete profile of customers and their behavior. Data lakes also make data available sooner than it might be available in a conventional EDW architecture. Our data and analytics in the cloud benchmark research shows that one in five (21%) organizations are integrating their data in real time. The research also shows that those who integrate their data more often are more satisfied and more confident in their results. Granted, a data lake contains raw information, and it may require more analysis or manipulation since the data is not yet cleansed, but time is money and faster access can often lead to new revenue opportunities. Half the participants in our predictive analytics benchmark research said they have created new revenue opportunities with their analytics.

Cognizant of the lack of governance and management tools some organizations hesitated to adopt data lakes, while others went ahead. Vendors in this space have advanced their capabilities in the meantime. Some, such as Informatica, are bringing data governance capabilities from the EDW world to data lakes. I wrote about the most recent release of Informatica’s big data capabilities, which it calls Intelligent Data Lake. Other vendors are bringing their EDW capabilities to data lakes as well. Information Builders and Teradata both made data lake announcements this spring. In addition, a new category of vendors is emerging focused specifically on data lakes. Podium Data says it provides an “enterprise data lake management platform,” Zaloni calls itself “the data lake company,” and Waterline Data draws its name “from the metaphor of a data lake where the data is hidden below the waterline.”

Is it safe to jump in? Well, just like you shouldn’t jump into a lake without knowing how to swim, you shouldn’t jump into a data lake without plans for managing and governing the information in it. Data lakes can provide unique opportunities to take advantage of big data and create new revenue opportunities. With the right tools and training, it might be worth testing the water.

Regards,

David Menninger

SVP & Research Director

Research Uncovers Keys to Using Predictive Analytics

As a technology, predictive analytics has existed for years, but adoption has not been widespread among businesses. In our recent benchmark research on business analytics among more than 2,600 organizations, predictive analytics ranked only 10th among technologies they use to gene­rate analytics, and only one in eight of those companies use it. Predictive analytics has been costly to acquire, and while enterprises in a few vertical industries and specific lines of business have been willing to invest large sums in it, they constitute only a fraction of the organizations that could benefit from them. Ventana Research has just completed a benchmark re­search project to learn about how the organizations that have adopted predictive analytics are using it and to ac­quire real-world information about their levels of maturity, trends and best practices. In this post I want to share some of the key findings from our research.

As I have noted, varieties of predictive analytics are on the rise. The huge volumes of data that organizations accumulate are driving some of this interest. Our Hadoop research highlights the intersection of this big data and predictive analytics: More than two-thirds (69%) of Hadoop users perform advanced analytics such as data mining. Regardless of the reasons for the rise, our new research confirms the importance of predictive analytics. Participants overwhelmingly reported that these capabilities are important or very important to their organization (86%) and that they plan to deploy more predictive analytics (94%). One reason for the importance assigned to predictive analytics is that most organizations apply it to core functions that produce revenue. Marketing and sales are the most common of those. The top five sources of data tapped for predictive analytics also relate directly to revenue: customer, marketing, product, sales and financial.

Although participants are using predictive analytics for important purposes and are generally positive about the experience, they do not minimize its complexities. While now usable by more types of people, this technology still requires special skills to design and deploy, and in half of organizations the users of it don’t have them. Having worked for two different vendors in the predictive analytics space, I personally can testify that the mathematics of it requires special training. Our research bears this out. For example, 58 percent don’t understand the mathematics required. Although not a math major, I had always been analytically oriented, but to get involved in predictive analytics I had to learn new concepts or new ways to apply concepts I knew.

Organizations can overcome these issues with training and support. Unfortunately, most are not doing an adequate job in these areas. Not half (44%) said their training in predictive analytics concepts and techniques is adequate, and fewer than one-fourth (24%) provide adequate help desk resources. These are important places to invest because organizations that do an adequate job in these two areas have the highest levels of satisfaction with their use of predictive analytics; 89% of them are satisfied vs. 66% overall. But we note that product training is not the most important type. That also correlated to higher levels of satisfaction, but training in concepts and the application of those concepts to business problems showed stronger correlation.

Timeliness of results also has an impact on satisfaction. Organizations that use real-time scoring of records occasionally or regularly are more satisfied than those that use real-time scoring infrequently or not at all. Our research also shows that organizations need to update their models more frequently. Almost four in 10 update their models quarterly or less frequently, and they are less satisfied with their predictive analytics projects than those who update more frequently. In some ways model updates represent the “last mile” of the predictive analytics process. To be fully effective, organizations need to build predictive analytics into ongoing business processes so the results can be used in real time. Using models that aren’t up to date undermines the whole effort.

Thanks to our sponsors, IBM and Alpine Data Labs, for helping to make this research available. And thanks to our media sponsors, Information Management, KD Nuggets and TechTarget, for helping in gaining participants and promoting the research and educating the market. I encourage you to explore these results in more detail to help ensure your organization maximizes the value of its predictive analytics efforts.

Regards,

David Menninger – VP & Research Director