For data lakes, as with any valuable enterprise data set, architecture is a requirement. But it is also a moving target, due to ongoing evolution.
The data lake has come a long way since its origins around 2015. Today it is a well-established design pattern and data architecture for profound applications in data warehousing, reporting, data science, and advanced analytics as well as operational environments for marketing, supply chain, and finance. Over the years, users’ expectations, best practices, and business use cases for the data lake have evolved, as have the available data platforms upon which a data lake may be deployed.
This evolution is forcing changes in how data lakes are designed, architected, and deployed. In short, TDWI sees many corporations, government agencies, and other user organizations modernizing their data lakes to adapt them to today’s data and business requirements instead of those of 2015. Similarly, «greenfield» data lakes today are quite different from early data lakes.
Many of the recent changes in the data lake are occurring at the architectural level. In particular, TDWI sees many organizations replatforming their data lakes as they abandon older database management systems and other data platforms in favor of modern ones. This forces changes in the systems architectures of a data lake, whether an old platform is replaced by a new one or left in place and augmented by a new one. Replatforming also leads to changes in a data lake’s data architecture when data is redistributed among the new mix of platforms or remodeled and improved during data migration.
Replatforming and other drivers for data lake architecture evolution take various forms:
Data lakes started on Hadoop but are migrating elsewhere.
In fact, the earliest data lakes were almost exclusively on Hadoop. The current wave of dissatisfaction with Hadoop is driving a number of lake migrations off of Hadoop. For example, after living with their lakes for a year or more, many users discover that key use cases demand more and better relational functionality than can be retrofitted onto Hadoop. In a related trend, many organizations proved the value of data lakes on premises and are now migrating to cloud data platforms for their relational functionality, elasticity, minimal administration, and cost control.
A modern data lake must serve a wider range of users and their needs.
The first users of data lakes were mostly data scientists and data analysts who program algorithm-based analytics for data mining, statistics, clustering, and machine learning. As lakes have become more multitenant (serving more user types and use cases), set-based analytics (reporting at scale, broad data exploration, self-service queries, and data prep) has arisen as a requirement for the lake — and that requires a relational database.
Cloud has recently become the preferred platform for data-driven applications.
Cloud is no longer just for operational applications. Many TDWI clients first proved the value of cloud as a general computing platform by adopting or upgrading to cloud-based applications deployed in the software-as-a-service (SaaS) model. Data warehousing, data lakes, reporting, and analytics are now aggressively adopting or migrating to cloud tools and data platforms. This is a normal maturity life cycle — many new technologies are first adopted for operational applications, then for data-driven, analytics applications.
New cloud data platforms are now fully proven.
The early adoption phase is over, spurring a rush of migrations to them for all kinds of data sets. As mentioned earlier, cloud data warehouses and other data platforms have the relational functionality that users need. In addition, they support the push-down execution of custom programming in Java, R, and Python. Early adopters have corroborated that the platforms perform and scale elastically, as advertised, while maintaining high availability and tight security. This gives more organizations the confidence they need to make their own commitments to cloud data platforms.
User best practices for data lakes are far more sophisticated today.
Early data lakes suffered abusive practices such as data dumping, neglect of data standards, and a disregard for compliance. Over time, lake users have corrected these poor practices. Furthermore, users have realized that the data lake — like any enterprise data set — benefits from more structure, quality, curation, and governance.
The catch is to make these improvements in moderation without harming the spirit of the data lake as a repository for massive volumes of raw source data fit for broad exploration, discovery, and many analytics approaches. It’s a bit of a balancing act, but data lake best practices are now established for maintaining detailed source data for discovery analytics while also providing cleansed and lightly standardized data for set-based analytics.
By: Philip Russom