I wrote earlier that data is structured in multiple forms. In fact, it is the structure of data that allows applications to handle it “automatically” – as an automaton, i.e., programmatically – rather than relying on humans to handle it “semantically”.
Thus a search engine can search for words, propose completion of partially typed words, do spell checking, and suggest grammar corrections “automatically”.
In the last 30 years, we’ve built specialized systems to handle each data structure differently at scale. We index a large corpus of documents in a dedicated search engine for searches, we arrange lots of words in a publishing framework to compose documents, we store relational data in a RDBMS to do reporting, we store emails in an e-discovery platform to identify emails that satisfy a certain pattern, we build and store cubes in a MOLAP engine to do interactive analysis, and so on.
Each such system is a silo – it imposes a particular structure on big data, and then it leverages that structure to do its tasks efficiently at scale.
The silo approach imposes fragmentation of data assets. It is expensive to maintain these silos. It is inefficient for humans and programs to master these silos – they have to learn the nuances of each silo to become an expert in exploiting it. As a result, we have all kinds of data administrators – a cube expert, a text expert, a spreadsheet expert, and so on.
The state of data fragmentation reminds me of the “dedicated function machines” that pre-dated the “Personal Computer”. We used to have electronic type-writers that would create documents, calculators that would calculate formulae, fax machines that would transmit documents, even tax machines that would calculate taxes. All of these machines were booted to relic-status at a museum by a general-purpose computer – the functions were ported on top of its computing framework and the data was stored in its file system. The unity of all of these functions and its data on the general-purpose computer gave rise to “integration” benefits. It made tasks easier: we can now fill our tax forms in (structured form-based) PDF documents, do tax calculations, and file taxes by transmitting the document – all on one platform. Our productivity has gone up. Indeed, the assimilation of data is leading to net new tasks that were not possible before. We can let programs search for previous year’s filings, read the entries, and populate this year’s forms from previous year’s filing to minimize data-entry errors.
We have the same opportunity in front of us now in the field of big data. For too long, have we relegated functions that work on big data to isolated “dedicated function machines.” These dedicated function machines are bad because they are not “open.” Data in a search engine can only be “searched” – it cannot be analyzed for sentiments or plagiarism or edited to insert or remove references. The data is the same, but each of these tasks requires a “dedicated function machine.”
We have the option to build a general purpose machine for big data – a multi-structured big data platform – that allows multiple structures of data to co-exist on a single platform that is flexible enough to perform multiple functions on data.
Such a platform, for example, would allow us to analyze structured payments data to identify our valuable customers, interpret sentiments of calls they made to us, analyze the most common problem across negative sentiment interactions, and predict the loss in revenue that can be prevented by solving that problem and the cost of acquiring net new customers to overcome the losses. Without a multi-structure big data platform, the above workflow is a 12-18 month cycle performed by a cross-functional team of “dedicated function experts” (CFO group, Customer Support group, Products group, Marketing group) – a bureaucratic mess of project management that produces results too expensively, too infrequently and too inaccurately, making simplifying assumptions at each step as they cannot agree on even basic metrics.
An open “Multi-Structured Big Data Platform” would be hugely enabling and open up vast efficiency and functionality that we can’t imagine today.