We hear this a lot. We hear it from very smart people. Just the other day we heard someone say they had tried RDF twice at previous companies and it failed both times. (RDF stands for Resource Description Framework, which is an open standard underlying many graph databases). It’s hard to convince someone like that that they should try again.
That particular refrain was from someone who was a Neo4j user (the leading contender in the LPG (Labeled Property Graph) camp). We hear the same thing from any of three camps: the relational camp, the JSON camp, and the aforementioned LPG camp.
Each has a different reason for believing this RDF stuff is just too hard. Convincing those who’ve encountered setbacks to give RDF another shot is also challenging. In this article, I’ll explore the nuances of RDF, shedding light on challenges and strengths in the context of enterprise integration and application development.
For a lot of problems, the two-dimensional world of relational tables is appealing. Once you know the column headers, you pretty much know how to get to everything. It’s not quite one form per table, but it isn’t wildly off from that. You don’t have to worry about some of the rows having additional columns, you don’t have to worry about some cells being arrays or having additional depth. Everything is just flat, two-dimensional tables. Most reporting is just a few joins away.
JSON is a bit more interesting. At some point you discover, or decree if you’re building it, that your dataset has a structure. Not a two-dimensional structure as in relational, but more of a tree-like structure. More specifically, it’s all about determining if this is an array of dictionaries or a dictionary of arrays. Or a dictionary of dictionaries. Or an array of arrays. Or any deeply nested combination of these simple structures. Are the keys static — that is, can they be known specifically at coding time, or are they derived dynamically from the data itself? Frankly, this can get complex, but at least it’s only locally complex. A lot of JSON programming is about turning someone else’s structure into a structure that suits the problem at hand.
One way to think of LPG is JSON on top of a graph database. It has a lot of the flexibility of JSON coupled with the flexibility of graph traversals and graph analytics. It solves problems difficult to solve with relational or just JSON and has beautiful graphics out of the box. Maybe link to your blog post about LPGs as training wheels?
Each of these approaches can solve a wide range of problems. Indeed, almost all applications use one of those three approaches to structure the data they consume.
And I have to admit, I’ve seen a lot of very impressive Neo4j applications lately. Every once in a while, I question myself and wonder aloud if we should be using Neo4j. Not because RDF is too hard but because we’ve mastered it and have many successful implementations running at client sites and internally. But maybe, if it is really easier, we should switch. And maybe it just isn’t worth disagreeing with our prospects.
Enterprise Integration is Hard
Then it struck me. The core question isn’t really “RDF v LPG (or JSON or relational),” it’s “application development v. enterprise integration.”
I’ve heard Jans Aasman, CEO of Franz, the creators of AllegroGraph, make this observation more than once: “Most application developers have dedicated approximately 0 of their neurons contemplating how what they are working on is going to fit in with the rest of their enterprise, whereas people who are deeply into RDF may spend upwards of half their mental cycles thinking of how the task and data at hand fits into the overall enterprise model.”
That, I think, is the nub of the issue. If you are not concerned with enterprise integration, then maybe those features that scratch the itches that enterprise integration creates are not worth the added hassle.
Let’s take a look at the aspects of enterprise integration that are inherently hard, why RDF might be the right tool for the job, and why it might be overkill for traditional application development.
One of the biggest issues dealing with enterprise integration is complexity. Most mid to large enterprises harbor thousands of applications. Each application has thousands of concepts (tables and columns or classes and attributes or forms and fields) that must be learned to become competent either in using the application and/or in debugging and extending it. No two application data models are alike. Even two applications in the same domain (e.g., two inventory systems) will have comically different terms, structures, and even levels of abstraction.
Each application is at about the complexity horizon that most mere mortals can handle. The combination of all those models is far beyond the ability of individuals to grasp.
Enterprise Resource Planning applications and Enterprise Data Modeling projects have shone a light on how complex it can get to attempt to model all an enterprise’s data. ERP systems now have tens of thousands of tables, and hundreds of thousands of columns. Enterprise Data Modeling fell into the same trap. Most efforts attempted to describe the union of all the application models that were in use. The complexity made them unusable.
What few who are focused on solving a point solution are aware of, is that there is a single simple model at the heart of every enterprise. It is simple enough that motivated analysts and developers can get their heads around it in a finite amount of time. And it can be mapped to the existing complex schemas in a lossless fashion.
The ability to posit these simple models is enabled by RDF (and its bigger brothers OWL and SHACL). RDF doesn’t guarantee you’ll create a simple or understandable model (there are plenty of counterexamples out there) but it at least makes the problem tractable.
An RDF based system is mostly structure-free, so we don’t have to be concerned with structural disparities between systems, but we do need a way to share concepts. We need to have a way to know that “employee,” “worker,” “user,” and “operator” are all referring to the same concept. Or if they aren’t, in what ways they overlap.
In an RDF-based system we spend a great deal of time understanding the concepts that are being used in all the application systems, and then creating a way that both the meaning and the identity of the concept can be easily shared across the enterprise. And that the map between the existing application schema elements and the shared concepts are also well known and findable.
One mechanism that helps with this is the idea that concepts have global identifiers (URIs /IRIs) that can be resolved. You don’t need to know which application defined a concept; the domain name (and therefore the source authority) is right there in the identifier and can be used much like a URL to surface everything known about the concept. This is an important feature of enterprise integration.
Instance Level Integration
It’s not just the concepts. All the instances referred to in application systems have identifiers. But often the identifiers are local. That is, “007” refers to James Bond in the Secret Agent table, but it refers to “Ham Sandwich” in the company cafeteria system.
The fact that systems have been creating identity aliases for decades is another problem that needs to be addressed at the enterprise level. The solution is not to attempt, as many have in the past, to jam a “universal identifier” into the thousands of affected systems. It is too much work, and they can’t handle it anyway. Plus, there are many identity problems that were unpredicted at the time their systems were built (who imagined that some of our vendors would also become customers?) and are even harder to resolve.
The solution involves a bit of entity resolution, coupled with a flexible data structure that can accommodate multiple identifiers without getting confused.
Data Warehouse, Data Lake, and Data Catalog all in One
Three solutions have been mooted over the last three decades to partially solve the enterprise integration problem: data warehouses, lakes, and catalogs. Data warehouses acknowledged that data has become balkanized. By conforming it to a shared dimensional model and co-locating the data, we could get combined reporting. But the data warehouse was lacking on many fronts: it only had a fraction of the enterprise’s data, it was structured in a way that wouldn’t allow transactional updates, and it was completely dependent on the legacy systems that fed it. Plus, it was a lot of work.
The data lake approach said co-location is good, let’s just put everything in one place and let the consumers sort it out. They’re still trying to sort it out.
Finally, the data catalog approach said: don’t try to co-locate the data, just create a catalog of it so consumers can find it when they need it.
The RDF model allows us to mix and match the best of all three approaches. We can conform some of the enterprise data (we usually recommend all the entity data such as MDM and the like, as well as some of the key transactional data). An RDF catalog, coupled with an R2RML or RML style map, will not only allow a consumer to find data sets of interest, in many cases they can be accessed using the same query language as the core graph. This ends up being a great solution for things like IoT, where there are great volumes of data that only need to be accessed on an exception basis.
We hinted at query federation in the above paragraph. The fact that query federation is built into the spec (of SPARQL, which is the query language of choice for RDF, and also doubles as a protocol for federation) allows data to be merged at query time, across different database instances, different vendors and even different types of databases (with real time mapping, relational and document databases can be federated into SPARQL queries).
Where RDF Might Be Overkill
The ability to aid enterprise integration comes at a cost. Making sure you have valid, resolvable identifiers is a lot of work. Harmonizing your data model with someone else’s is also a lot of work. Thinking primarily in graphs is a paradigm shift. Anticipating and dealing with the flexibility of schema-later modeling adds a lot of overhead. Dealing with the oddities of open world reasoning is a major brain breaker.
If you don’t have to deal with the complexities of enterprise integration, and you are consumed by solving the problem at hand, then maybe the added complexity of RDF is not for you.
But before you believe I’ve just given you a free pass consider this: half of all the work in most IT shops is putting back together data that was implemented by people who believed they were solving a standalone problem.
There are many aspects of the enterprise integration problem that lend themselves to RDF-based solutions. The very features that help at the enterprise integration level may indeed get in the way at the point solution level.
And yes, it would in theory be possible to graft solutions to each of the above problems (and more, including provenance and fine-grained authorization) onto relational, JSON or LPG. But it’s a lot of work and would just be reimplementing the very features that developers in these camps find so difficult.
If you are attempting to tackle enterprise integration
issues, we strongly encourage you to consider RDF. There is a bit of a step
function to learn it and apply it well, but we think it’s the right tool for