So, we have a phrase we like to use around here borrowed from the legal academic world. Used to describe an action or conduct in analyzing a nuance in tort negligence, is the phrase “frolic and detour.” I am taking a bit of detour and frolicking in an increasingly noisy element of explaining the complexity of our work here. (The detour comes from the fact that as “Development Officer” my charge is ensuring the Foundation and projects are financed, backed, supported, and succeed in adoption. The frolic is in the form of commentary below about software development methodologies although I am not currently engaged or responsible for technical development outside of my contributions in UX/UI design.) Yet, I won’t attempt to deny that this post is also a bit of promotion for our stakeholders — elections IT officials who expect us to address their needs for formal requirements, specifications, benchmarks, and certification, while embracing the agility and speed of modern development methodologies.
This post was catalyzed by chit-chat at dinner last evening with an energetic technical talent who is jacked-up about the notion of elections technology being an open source infrastructure. Frankly, in 5 years we haven’t met anyone who wasn’t jacked-up about our cause, and their energy is typically around “damn, we can do this quick; let’s git ‘er done!” But it is about at this point where the discussion always seems to get a bit sideways. Let me explain.
I guess I am exposing a bit of old school here, but having had the formal training in computer systems science and engineering (years ago) I believe data modeling — especially for database-backed enterprise apps — is an absolute imperative priority. And the stuff of elections systems is serious technology, containing a significant degree of fault tolerance, integrity and verification assurance, and perhaps most important a sound data model. And modeling takes time and requires documentation, both of which are nearly antithetical in today’s pop culture of agile development.
Bear in mind, the TTV Project embraces agile methods for UX/UI development efforts. And there are a number of components in the TTV elections technology framework that do not require extensive up-front data modeling and can be developed purely in an iterative environment.
However, we claim that data modeling is critical for certain enterprise-grade elections applications because (as many seasoned architects have observed): [a] the data itself has meaning and value outside of the app that manipulates it, and [b] scalability requires a good DB design because you cannot just add in scalability later. The data model or DB design defines the structure of the database and the relationships between the data sets; it is, in essence the foundation on which the application(s) are built. A solid DB design is essential to achieve a scalable application. Which leads to my lingering question: How do agile development shops design a database?
I’ve heard the “Well, we start with a story...” approach. And when I ask those who I really respect as enterprise software architects with real DB design chops, who also respect and embrace agile methodologies, they tend to express reservations about the agile mindset being boorishly applied to truly scalable, enterprise grade relational DB design that results in a well performing application, and related data integrity.
Friends, I have no intention of hating on agile principles of lightweight development methods — they have an important role in today’s application software development space and an important role here at the Foundation, but at the same time, I want to try to explain why we cannot simply just “bang out” new elections apps for ballot marking, tabulation, or ballot design and generation in a series of sprints and scrums.
First, in all candor, I fear this confusion rests in the reality that fewer and fewer developers today have had a complete computer science education, and cannot really claim to be disciplined software engineers or architects. Many (not all) have just “hacked” with, and self-taught themselves, development tools because they built a web site or implemented a digital shopping bag for a friend (much like the well intentioned developer my wife and I met last evening).
Add in the fact, the formality and discipline of compiled code has given way to the rapid prototyping benefits of interpreted code. And in the processes of this new modern training in software development (almost exclusively for the sandbox of the web browser as the UX/UI vehicle) what has been forgotten is that data modeling exists not because it creates overhead and delays, but because it removes such impediments.
Look at this another way. I like to use building analogies — perhaps because I began my collegiate studies long ago in architectural engineering before realizing that computer graphics would replace drafting. There is a reason we spend weeks, sometimes months traveling by large holes in the ground with towers of re-bar, forms, and concrete pouring without any clue of what really will stand there once finished. And yet, later as the skyscraper takes form, the speed with which it comes together seems to accelerate almost weekly. Without that foundation carefully laid, the building cannot stand for any extended period of time, let alone bear the dynamic and static weights of its appointments, systems, and occupants. So too, is this the case with complex, highly scalable, fault tolerant enterprise software — without the foundation of a sold data model, the application(s) will never be sustainable.
I admit that I have been out of production grade software development (i.e., in the trenches coding, compiling; link, load, dealing with lint and running in debug mode) for years, but I can still climb on the bike and turn the pedals. The fact is, data flow and data model could not be more different. The former cannot exist without the latter. It was well understood and data modeling has demonstrated many times that one cannot create a data flow out of nothing. There has to be a base model as a foundation of one or more data flows, each mapping to its application. Yet in our discussion punctuated by a really nice wine and great food, this developer seemed to want to dismiss modeling as something that can be done later… perhaps like refactoring (!?)
I am beginning to believe this fixation of modern developers with “rapid” non-data-model development is misguided, if not dangerous for its latent time shifted costs.
Recently, a colleague at another Company was involved with the development of a system where no time whatsoever was spent on data model design. Indeed, the screens started appearing in record time. The UX/UI was far from complete, but usable. And the team was cheered as having achieved great “savings” in the development process. However, when it came time to expand and extend the app with additional requirements, the developers waffled and explained they would have to recode the app in order to meet the new process requirements. The data was unchanged, but processes were evolving. The balance of the project ground to a halt in the dismissal of the first team over arguments about why requirements planning up front should have been done, and they figured out who to hire in to solve it.
I read somewhere of another development project where the work was getting done in 2 week cycles. They were about 4 cycles away from finishing when on the tracker schedule a task called “concurrency” appeared for the next to last (penultimate) cycle. The project subsequently imploded because all of the code had to be refactored (a core entity actually was determined to be two entities.) Turns out that no upfront modeling led to this sequence of events, but unbelievably, the (agile) Development Firm working on the project, spun this as a “positive outcome;” that is they explained, “Hey, its a good thing we caught this a month before go-live.” Really? Why wasn’t that caught before that pungent smell of freshly cut code started wafting through the lab?
Spin doctoring notwithstanding, the scary thing to me is that performance and concurrency problems caused by a failure to understand the data are being caught far too late in the Agile development process, which makes it difficult if not impossible to make real improvements. In fact, I fear that many agile developers have the misguided principle that all data models should be:
create table DATA
Actually, we shouldn’t joke about this. That idea comes from a scary reality: a DBA (database architect) friend tells about a development team he is interacting with on an outsourced State I.T. project that has decided to migrate a legacy non-Oracle application to Oracle using precisely this approach. Data that had been stored as records in old ISAM type files, will be stored in Oracle as byte sequences in Blobs, with an added surrogate generated unique primary key. When he asked what’s the point of that approach, no one at the development shop could give him a reasonable answer other than “in the time frame we have, it works.” It begs the question: What do you call an Oracle Database where all the data in it is invisible to Oracle itself and cannot be accessed and manipulated directly using SQL? Or said differently, would you call a set of numbered binary records a “database,” or just “a collection of numbered binary records?”
In another example of the challenges of agile development in a database-driven app world, a DBA colleague describes being brought in on an emergency contract basis to an Agile project under development on top of Oracle, to deal with “performance problems” in the database. Turns out the developers were using Hibernate and apparently relied on it to create their tables on an as-needed basis, simply adding a table or a column in response to incoming user requirements and not worrying about the data model until it crawled out of the code and attacked them.
This sort of approach to app development is what I am beginning to see as “hit and run.” Sure, it has worked so far in the web app world of start-ups: get it up and running as fast as possible, then exit quickly and quietly before they can identify you as triggering the meltdown when scale and performance start to matter.
After chatting with this developer last evening (and listening to many others over recent months lament that we’re simply moving too slowly) I am starting to think of Agile development as a methodology of “do anything rather than nothing, regardless of whether its right.” And this may be to support the perception of rapid progress: “Look, we developed X components/screens/modules in the past week.” Whether any of this code will stand up to production performance environments is to be determined later.
Another Agile principle is of incremental development and delivery. It’s easy for a developer to strip out a piece of poorly performing code and replace it with a chunk that offers better or different capabilities. Unfortunately, you just cannot do this in a Database. For example: you cannot throw away old data in old tables and simply create new empty tables.
The TrustTheVote Project continues to need the kind of talent this person exhibited last evening at dinner. But her zeal aside (and obvious passion for the cause of open source in elections), and at the risk of running off the (Ruby) rails here, we simply cannot afford to have these problems happen with the TrustTheVote Project.
Agile methodologies will continue to have their place in our work, but we need to be guided by some emerging realities, and appreciate that for as fast as someone wants to crank out a poll book app or a ballot marking device, we cannot afford to short-cut simply for the sake of speed. Some may accuse me of being a waterfall Luddite in an agile world; however, I believe there has to be some way to mesh these things, even if it means requirements scrums, data modeling sprints, or animated data models.