A quick heads-up on the Atlantic article

The challenges of software development are hitting the mainstream! A long, detailed article was recently published in The Atlantic; it is maybe a bit slow-paced for my taste, but I still recommend reading it. I would summarise it as follows:

Software is getting more and more complex (while at the same time being responsible for more and more critical aspects of systems and society). The methods used by developers in industry are no longer able to deal with the complexity, leading to potentially really dangerous failures. Developers need to change the way they write software. Two basic paths forward exist: the first one is making (the representations of) programs less complex and more easily understandable, effectively reducing their accidental complexity. Models, DSLs, and code generation are the useful here, as are ideas from live programming that reduce the gap between the program code and its execution through simulation, realtime feedback and visualisations. The second one, for the essential complexity that cannot be reduced with the first, formal methods, such as model checking, SMT solving and proof assistants should be more heavily used.

I agree with the general premise of the article, both the diagnosis of increasing complexity and the approaches to address them. Models and DSLs are my bread-and-butter work. The synergies with formal methods are a very interesting field which we have explored over the last few years. Realtime feedback, visualisations and simulations have “saved the day” in several of our DSL projects, and we are putting more and more work into this aspect of languages and tools. For example, in KernelF, our embeddable functional language, we have an in-IDE interpreter to run tests, we overlay execution values over the source in a tracer/debugger, and we are working on a reactive framework for incrementally updating computed values in the IDE.


There are a few places where I would be a bit more differentiated. For example, the promises of live programming, inspired by Bret Victorsvideo, haven’t quite panned out. All his examples, as well as the examples mentioned in the Atlantic article (image processing, web pages, animations, visualisations) are systems where the output is easily representable graphically. How would you do this for an airline reservation system? Or a medical diagnostics app? Those systems have such complex behaviors that you cannot “show” them. You have to define (all!) scenarios and illustrate those. This approach is called simulation, and is not really a new idea (though it is underused in the software space; we should look at systems engineers for inspiration). And yes, overlaying program values over the code and letting the user move back and forth in time is cool and useful, but it’s far from Victor’s vision.

Regarding formal methods, yes, I agree, it must play a bigger role in the future. However, as Benjamin Pierce explained in this omega tau episode, the effort to fully verify programs is still huge. And it is very much expert work, not easily accessible to mainstream developers (including yours truly!). This should not be an excuse for not at least using formal methods for the low-hanging fruits, for trying to educate future developers in formal methods, and for spending the effort on infrastructure components (such as operating systems, network stacks or web servers) to at least make those platforms more reliable. It will come, but it will take a bit longer than we might wish. In the meantime, by the way, it would be useful to at least test software systematically!

Even for the DSL and modelling stuff, we notice how hard it is to change the established habits of programmers. We have run many DSL projects where all the stakeholders involved in an initial proof-of-concept concluded that this is the way to the future, only to notice that the organisation as a whole is not willing to make the necessary process changes and supply the required education and training.

A final point I disagree with is this sentence:

The serious problems that have happened with software have to do with requirements, not coding errors.

While it is correct that a system that does the wrong thing is a problem (they provide some examples of well-known system failures in the article), coding errors, as they call them, are still serious issues. Many of the well-known security exploits are a result of low-level errors, often symbiotically related to quirks/design flaws of the programming language that is employed. It is completely mysterious to me why we still write many safety-critical systems in C, a language that is famous for its pitfalls.

In some sense, the summary of the summary would be that we as developers should rely more on computers and tools in the process of software development. Except for compilers, we tend to use tools for ancillary tasks such as build, packaging and executing tests or to help us get the structure of programs correct (IDEs, editors). The core activity of programming, and in particular, understanding of what the programs do, is still mostly happening in developers’ brains. This is kinda funny, because as software developers we often write tools that help other domains become more efficient (writers, presenters, technical designers, systems engineers). In this sense, we should focus more on ourselves.