Brooks asserts that:
There is no single development, in either technology or management technique, which by itself promises even one order of-magnitude improvement within a decade in productivity, in reliability, in simplicity.
Brooks is primarily concerned with programmer productivity, which is a critical topic as software is becoming far more complex and therefore the number of ways in which it can fail is also increasing dramatically. We’re far past the point where a single person can understand the whole environment in which they work.
Programming, Brooks argues, can be separated into distinct tasks. The first is dealing with incidental complexity, and the second managing essential complexity. We’d like to spend as much time as possible on the essential tasks, than on the incidental.
Software projects will always be complex, as what makes software useful is it’s complexity. Brooks argues that:
The essence of a software entity is a construct of interlocking concepts, data sets, relationships among data items, algorithms, and invocations of functions. The essence is abstract … but is nonetheless highly precise and richly detailed.
Does it have to be hard?
The hard part about software is figuring out how these concepts relate, not the actual act of programming. It’s fundamentally about getting the relationships between complex items right. This is what Brooks considers to be the essential acts of programming.
Brooks focuses on a few aspects that make programming especially difficult.
Brooks starts by asserting that software systems are far more complex than any other human system. This is compounded by our inability to see their actions, or understand the system visually. Common programming principles like DRY focus on code reuse, this means that each item in a software system is unique.
In a physical system, repetition of common parts is far more common, and therefore results in a system that is much easier to understand despite its high number of moving parts. Take a car for example, say a car is composed of 100,000 parts. We can safely ignore 2/5 of those, due to repetition from the left side and the right. Furthermore, common items like bolts, or bushings don’t really need to be understood, so perhaps we can remove another 1/3. This brings our total part count to around 40,000. And that is of course a rough estimation.
usr directory contains 18,843 directories, and 237,244 files.
That isn’t even the count of total functions that exist on my computer, which is
probably at least 10 times the number of files, so around 2,372,440. That’s a
whole lot of functions, my guess is there are probably a bit more than that.
If we go a step further an look at the total number of files on my computer to allow it to “do its thing”, you get a mind boggling number of files, 2,399,961 to be exact on my machine. Though this includes data files, so it’s not really a fair estimate, but you get the idea.
Software is really complex, as things increase in complexity and also size, it becomes increasingly more difficult for people to understand the system, and talk about it with any authority. The possible state space for the system also increases dramatically. If a system have 500,000 possible states, you can’t really prove all of them to be safe, much to the lament of Dijkstra and Lamport. Brooks wonders how we will combat complexity going forward.
Unlike physics software objects have no fundamental principles upon which they can rest. And therefore, each software artifact must be evaluated in isolation. This makes each evaluation many times more difficult, as you cannot treat any two functions as truly equivalent until they have both been thoroughly investigated. As you can imagine, this slows things down enormously.
The fact that software is adapted throughout its life time is another source of complexity. Most physical products, argues Brooks, are not changed after manufacturing. Software, on the other hand, is guaranteed to require patches, and updates on a somewhat regular basis. Software is also constantly being pushed towards the edge of its use case, as popular software is often used in domains it was not explicitly designed for. Thus, developers feel the desire to support that additional functionality, and the software is no longer tailored towards a very specific use case.
In some ways this is what’s nice about the Unix Philosophy.
- Programs should do one thing and do it well.
- Expect the output of every program to become the input to another, as yet unknown, program.
This principles allow for small utilities to be pulled together through piping the output of one into the input of another.
Software is also hard because we can’t see it. We have to manipulate everything inside our own heads, which leads to mistakes and conceptual errors. There is also no easy way to represent the content of a program in space. Object Oriented Languages can have UML Diagrams but these don’t really solve the problem, as they’re static and focus on class design, not how the program interacts dynamically. I’ve been working towards a better way to visualize outputs of functions called vtest, but visualizing programs is a hard problem.
Most gains have come from removing artificial barriers, and by their very nature these approaches limit themselves.
High Level Languages
High level languages abstract away repeated patterns. Chances are you’re going to have to read in a file more than once, this operation is going to be very very similar each time you want to execute that procedure. In this case you should not have to rewrite a bespoke function to accomplish an essential repetitive task.
Similarly the language allows the best version of each pattern to be used by everyone without any additional cost. Most incidental complexity comes in the form of rewriting the same patterns, and dealing with solved problems. This is a huge benefit, and increases productivity, according to Brooks estimation, by a factor of 5.
Time-sharing is not appreciated as much as it was during its introduction in the early 1960s. Multics was the first of such systems, and if you’re not familiar with the concept Computerphile has an excellent video on the topic: Mainframes and the Unix Revolution.
Time-Sharing essentially allows for multiple users to get back results from the programs in minutes rather than days. This delay time resulted in quicker iteration for programers, and therefore increased efficiency. Currently of course, the response time is nearly immediate, and yet programming is still difficult. Brooks expands on time sharing, noting that as the time decreases the amount of benefit also decreases.
Unified programming environments
Unified programming environments like Unix and Interlisp, provided consistent metaphors and support for common tasks. This results in decreased computational overhead for the programmer, and results in programs that can easily be composed.
Brooks contemplates future ways in which we can address common issues, and considers if those solutions address incidental or essential complexity. I’m not going to go into all these approaches, but if one piques your interest take a look at the full text.
|OOP||Automatic Programming||Environments and Tools|
|Ada||Expert Systems||Program Verification|
Attacks on Essential Complexity
Brooks believes that the conceptual aspects of programming are taking the majority of the time. Therefore, he focuses on methodologies to limit either the amount of time spent on a task, or how often it has to be accomplished.
Buy not Build
Brooks notes that ”The most radical possible solution for constructing software is to not construct it at all.” Therefore, when a simple off the shelf solution exists, use it! Even if it’s not done exactly the way you would like, its better to use a well tested, proven solution the embarking to reinvent the wheel. Brooks, I’m sure, would look at the FOSS solutions we have today and be delighted in the amount of code reuse.
Brooks elaborates that most use cases are not significantly different from each other, resulting in generic products that each user can integrate into their workflow. The key is keeping these products small and modular, which allows for much easier composition.
One of the most difficult tasks when working on a software project is deciding what exactly to build, and how the user interacts with it. It’s much easier to start small in this case, and get feedback quickly.
Rapid prototyping has also become more and more prevalent, as programmers and Designers have warmed up to the idea of showing people Minimum Viable Products, that the users can interact with. This gives excellent feedback to keep the software in light with the goals of the users.
Incremental Growing, not building.
Brooks argues that software should be grown organically.
I still remember the jolt I felt in 1958 when I first heard a friend talk about building a program, as opposed to writing one. In a flash he broadened my whole view of the software process.
Brooks however thinks the building metaphor is not well equipped to handle the current projects we’re developing. Instead of building, which requires adequate plans and foresight, we should focus on growing a program organically. Once even a very simple program is up and running, developers are much more enthusiastic about the progress.
In order to develop better systems, we must have better designers. It’s not surprising that great designers produce software that is elegant, simple and cohesive. Great designers have a vision regarding what computation should look like. Take this table from No Silver Bullet showing languages that were designed vs those done by committee.
Languages with great designers have a certain clarity to them, its something we all should aim to have within our own programs.
We ought to focus on making simple, elegant pieces of software that can compose easily with other systems when we have to. If a solution exists, and it satisfies your need then use it!