However, complexity can be managed. Humans do this using abstraction as a tool. We divide the complex problem up into a sequence of several simpler states, and we find it easier to understand the simpler problems along with the sequence within which they lie.
Good software uses this same approach to reduce complex issues to a manageable process. A good tool makes the simple things easy and the complex thing possible, the design of the tool reflects the effort and work that the designer out in to u the problem they’re trying to solve, and to produce something that helps guide others along that self-same path of understanding, without them having to put in the same level of effort; it establishes the golden path through the marshes and bogs of difficulties that the problem domain throws up.
“Embracing complexity” is a measure of last resort, IMHO. It means the tool developer could not analyze the problem and come up with a good solution; it means “here, you figure it out”; it means giving up on one of the fundamental reasons for the tools existence.
Sometimes, embracing complexity and the ensuing struggle that this necessitates is simply what you have to do, but not often. Maybe, maybe this is one of those times, but I always start off with a critical eye when someone tells me that a complicated thing is “the only way it can be done”. Colour me sceptical.
Heres an amusing and simple example of manufactured complexity: https://github.com/EnterpriseQualityCoding/FizzBuzzEnterpris...
That said, there's no denying that there's probably a lot of unnecessary complexity in most software.
If anyone proved such a result, I'm sure they would be awarded a prize of some kind.
Introduce enterprise into any system and complexity goes over the roof ASAP - suddenly its not only your machine but TEAM of people, history, dashboards with metrics, logs, automation, security etc...
This one isn't even approaching any of it except dev side and CI.
However, any degree of functionality requires at least a matching degree of complexity. In that sense, some complexity is 'essential' and that complexity does need to live somewhere in your project.
There are always cross-cutting concerns that aren't easily transformed, or become that way due to code changes. I think this area is where managing complexity is most difficult.
Take for example this function which adds one to an integer:
function addOne(x: int) -> int {
return (x * 1) + 0 - x + x + 1;
}
function addOne(x: int) -> int {
return x + 1;
}
But you can't make the solution simpler than the problem. There's kind of a "by definition" aspect to this, so it's kinda circular, but it's still a valuable observation, because we also see this in code all the time. For instance, it plays a significant role in the leakiness of abstractions; that leakiness is typically the attempt to make a solution that is simpler than the entire problem space. You can be fine as long as your particular problem fits in the space of the leaky solutions, but once your problem pokes out of that you can be in big trouble.
> you can "create" complexity
AKA "the complicator's gloves" https://thedailywtf.com/articles/The_Complicator_0x27_s_Glov...
But I do agree with you in spirit of what you said, that there is some essential “hardness” of the problem that is part-and-parcel with the difficulty in stating or explaining the problem. I see a similar complexity in representing the problem in code or functions or algorithms etc. Then comes a related issue with reducibility; some concepts or problems cannot be broken down into smaller discrete parts. They are already as simple as they can be. This is ideal in a certain sense, in that the concepts can be explored with fewer assumptions. However when someone asks you to explain like they’re 5 years old, interpretation and perception starts to creep in through intuitions and abstractions.
Even that can be deceptive. For example, I can state very simply:
I want to move my house 3 inches to the left
It's about how the problem maps onto the constraints of implementation.
For example, newtons laws of motion, seemingly is a simple solution to finding a formal method of describing motion. It seems to fit the problem but there is no way to verify that it is the simplest solution to the problem and we don't even really know what "simple" is.
We know through experimentation that relativity fits observations more accurately but even with relativity we don't know if it's the most "simple" solution or even if its' fully correct.
Because you cannot know or ever know whether a solution is at its' simplest form, there is ALWAYS an opportunity for a reduction to be discovered.
You absolutely cannot say that "The solution can be as simple as the problem but no simpler" because we don't know what "simple" is and we don't know whether or not a "simpler" solution may be discovered in the future.
Any normal task has an inherent level of complexity (“add one” is a very simple task, so can be reduced down as you point out). “Invert a matrix” is somewhat less reducible. “Calculate a quaternion” even less so, etc.
We generally string dozens or hundreds of these simpler tasks into sequences to get what we want, and human understanding being what it is, there’s quite a difference in conceptual complexity between “fly the plane to [here]” and doing all that maths.
There’s probably a relationship to subitising here as well, the brain does seem to be able to process small groups of things, even if it knows those things represent larger concepts, better than larger groups of things. This is just speculation on my part though.
Am I missing something?
So I took 2 days, made a chart of every path and pattern of events (restarting a timer from a timer callback; having another time interval expire while processing the previous; restarting while processing and another interval expires; restarted timer expires before completing previous expiration and on and on). Then writing exhaustive code to deal with every case. Then running every degenerate case in test code until it survived for hours.
It never had to be addressed again. But it did have to be addressed. So many folks are unwilling to face the music with complexity.
Refactoring too soon, before understanding the entire system (and all the possible changes/additions/issues that could arise) is probably worse (time-wise) than thrashing for a bit.
Developers know about the Boiled Frog problem. But we sometimes don't know when we are the frog. We keep hill-climbing one solution because we are only trying to do one little thing so of course I should be able to do a small to medium thing to accomplish it.
We also send out mixed messages on rewrites, and I'm guilty of it too. But there's gotta be some litmus test where we compare the number of interacting states in the requirements versus the number of interacting states in the code and just call Time of Death.
> made a chart of every path and pattern of events
> It never had to be addressed again
Am I reading you correctly--every possible issue was knowable and preventable with enough prior thought?
If all your logic is immutable apart from a small area around the state machine the state gets much easier to manage. You write everything as f(state, inputs) => next_state. You can at any point inspect or print the state of the state machine, and record its transitions so you have a clearer idea of how it got there if there was indeed a bug.
Essentially, the author is, I think, arguing about what the paper calls "Essential complexity", complexity inherent to the problem one is trying to solve. And with that, I agree.
I think the author should acknowledge accidental complexity (or provide some argument as to why that must live somewhere), and I think a lot of comments here on HN are pointing out the fact that accidental complexity exists, and doesn't have to live somewhere. But my guess is that that's not what the author is saying, and that the author is only arguing about essential complexity.
[1]: https://github.com/papers-we-love/papers-we-love/blob/master...
¹I personally found this paper somewhat mixed. The definitions of complexity are what make it worth reading. Its conclusion of "functional programming languages will fix all the woes" I think is not practical.
Stuff like over-abstracting, insane default values, duplicated state that needs to be synchronized (e.g in React components) or just overly-repetitive code, for example.
Yes, it can be eliminated, but only to some extent. Programming languages add a baseline of complexity by themselves that _can't_ be removed: programming languages are not infinitely flexible so there will always be problem characteristics (invariants, data structures, etc) that will not be easily expressible. Those are new sources of complexity.
Think how Java simplifies manual memory management when compared with C or C++. Or how the Rust borrow-checker provably prevents a whole class of bugs. But these are all tradeoffs, both Java and Rust are ill-suited to express other problems.
Some kinds of type systems similarly lead to accidental complexity, for example in the form of repetitive code because the type system makes sufficient abstraction impossible. It's possible for some kinds of requirements to mandate very concrete types with no room for abstraction in some parts of a program, but even then the compiler can do most of the work with making sufficiently abstract code concrete and performant.
The complexity of getting different pieces to talk to each other? I can see how that's only conditionally "essential" in that, if you have enough control over the environment, you can make the various distributed components as similar as possible, never deal with version skew, and never deal with varying interpretations of the same protocol standard.
OTOH, if you're writing a web page, and you have to make that page useful across multiple browser types and versions, some way of dealing with the vagaries of different browser implementations becomes an essential complexity: You either push it into a library or framework or you handle it in the main codebase, but something has to know what MSIE can and cannot handle.
Further, there's different levels and causes of “coordination didn't happen”, such as “we tried, but got confused and didn't have time to understand each other properly” or “actively sabotaging everyone else's ability to deal with the aggregate system so that I can get ahead”… and somewhere in there is one that stands out to me as specifically interacting with technical complexity, along the lines of “I want my piece to not have to be the one that changes and/or takes on other costs”. Which is part of the social dynamic of moving externally-essential complexity around within the aggregate system and potentially creating ripples of accidental complexity along the way.
(Whether and how this also happens when the information systems in question are human minds rather than software components might be relevant but also an exercise for the reader…)
The failing of this point is that much of what we call complexity is disorganization. Certainly, there is a fundamental level of logic in any desired system that can not be willed away with cute patterns, but to consider all complexity of an existing system to be necessary is a fallacy. Dividing systems into problem domains does not inherently reduce the total complexity of a system. It probably usually adds to it. But organizing systems this way can drastically reduce the scope of complexity into manageable pieces so that mere mortals can work on it with out having to hold the entire system in their mind at one time.
It’s like saying that you can’t make a garage full of junk any less complicated, because no matter how you arrange it, it will still contain all the same junk. In fact, organizing all the junk into manageable storage can make it much easier to understand, work with, sort through, clean, and identify items that may be unnecessary.
The inherent complexity can be taken on by the solution (and the inherent complexity may be hidden behind a simple interface, or it may be introduced in the interface, but that's another problem), or it gets removed of the problem domain, and then needs to be dealt with by the user.
There's no correct answer; half the problem is defining the appropriate problem domain, and even then there are only better or worse solutions. The incidental complexity of the system comes purely down to the implementation, which often comes down to how well the problem domain is defined in the first place.
If you want to reduce the complexity of retrieving items out of the garage, you can move some of that complexity to when you store things in the garage by organizing them in a certain way. Without a doubt this is more complex when storing, but we would all agree that it is a creates a good balance in complexity between tasks.
However, I would argue that if you ever only store 10 items in a garage, then spending time organizing them in a way that reduces retrieval time would be an utter waste of time, and hence (to completely jump out of my/our analogy) why a good business person takes all of this into consideration when making decisions on how to structure their code and business and what complexity to take on.
There are also ways to design things better (or worse) with regards how they handle complexity. Reuse UI, patterns, workflows and languages. Keep things consistent, make things discoverable.
Point is, the slogan "Complexity has to live somewhere" could also be used as an excuse to do sloppy work. Then again, this keeps us all employed.
I think features are definitely worth servers being installed in a data center, and most people coming up with product requirements are fine with the monetary costs.
Features leading to bugs and wasted development time... that's where the point of the article lives. There are ways to build software that can more easily accommodate changes.
However, it's more common that developers try to hide the complexity of features behind abstractions that hinder more than they help. And that's what results in breakages and lost time.
Of course your software will also add some accidental complexity as well (perhaps a lot) but ultimately the software can't be any simpler than the business process it's trying to model. And if that process is a spaghetti nightmare, well, you're just stuck with that.
And true to the article, that is only half of the equation. Complexity is not created as much as it is moved around. By adding code that makes old emails searchable instead of simply archiving them moves the business complexity of customers with this specific edge case to your code base.
It is then a business decision. And I agree with you that it is often not worth it, but the art of business is in understanding when taking on the added complexity is worth it to differentiate your product over your competitors. Or stated another way, strategically deciding to reduce the complexity of your customer's business and increasing the complexity of your own.
This is not always an obvious decision, hence why the developer rank and file often look at it from one side only, and the product manager's/sales staff often look at it from their side.
It takes a strong leader who understands both sides to make the right strategic decision. When this decision is not made thoughtfully you end up with an overburden of complexity for a subset of customers that don't really move the needle on your business.
https://news.ycombinator.com/item?id=18774619
One of my highest upvoted comments with 161 upvotes.
But I've come to another idea too. Part of all this complexity is dealing with change. We could simplify things if some aspects of our software hit a final point where only security updates were published after that. Imagine a programming language that was specified and actually finished without the constant roll of changing patterns and practices. Or a web framework. Or a database. Intentionally designed to be robust and secure from the first day then intentionally set to minimal patches for bugs and security fixes.
Part of the issue though is that things keep changing. New characters are added to unicode, new timezones emerge or existing ones change. It's a hard problem to crack.
An E-program is written to perform some real-world activity; how it should behave is strongly linked to the environment in which it runs, and such a program needs to adapt to varying requirements and circumstances in that environment
Furthermore, as a part of its own environment, the software's own existence has an impact on what people expect and want to do with it.
Many programming languages like that exist: ANSI C, Pascal, PHP 1.0, Ruby 1.0, Python 1.0, etc. No one uses them. They could simply choose not to upgrade, but they don't.
If you assume people, on aggregate, are reasonable, then it seems that the value of changing languages outweighs the cost of the churn.
* Any language has its paradigms that need to be learned, and golang is no different. Just because you can learn the syntax in a couple of sittings does not mean you know how to use the language in an effective matter. Something I see many people not bring up.
This article has some nice specific examples of "Simple" APIs that push the complexity onto the programmer. https://fasterthanli.me/blog/2020/i-want-off-mr-golangs-wild...
Another common example I've seen cited is the need for Go code generation tools in the community (lack of generics pushing the complexity to external tools).
Otherwise this essay is spot on when it comes to essential conplexity.
Incidentally, the question of “where complexity lives” is one of the focal points of “A Philosophy of Software Design,” which comes highly recommended if you’re trying to come up with your strategy for managing complexity from first principles.
Their mental models and their understanding of everything is not fungible, but is still real and often what lets us shift the complexity outside of the software.
The teachings of disciplines like resilience engineering and models like naturalistic decision making is that this tacit knowledge and expertise can be surfaced, trained, and given the right environment to grow and gain effectiveness. It expresses itself in the active adaptation of organizations.
But as long as you look at the software as a system of its own that is independent from the people who use, write, and maintain it, it looks like the complexity just vanishes if it's not in the code.
Yes, this is Larry Wall's waterbed theory: https://en.wikipedia.org/wiki/Waterbed_theory
I do think it's important to distinguish accidental and essential complexity. Some complexity is inherent and if you think you've eliminated it, all you have really done is made it someone else's problem.
But there is also a lot of complexity that is simply unnecessary and can be eliminated entirely with effort. Humans make mistakes and some of those mistakes end up in code. Software that does something that no one ever intended can be simplified by having that behavior removed.
Along the same lines, there's a great quote from many years ago that I unfortunately can't find the exact text of, but it goes like this (paraphrasing):
"Most Microsoft Word users only use 5% of its features."
"So why don't we get rid of the other 95%, since it's so bloated and complex?"
"Because each user uses a different 5%."
It was 80/20, but the sentiment holds.
Dirt has to live somewhere, too. That's a good motto for a garbage company, but if you hear a chef saying it in the kitchen while preparing your meal, you might be worried.
Certain knots can be solved without changing the topology! They would just add artifical complexity.
Also, some truths can have both very simple and very involved proofs. They are a perfect example how just the right approach can reduce much complexity. Just formulating the problem in a different way can already reduce much artifical complexity.
When I want to search for a string within some data, I don't want to have to think about the algorithm it uses (naive, Boyer–Moore [1], etc.). I just want to know that somebody cared about it and that it will work great.
Bad abstractions, on the other hand, make you think about the underlying details that you have to be aware of. For one layer that might not be too bad, but the more you build on top of such things, the more fragile the whole construct becomes.
[1] https://en.wikipedia.org/wiki/Boyer%E2%80%93Moore_string-sea...
The point of the text is, that you have to define where to put the complexity:
If you are lucky, it lives in well-defined places.
With nowhere to go, it has to roam everywhere in your system, both in your code and in people's heads.
When defining the abstraction you focus on the practical use-cases so that it is as simple as possible but not simpler. The main task here is to find the right focus. In doubt, it might be better to use multiple different abstractions (e.g. if there are two distinct use-cases).
The goal is to find a construct where everything looks simple from the outside because the complexity is hidden by the abstraction.
The real question of design, especially the design and organization of computer programs, is where that complexity lives, and how do you organize that complexity so that it doesn't introduce extra complexity.
Usually the way we handle this is through layers, with the initial layers being simple, the middle layers being complex and the top layers being simple again. There's no theory about why this is better, it is just usually how it's done and it seems to work when we can pull it off.
The bottom layers are your primitives and axioms. Simple building blocks. This is where designers want to create a most minimal and simple set of tools that can help facilitate unlimited complexity at the upper layers (Think a minimal set of language primitives that allow for Turing completeness)
The middle layers are your theorems. Different permutations and compositions of your axioms to implement things like business logic or whatever logic you want. This is where good designers should try to stuff as much complexity into as possible.
The Top and final layer is the interface. Something that is built to directly access the middle layer while providing an interface that is simple and more understandable from the perspective of a user. This is probably the least theoretical aspect of the stack as you have to factor in things like art and human psychology into how you build an interface.
A good example of this will be your phone. The primitives are assembly language instructions, RISC. The middle layer is all the code that facilitates programming and applications and the interface layer is the touch screen.
Note that you don't have to look at it from a birds eye view from CPU instructions all the way to touch screen. You can zoom in and look at just a web stack: Java on the backend all the way up to react on the front end.
By zooming in you will see systems that violate the design principle I describe. Java as a primitive has become incredibly bloated and so has React and the javascript tool chain. There are several reasons for why this happens: Lack of planning, optimization requirements inevitably violate design principles, lack of knowledge, lack of central direction and more.
Either way, always remember that the ideal design that most people should strive for is complexity sandwiched by simplicity at the primitive layer and the interface layer. This goes for most things within computer programming and outside of it, like your car.
I can't tell you how many engineers and designers I've worked with for whom it is not obvious.
Unfortunately, the lesson that complexity has to exist seemingly needs to be re-taught constantly.
A lot of people really truly believe that a beautiful, minimalist, elegant product is always the best solution -- and that the users needing it to do other/extra things are the ones who are wrong.
The real question is what to do with complexity. Like you said some people think that complexity should be pushed onto users other people like you and I disagree.
I'm not sure I'd agree. Some complexity exists because the effort to simplify was too great. Either cost or skill prevented the refactor. As the saying goes, "sorry this letter is so long, I didn't have time to make it shorter."
I've repeatedly found that the first iteration of a solution tends to be more complex. One culprit is that sufficient abstractions were either not recognized or not implemented. Put those abstractions in, and you simplify the system.
So that refactoring step after an initial solution is created is crucial and unfortunately often just not done.
If we accept that the integration itself is necessary, then the complexity necessarily comes along with it. Not to mention all the complexity generated by the client base: I'm paying for this system, it's mine, so make it do X, even if you think X is too complicated/impossible/etc.
So, no, it doesn't have to live somewhere. It can be created and destroyed, this happens every day. Probably some of it can not be destroyed, but nobody knows what part so any article about it will be useless.
> We try to get rid of the complexity, control it, and seek simplicity.
Well, not really enough. Complexity is a beast that takes many years to understand, and that's when you are really trying. Many devs don't. And companies even less. So while it's true some complexity is unavoidable, I think we still have a long way to go in being aware of it first. We write software once, and that's not a proper strategy to manage complexity. Anyone who has looked into computers from top to bottom knows we have tons of problems with complexity that we really haven't solved properly, and that bleed into day to day development in the ugliest ways.
My particular take when I don't have the massive amount of time required to properly deal with complexity is write something like this: "hey, this is very complex. please don't touch it. if you have to, we have this much headroom. if you need to go beyond that, please rewrite this entirely / find a better way. if the code doesn't bite you immediately I will".
If you are lucky, you have someone who is able to analyze the entire system, can identify all of the stakeholders, and can drive consensus when the change in abstractions is necessary, and has the budget to do it.
The danger is that people pick the solution for a large solution first because "we will need this someday" rather than waiting for the accidental complexity to build, knowing that the rework necessary to move to an intermediate system is less expensive than the cost of delay in getting the simple solution out first.
My dream is that we can build incremental complexity systems that could support simple solutions quickly and highly complex solutions eventually. The problem is that these are hard to build. :)
Now look at this app I can write in 30 lines of code with this framework! Oh and 500 lines of yaml across 2 dozen files but look the "code" is so sleek and sexy.
Joking aside I think he could have at least mentioned the difference between required complexity and unneeded complexity. Ironically it seems to spawn from attempts to reduce the first type of required complexity in a lot of cases.
Simplicity is the art of taking complex things and stripping them down to their essential nature. Simple things can be complex, but no more complex than they have to be.
With that said, this piece talks a lot about frameworks and tools and patterns, and I really hope people don't think that's the right way to think about architecture. Abstractions do not make things simpler, they make them abstract.
I am fully on board with the overall sentiment of the article that there is some irreducible complexity that one cannot avoid. Often it comes straight from the business domain or entity the code is modelling, and then, sure, you cannot make it any easier, otherwise you are not solving the problem and users will be unhappy.
However, then the author goes too far: > Accidental complexity is just essential complexity that shows its age.
You can absolutely add heaps of unnecessary complexity; in fact, this is almost surely what you will get by attempting to cover every possible future use case and evolution scenario, or to cater to every minor aesthetic concern (e.g., "I need to be able to replace my cloud provider / DB / message queue / user notification medium / etc by changing just one line!").
It takes humility to admit that getting the balance "right" from the start is difficult, and often the only way to improve is to accept some badness now and revisit the decision later with more information.
The quote however seems to be saying that we (as software engineers) always get it right, just later things change and our choice does not appear right anymore. This mindset is counterproductive, as admitting flaws is the first step to improve.
Doesn't that hold that the inverse must be true? In some cases, you must be able to remove complexity from the equation, without introducing it somewhere else.
I agree with the spirit of the post, but I think it's eliding some of the complexity involved in identifying essential complexity.
I tried to describe them here. I'm not sure I did a good job:
https://medium.com/@b.essiambre/product-market-crossfit-c09b...
The gist:
Low resolution fit: The product vaguely and simply fits its market or domain.
High resolution fit: The product is complex and is tightly tailored to the market or domain.
Overfit: The product is so tightly tailored to some users that it only fits a small part of the market.
Underfit: The product is so generic that it’s missing important features and corner cases.
And then another qualifier 'crossfit' which is a bit harder to grasp, inspired by cross-entropy, that has to do with whether the design language, which can be seen in some ways as 'encoding' the problem, fits the domain or market well.
Yet for other people it's a simple as fuck thing that accomplishes things for them and lets them move forward with getting shit done.
So really, it's how you look at it.
The human body is so freaking complex, it is entirely possible we'll never even understand how half of it works. Yet from another point of view we easily hire a bunch of humans to do tasks (programming, making food, yard work, etc..). And even though we don't know how the body works, we have a simple understanding of how that body can accomplish those tasks, which is enough to move forward.
Sometimes there is /essential complexity/, but it is far from always. I think the line of thinking espoused by the author leads to less simplification and encourages a "well, the complexity has to live somewhere, so why bother?" type of approach. However, reducing complexity is often possible, and it usually comes before the implementation stage. Reducing complexity sometimes means solving a different problem.
If you write anything, even "hello, world", complexity lives somewhere! Just the complexity of the print function path all the way to the hardware would baffle some.
Keeping them orthogonal seems the first step in the process of trading the problem for a smaller problem.
As TFA notes, there is a lower limit to how much complexity can be squeezed out of a system.
Go : Make the language simple, make achieving a simple thing complex.
C : The machine is complex, be careful what you do if you overflow into the inner workings.