[Bratmon]

This article is very rambling and somehow manages to miss the most important driver of reduced software quality: shifting requirements.

You can have the most beautiful perfect design that leverages the greatest abstractions in the world, and then have it all entirely destroyed by a single requirement shift that totally kills your abstractions.

[tcfhgj]

well you cannot design for the unpredictable, but well designed code will make refactoring easier anyways

[adityaathalye]

Post author here. Guilty as charged re: rambling. Brevity is not my strong suit.

To the second point, I'd say I'm in violent agreement with you. In fact, my whole thesis is that because software mutates all the time, it is observed that successful software mutates successfully.

Obnoxiously quoting self:

> Unlike all other machines, it is pure concept, and as such it is infinitely malleable and mutable. And mutate it does, all the time.

and several paragraphs below that...

> Not only does all software mutate, we also end up performing all kinds of deep surgery on the organisation that produces it. The whole thing—product and org—is simultaneously flexed, reconfigured, and even totally redesigned in-place with rapidity that is very uncommon in other industries. Why? Because software fundamentally is peoples' thoughts being played on repeat.

[timw4mail]

I think a good way to put it is this: every abstraction is a gamble.

There are likely a lot of safe bets, but even choices that once made sense can become an issue later.

[evan_a_a]

If the design is so inflexible that a single requirement shift destroys it, then it wasn't a good design.

[PaulHoule]

That's a refutation of agile, Vibe Coding and the last 25 years of the software industry.

One of my own quality metrics that almost everyone rejects is a "a good design is a design where a change that seems like a small change to management is... a small change"

It plays out like this. To get to product-market fit you are going to have to zig and zag 20 times. Everybody optimizes for the first step of this journey and then on the second step they optimize for that... and by step 5 or so the archiecture is exhausted and you have no hope of getting to step 20. The rare team that is looking ahead 20 steps is the one that survives.

[pydry]

>It plays out like this. To get to product-market fit you are going to have to zig and zag 20 times. Everybody optimizes for the first step of this journey and then on the second step they optimize for that... and by step 5 or so the archiecture is exhausted and you have no hope of getting to step 20.

It not being "exhausted" is just what distinguishes a good programmer from a bad one. If you're good you can keep a lid on tech debt in the face of rapidly changing requirements. If you suck at it, well... you can't.

>The rare team that is looking ahead 20 steps is the one that survives.

lol no good software engineering means assuming that you don't have a crystal ball (because you never will) and building accordingly.

This usually means YAGNI, minimizing up front design and refactoring aggressively. These are the things that keep you agile enough to move in many different potential directions in response to different potential futures.

The time when this really got rammed home for me was working on a startup where a political black swan event ended up changing key requirements of our software overnight. This pulled the optimal architecture in a wildly different direction to where we anticipated it was going in the previous week.

[xedrac]

> If you're good you can keep a lid on tech debt in the face of rapidly changing requirements. If you suck at it, well... you can't.

Every drop of tech debt (aside from AI slop) I have ever been a part of has manifested simply because some manager or executive made the decision to sacrifice quality for short-term speed. Sometimes this looks like an unskilled developers producing something that management won't let you fix. Sometimes it's a completely unrealistic deadline that requires extreme measures to meet. If you are able to keep a lid on tech debug in the face of rapidly changing requirements, it likely means that you have enough slack in your job to do so.

[pydry]

Asking permission to fix tech debt is a rookie mistake. You need to bake the fixes into the ticket.

Half of my job as a lead is to cajole, threaten and convince juniors into spending 5 days on a ticket to do it properly rather than spending 2 in order to get brownie points for "looking" productive.

Ive had more than one call where the junior says "we don't have time for that" and I have to call in the PM to let them know that it's OK, we do.

There are environments where management put so much pressure on you that you can't help but deliver crap but not everywhere has to be like that.

[GuB-42]

The problem is that you don't know in advance what the 20 steps will be, and you may spend time solving the wrong problem.

Make step 1 as simple as possible, step 1 is small and changes are cheap, you can even rewrite everything with little effort, no need to prepare for anything at this point.

Step 2-4 will inevitably break something, no big deal, rewrite the parts that need to be rewritten, it is still cheap.

At step 5 rewrites are starting to get costly, but now you have a better idea of what is here to stay, so it is now time to invest in things like regression testing, style guides, etc...

At step 10, you have quite a history, rewrites become a really bad idea, no turning back after that. So now it is time to think about the future. The difference from doing that in step 1 is that you now have 10 steps of experience and only need to look ahead 10 steps.

[deterministic]

Nope not true at all (30+ years of experience here working on very large scale software).

The whole point of being agile is to design your software architecture to be agile. That's what I do. And it works.

[AlotOfReading]

Then I guess no design is good. Here's an easy example: take any software system you can think of. Let's add a single new requirement: it must produce an output in 1us. I'll bet your system has to change.

[regular_trash]

Well that's a pretty contrived example. There are fair amount of requirements that can be predicted as "possibly coming" in a given problem space, and a good design typically takes them into consideration and plans contingencies such that a design doesn't crumble.

[AlotOfReading]

It's extreme to demonstrate the point, but it's not contrived. Realtime deadlines are a completely normal requirement for the systems I work on. I have change reqs on my desk right now that specify a full reboot back to network responsiveness in <=300ms. That definitely changes how the system is designed.

[pydry]

It being extreme is what makes it contrived.

It's also a requirement programmers often overemphasise due to the McNamara fallacy.

[Supermancho]

Contrived does not mean unrealistic. I have seen many extreme requirements come down on short notice. Over time, the peter priciple guarantees that there will be a middle manager that miscommunicates an agenda they dont fully understand, then push an emergency change down the hierarchy.

[AlotOfReading]

You missed the point of the example. The parent said any design that has to change with a single requirement shift is a bad design. So I gave a requirement that virtually any design would have to change to accommodate. It doesn't preclude the existence of other, more plausible requirements that shift designs. I even provided a real example of one, which you've ignored.

[akoboldfrying]

Right, the point is that GP's definition should take the nuance you mention into account (namely, that some types of changes are more likely, and thus more important to handle nicely, than others).

[Rzor]

I think you are interpreting his claim in the most extreme possible way. You are doing the equivalent of asking a civil engineer what if his bridge needs to support flying cars, or weight half as much while carrying ten times the load. Of course there are architecture breaking requirements, but his point still stands for reasonable and foreseeable changes. If it doesn't survive that, then it wasn't a good design begin with.

[phailhaus]

Completely untrue. Every design can be destroyed by a single poorly-thought out requirement. I can take down Facebook's design just by saying "any edit must immediately be visible to all users".

[deterministic]

That's a straw man argument. Your new requirement is physically impossible. No software team can break the laws of physics.

So the real question is what exactly do you mean by "immediately"? There are many answers with different pros/cons and cost. And I bet the solution won't impact 99% of the software Facebook has already written.

[phailhaus]

My new requirement is NOT physically impossible. I did not mean "immediately" as in "zero milliseconds", it meant strong read-after-write consistency. It makes facebook so slow and unscalable as to be unusable. Facebook deliberately decided that stale reads are worth it for responsiveness and scale.

[deterministic]

So your argument is that no solution could satisfy your new requirement, not even if Facebook rewrote the entire system from scratch.

If that's the case, then your example doesn't really address what we're discussing: whether software can be designed to remain flexible under realistic requirement changes.

A stronger argument would be to give an example of a realistic requirement change that would force all possible software architecture to be rewritten from scratch. Do you have an example like that?

[deterministic]

If that ever happens to you, you're doing it wrong.

I've spent 30+ years building large-scale software, and it has never happened to me. Good software design lets you accommodate even major requirement changes with minimal changes to the code.

[xedrac]

If you're using Lisp, then sure. If you're using C++, then chances are you're going to have to heavily overengineer the thing to cover all possible requirement changes. That comes at a great cost - in most cases, you are unlikely to benefit from the overly abstract design, but you will pay for that additional complexity for the life of the project. I am convinced that you are better off designing the project to solve the problem in front of you, and then if the requirements shift, pay the cost to solve the new problem, rather than forever pay the cost of solving a host of unrealized problems.

[deterministic]

> If you're using Lisp, then sure.

So are you saying Lisp somehow makes software flexible enough to handle any requirement change, no matter how badly it's written?

Really?

Lisp doesn't have any unique feature that makes this possible. Macros, meta programming, and single type systems all exist in other languages too.

> If you're using C++, then chances are you're going to have to heavily overengineer the thing to cover all possible requirement changes.

I disagree. My experience has been the opposite.

I maintain a large C++ codebase used by companies around the world. I spend almost all my time adding features and almost none fixing production bugs. The code is flexible because it's well designed and well tested.

The language isn't what's important. Good design and good testing make software flexible. Poor design doesn't, no matter what language you use.