Ask HN: Bottom Up Programming Question

Personally, I'm against HtDP because it strips out all of the romance in programming for me, but it does work.

A much better way of thinking about "bottom-up design" is exploratory programming - where one decides what to write by writing it [1]. This means you start by discovering and solving sub-problems, by creating components which work in their own right, documenting and unit testing them (if you're working in an environment that needs it), and moving on to solving bigger problems using said components.

On the other hand is the "architecture approach" - where you measure twice and cut once. By "measure", I mean decide what all of your components will be, write out UML diagrams for them, figure out your types, your interfaces, and the contracts between them, write unit tests, and only then begin to write run-time code.

There are much more experience programmers than me who advocate both ways, so it's hard for me to make a call [2].

Another great metaphor I find analogous is "Mozart vs. Beethoven" style programming: http://stackoverflow.com/questions/292682/understanding-dijk...

I like this one because suggests that you can make "great" programs either way. I use the quotes, because obviously, music is a matter of taste, as is the quality of software we develop. This is important to remember - for some people, Windows is the perfect OS out there (for example, those concerned with Microsoft's profits over the OS's usability).

1. http://arclanguage.org/

2. Although the ones who advocate exploratory design seem more well-rouded: Alan Kay, Paul Graham, and the 37signals guys come to mind. Since I want to be well-rounded and not a specialist, I try to follow in their example. Obviously, this is a subjective judgment.

Duck 1.2: Eyes.

Duck 1.21: Eyes now see color.

Duck 1.22: Eyes now see colors including black. Apparently, some bugs are black. Damn users and their damn uncolored colored bugs.

Duck 1.5: Realized eyes and mouth both share a lot of code. Refactored as "nerves". Allowed consumption of turquoise bugs -- not exactly sure why they're safe when blue bugs are not, oh well. Added quack just for fun. Spent 2 weeks getting the quack to sound exactly right.

Duck 2.0: Now with feet and wings! Turns out to have feet you really need muscle tissue. Refactored again, stealing code from Mouth. Supported modes of operation include walking, running, and swimming. Option in menu entry for flying but if you push it the duck dies.

Duck 2.1: It glides!

Duck 2.11: So it turns out that to utilize the gliding feature users were driving ducks off of tall cliffs. This resulted in downward acceleration being sufficient at end of glide to overly stress feet muscles, causing critical duck failure. Added in Bones. Damn users and their damn cliffs. As long as I had Bones I implemented them in the Mouth too, can finally eat blue bugs.

Duck 2.12: Some blue bugs are poisonous?! WTF kind of bugs do you have in Australia. Added in Tongue. Borrows heavily from existing Nerve classes. More refactoring.

...

btw, duck can't fly so u may want to cut that feature

you would start with cells - sounds good

you would first have had to have divided the duck up conceptually into different parts in order to know that different cell types are needed - does not follow.

I noticed that your premise assumes a requirements analysis that already nails down what the program is supposed to do. While that way of thinking is common in the corporate world (where, I note in passing, effective software development is not common - coincidence?), it doesn't fit well with the style of programming you're asking about. If you've read PG's piece on the subject (http://www.paulgraham.com/progbot.html) you may recall that there's one passage in it that he emphasized. It begins:

It's worth emphasizing that bottom-up design doesn't mean just writing the same program in a different order. When you work bottom-up, you usually end up with a different program.

The program changes as it evolves and your understanding of what the program is changes as it evolves [~]. All sorts of interesting feedback loops result.

When I work this way, I listen to what the code is telling me. That is, I start with something primitive (your "cells" in the analogy) and sense where it's leading. Perhaps it's akin to riding a horse where there's a balance between what you want to do and what the horse wants to do; or to creating a piece of art where the creative process incorporates feedback from the medium itself and from the work as it's emerging.

That doesn't mean it's all spontaneous, with no guiding factor. If that were true, no system would emerge - you'd just keep going around in circles and finding yourself back at the primitive level. But the guiding factor is not at all like "dividing the duck up conceptually into different parts". It is more like having a felt vision of what a duck would be like. You have that vision and you keep referring to it as you go along. But the vision itself clarifies, and sometimes changes, in response to what happens in the code. You discover what the system is as you make it.

Learning to listen to what the code is telling you is one of the most interesting things you can do as a programmer.

([~] PG's piece doesn't say that your concept of what the program is changes, only that the constructs it's composed of will be different. I think the two are related though.)

In code, a function to trim whitespace from a string is perfectly functional on its own. When you actually try to write code this way, a rather large number of things can be written, larger than you may think.

Rather than write a lengthy explanation, my suggestion is to write a non-trivial program in Haskell. After that you'll better understand how it is possible to have many functions that can stand on their own, and I think this will help clarify the situation in your mind.

You can do this in any other language, but most languages make it relatively easy to reach out and touch lots and lots of state. Real-world evolved systems tend to do the biological equivalent all over the place, to a level even the worst spaghetti code ever written doesn't even come close to aspiring to, and do not resemble human-engineered code in the slightest. The Haskell advantage here is that it won't let you. Achieving this in other languages right now is a chicken-and-egg problem; if you knew how to do that you wouldn't be using the metaphor you cited.

(BTW, just to state it explicitly: No, I didn't directly answer your question. As others point out, it's not a very informative question. I'm going for the root of the problem, how you are conceiving of functional units in a program.)

Ever heard of something called... LINUX!?

I'd start with the "outside" of the program and work inwards. This forces you to categorize features as greatest to least important.

Refactoring can be significant -- much more than top-down. When used poorly, this can lead to the "design a 747 by building a wheel first" mentality, however. In other words, like everything else there's a place for it in the toolbelt.

You know, you can always model the problem domain first and still code bottom-up. The structure of the data is probably the most important factor anyhow, no matter how you code or what the qualities of the language are.

This will have two effects: first, the complexity of the design is equal to the complexity of the duck. That's A LOT of paper. And second, most mistakes in the design will cause the duck to fail spectacularly.

On the other hand, with bottom-up design you start with something which eats and reproduces, and improve it constantly until it has a beak and feathers and quacks. It will certainly not look anything like a duck at fist, but it is very likely that from the second or third iteration it will solve 80% of the client's problems. Sometimes even from the first, but that's just luck. It's messy, but it's at least doable.