Comment by agentultra | Hacker News Reader

Comment by agentultra | Hacker News Reader

Great talk.

Simplicity is, of course, key; but a few of his applications of these principles are misguided IMO.

Ex: The "Parens are hard!!" slide. He suggests that parens are "overloaded" in CL/Scheme because they are used to wrap functions, data, and structures all the same. However he completely misses the fact that by representing everything with parens, CL/Scheme remove a lot of complexity in software written in those languages.

AFAIK, the only languages that do macros right are homoiconic. Anything else is too complicated. Just look at Perl 6 and Haskell macros. They require learning special syntax and really crazy semantics. Using them will probably just make your program more difficult to understand.

He also "rails" against testing. He misses the virtues of a proper testing infrastructure in producing reliable software: if you don't test it, how do you know if it works? Because you reasoned about it? How do you know you're reasoning is correct?

True, "guard rail" testing isn't a crutch that will automatically produce reliable software. But I think Rich relies too much on this narrow view of testing to make his point. Testing is good and necessary.

And the jab to the Unix philosophy? ("Let's write parsers!"). Isn't that what we do on the web anyway? AFAIK, HTTP is still a text-based protocol. Any server that speaks the protocol has to parse a string at some point. So what was he getting at there? The Unix philosophy is about simplicity and has a lot to offer in terms of how one can design software to be simple.

Overall though, it's a great talk. I just think that if he wants to get pedantic then he could be a little more thorough and less opinionated. Everything he said about designing simple systems I pretty much agree with, but I think he glosses over some finer points with his own bias of what simplicity means.

Clojure is homoiconic just as much as CL and Scheme are. It just happens to use more than one datastructure to represent code.

The "oneness" in other lisps does not make things simpler, nor, in my opinion, easier. The reason why parens (lists) in traditional lisps are not simple is that they complect several different purposes. In contrast Clojure uses list forms (almost) exclusively for "active" expressions such as function and macro calls. For grouping and binding, vectors (and sometimes maps) are used instead.

In this manner Clojure manages to keep the roles of different data structures in code mostly simple.

The reason why parens (lists) in traditional lisps are not simple is that they complect several different purposes.

You've repeated what Hickey says in the talk, but I'm not sure I buy it. [Deleted boring stuff about boring parens here.]

The trouble with the general argument is that you can add constructs that are individually simpler but nevertheless make a system more complex. It's overall complexity that we should seek to minimize.

How does adding several independent simple things make systems more complex in a way that is unavoidable (ie. incidental complexity, not problem complexity)? The interaction between simple elements will be explicit, whereas complex elements by definition interact in ways that you can't control.

Of course, sometimes the complex tool might be exactly what you need to solve your problem, making things easier. But in cases where you need only a part of the functionality of this tool, the complexity bites you as all the unneeded functionality (along its cost) is forced on you.

What sort of "overall complexity" does having several data structures in code introduce? As Rich Hickey says in his talk, complexity is not about cardinality.

In Clojure's model, the elements are distinct, and as such there is more simplicity in the system than in lisps where you have to understand the difference ("untangle" the complexity) between several different uses of lists to get anywhere.

I also think having several datastructures makes code easier to read due to the visual differences, but that's a separate discussion. :)

Adding things means more things, which means more complexity. Doing so is a net win only if the new things subtract more complexity than they add. To figure out whether they do, you have to consider the system as a whole. That much is clear. Do we know how to do that? Not really. Empirical studies favor the simplest measurement of program complexity: code size. So that part is roughly taken care of. But I don't know of a good definition of "system". What counts as a single system and what counts as separate interacting systems, or subsystems? It's in the eye of the beholder. If I take 90% of a program and call it a library, has my program become 1/10 as complex?

Rich says his definition of simplicity is objective, but it's not. What determines whether a construct has a single purpose? It's whether or not, on inspection, you think it does. S-expressions seem to me to have a single purpose: to group things. How those groups of things are treated by the evaluator is a separate concern. You, following Rich, say no, they have 3 purposes. Ok. Why not 4? A function call and a macro call have different purposes; why conflate those? (I have no trouble reading function calls in s-expressions, but sometimes run into trouble not knowing that they are really macro calls, so this is not a made-up point.)

We have no objective basis for this type of discussion, only emotions and beliefs. Concepts like "readability" are hopelessly relative, but appeals to "ease of future change" are no better; to put it politely, they depend on individual perspective and experience; to put it bluntly, we imagine them.

I don't agree with your assessment that adding things necessarily increases complexity. Assuming these things are simple (in the objective sense as defined by Rich), actually suitable for the problem, and used correctly, then any complexity arising from their use is necessary to solve the problem, and can't be avoided.

The definition of a system in this case is anything with one or more components that accomplishes a particular purpose.

It still seems you're using a different definition of complexity than I am. To me, complexity implies unnecessarily intertwined elements.

In your hypothetical library situation, the answer is likely no. Your program still depends on the "library" code in ways that make the two not treatable as standalone entities, so no complexity has been removed.

An inherently complex tool may solve the specific problem it's built for, but it does not combine well with other tools. A simple tool tries to keep itself standalone so that can be freely combined with other simple tools to provide functionality that the original authors of either tool might not have envisioned. Clojure has many examples of this idea in action, but it's not the only language to exhibit simplicity.

I think Rich's definition of simple is straightforward and objective. For example, Clojure protocols fit the definition. They give you freely extensible polymorphism based on object types. Protocols don't even provide a direct way to do inheritance. That can be done by using a map of functions that you modify as needed, but requires no explicit support.

Your defense of s-expressions is rather puzzling. The evaluator defines what s-expressions (lists) mean in CL, and depending on context, there are multiple meanings that are completely separate. Certainly you can argue that Clojure conflates macro and function calls too (it is justified to me though, since they're all operators), but it has at least reduced complexity by not conflating binding and grouping with those elements.

As an added benefit, with few exceptions, whenever you see a list form in Clojure you can assume it's either a function call or a macro of some sort.

Having written a bit of Scheme and Clojure - Clojure's distinction between data structures make many things simpler - from writing code to writing macros.

As far as his comments on testing - I suggest you read this, http://blog.8thlight.com/uncle-bob/2011/10/20/Simple-Hickey....

Importantly, it should be mentioned that despite the way it distinguishes between different data structures, Clojure manages to retain the advantage of "representing everything with parens" by instead making sure that everything implements a common _interface_, (i.e., seq). In other words, as opposed to CL or Scheme, it separates logical list manipulation from the physical data structure, giving the best of both worlds. That's a big advantage for Clojure.

Big advantage for Clojure over what?

The point I was trying to make about parens is that I think Rich is creating a straw man from them. Parens in CL/Scheme don't have any special meaning other than demarcating lists. A Lisp compiler/interpreter just evaluates lists of symbols at the end of the day.

His point is that it puts together (complects.. ha!) lists, the data structure, and list-like operations into one and the same thing, when logically they are separate ideas. He's trying to point out that separating things brings simplicity, even if you end up with more of them, therefore unifying everything under one syntax for example is not necessarily "simpler". You may not agree, that's okay, but it's consistent with what he was saying, not a straw-man.

That's where I think he is 'discovering' complexity and in fact creating some.

That code and data can both be represented as lists is a major feature of what makes CL so compelling.

And I suppose that is the major contention -- is 'code as data' a complex idea? I say it's simple. There's no difference between the data structure of the code and the data structures the code acts on, so using the same operations on either should be trivial. With a handful of simple evaluation rules and a small number of special forms you can bootstrap an entire language written entirely in itself. Functions, classes, interfaces, namespaces, the whole nine-yards. The list is just an implementation detail and it's a very simple one that enables some very elaborate structures.

There's nothing you've expressed that doesn't apply to Clojure as well. I'm assuming you haven't done much Clojure and thus can't really illustrate what the problem is in practice.

Precisely. Clojure was my first Lisp, and I got spoiled by the "every collection is a seq" uniformity. In Clojure, functions like 'count' can operate on all collections, whereas in, say, Racket, you have an explosion of methods like:

length

vector-length

string-length

hash-count

set-count

...

Clojure embodies Alan Perlis' idea that "It is better to have 100 functions operate on one data structure than to have 10 functions operate on 10 data structures.", which is one of the many reasons why I enjoy the language so much.

> "It is better to have 100 functions operate on one data structure than to have 10 functions operate on 10 data structures."

Oddly enough though, you yourself are pointing out that it is better to have 1 function that operates on 10 data structures. ("length" operates on vectors, strings, hash tables, etc.) On the surface this seems opposed to the quote you chose.

However, it is clearer if you replace "data structure" with "interface". The is classic separation of concerns. When specifying the "what", we can get away with "100 functions operate on one interface", but the efficiency, the "how", can be specified independently based on the choice of data structure implementing that interface.

Yes, that is a better way of making the point. Thank you.

I think the Unix philosophy jab is about data. Instead of using string/text file with no form, I think he's suggesting that a standard format (like JSON) could be much simpler.

HTTP is just that, a protocol : a standard way for two entities to communicate. In Unix, there's no protocol, its just text without any standard form, which makes it difficult to write tools.

Without being confident enough to have voiced it earlier, I thought similar thoughts. Clojure appears to depart from something that characterizes both traditional Lisp and Unix systems – having one universal interface, with a big emphasis on one. Though by the definitions he outline, that would perhaps be easy more than simple.

Reply on news.ycombinator.com