Chaining vs. Nesting | Hacker News Reader

Chaining vs. Nesting

frankschmitt.org

43 pointsfrankus19 comments

Function composition and currying seem like a more natural translation to me.

Or you can think of it like this:

netstat_n >>= grep "tcp" >>= flip awk 5 >>= sort >>= uniq_c >>= sort_n

Glad someone posted that link, or I was going to :)

Pipes and method chaining "feel" the same. They're both expressions, but with the property that you could rewrite them as what looks like a series of statements without changing the semantics (although possibly changing the efficiency). e.g.

    usual_suspects.director.surname

has the same value ("Singer") as

    person = usual_suspects.director
    surname = person.surname
    surname

and

    cat /etc/passwd | grep root | cut -d: -f6

outputs the same thing (root's home directory) as

   cat /etc/passwd > tmp1
   grep root tmp1 > tmp2
   cut -d: -f6 tmp2 > tmp3
   cat tmp3

They feel the same because they're both monadic (at least when the chained methods or piped processes behave in certain 'normal' ways). '.' (the method invocation operator) and '|' (the pipe operator) are monadic combinators, just like Haskell's >>= and F#'s |>. And a monadic combinator is just a generalisation of function composition.

Monads are going to become a lot more important in the next few years, as programming languages get more expressive and capable of more abstraction. LINQ in C# and VB is a great example - while it looks like they've baked in language support for lots of concepts (querying databases and XML, for data parallelism and for event-driven programming), what they've really done is recognised that those are all forms of monadic computation, baked in language support for one thing (monadic expressions), then implemented those concepts as libraries that any user could have written. (Erik Meijer and Brian Beckman on MS's Channel 9 explain this really nicely.)

Here's the example from the article written using the Clojure thread macro:

(-> netstat_n (grep "tcp") (awk 5) sort uniq_c sort_n print)

I remember first running into this point with http://xahlee.org/UnixResource_dir/writ/notations.html and I'm sure Slava or somebody has pointed out that in postfix languages like Factor and Forth everything is essentially chained which very naturally leads to a function composition style. And that despite the "reverse" nature of RPN it often ends up being "forward", witness in Factor REPL actually being REPL instead of L(P(E(R))).

I've seen an interesting way to do method chaining in haskell: The '$' operator: Basically it's a low-priority right-associative binary operator that applies the function to the left to everything to the right.

Using the example from the article, one could do it as:

sort -n $ uniq -n $ sort $ awk '{ print $5}' $ grep tcp $ netstat -n

I.e essentially the same but in reverse order.

($) helps to avoid some parens. The (.) is more common to change functions, and closely related. ($) does not really compose functions, it is function application and right associative.

Your example would rather read something close to

  sort -n . uniq -n . sort . awk '{ print $5}' . grep tcp . netstat -n

(And `$ argument' at the end, if there was one.)

($) is actually the least interesting operator in the world :) But as you say, its parsing precedence makes it useful (because it lets you skip parens).

What I like is how ($) generalises to (<$>) for applicative functors, so you can write

   succ $ 5                -- -> 6
   succ <$> Nothing        -- -> Nothing
   succ <$> Just 5         -- -> Just 6
   succ <$> [1, 2, 3]      -- -> [2, 3, 4]

Who needs "for" loops? :)

I agree. Though I'd say id or const or even less interesting than ($).

(`to change functions' should have been `to chain functions'.)

I don't think ternary expression is ugly by itself, it's only ugly if parameters themselves are chains beacuse then you're back to having a tree. Which is the fundamental problem - chaining is not good for branches, so any branches send you back to nesting. On the higher level theproblem is that you are trying to express a tree as a list - ain't going to work.

Haskel side-steps this problem in a neat way - you can take advantage of delayed evaluation and give names to both arguments of the ternary expression so each individual piece looks linear. So your "tree" is broken down line by line:

  x=...
  y=...
  z=f?x:y

which is as neat and easy to read as a chain, but is more powerful because it can express a tree.

Once could enforce this sort of reader-friendly style by forbidding expressions in ternary operators.

Comparing a model in which the contract is 'everything is a stream of characters' to function application and composition in just about any programming language seems more than a little clueless.

"I think there's a connection. Certainly in the early days of Unix, pipes facilitated function composition on the command line. You could take an input, perform some transformation on it, and then pipe the output into another program. (...) [The mathematical approach of function composition is] exactly what I mean. (...) I think [that mathematical formalism] was right there from the start. Doug McIlroy, at least in my book, deserves the credit for pipes. He thought like a mathematician and I think he had this connection right from the start. I think of the Unix command line as a prototypical functional language."

-- from an interview with Alfred Aho, one of the creators of AWK

http://books.google.com/books?id=yB1WwURwBUQC&lpg=PT120&...

It's 2010, not the 'early days of Unix'. Yes there is a connection, a useful one, if you are sitting in front of a PDP-8.

http://groups.google.com/group/comp.lang.functional/msg/8476...

http://regex.info/blog/2006-09-15/247#comment-3085

I don't think unix pipes in the shell would be very useful to someone sitting in front of a PDP-8

The machine's panel switches are enough of a pain when programming in the absolute -- give a shell user a teletype at least!

A bit of a hyperbole, granted. Pipes -> any kind of vaguely modern runtime (clojure? srsly?) is a profoundly nonsensical comparison though.

I don't understand your hostility to pipes. It seems to be based on the fact that they were invented some time ago? On the contrary, the soundness of an idea famous enough to be known today is probably positively correlated with how old it is.

Pipes are a combination of dataflow programming and monadic computation, and are naturally parallelizable, functional and generally pipe processes don't share mutable state. All these concepts are highly relevant to modern programming.

pvg's dislike seems to be that pipes transfer a stream of octets with no inherent syntax or semantics. That means every program needs baroque options to impose structure, which is difficult to create, maintain, and debug. Also brittle.

And on the other hand, text streams are easy to work with, because ad-hoc tools can be written to work with them, rather than having to wrestle with strongly typed libraries and being constrained to languages compatible with the libraries that define those types. Imagine a web written using CORBA RPCs over rendering APIs, rather than HTML over HTTP.

You can argue that HTML has inherent syntax and semantics, but of course server and client can have slightly different ideas, and it all still works, mostly. The same is largely true of shell programming using pipes: different stages in the pipe expect certain formats, for regular expression or field extraction, and format pasting together, etc. The format is easy to eyeball and easy to test on the shell REPL, so in practice the problems aren't large.

I think my hostility was more towards the posted article than pipes.

Reply on news.ycombinator.com