The subsections of programs referred to in the previous section are often called functions. In mathematics, the term function has a fairly strict definition: a function is a mapping between one set of values and another. In computing, the definition of a function is looser: it's usually taken to mean a subsection which returns a value, but it can do all sorts of things which would be against the rules for a mathematical function. Functional programming, one of the oldest programming paradigms, relies on making computer functions, to a greater or lesser degree, obey the rules which apply to mathematical functions.
Functional programming is often seen as a bit marginal because the restrictions it places on the programmer by attempting to be like mathematics are fairly drastic. I'm going to talk about two of these: immutablility and removal of side-effects.
In terms of our imaginary workshop, immutability is the principle that an object can be placed in a box once, and once only. If box "A" contains that blue thing, it will contain the blue thing for all time (where "for all time" really means for the duration of this program, or subsection).
The boxes in our workshop are variables, and immutability is a hard concept for a lot of coders to get their head around: an immutable variable seems like a contradiction in terms, and if you've come from the imperative paradigm much of what your program (thinking in terms of the little homunculus) actually does is change variables: take some input, do some calculations, and then set the relevant variables to the correct value.
Immutability doesn't mean that the program can't perform calculations: just that the results of these calculations must be stored in a different variable, not used to overwrite the old one. It's hard to overstate the scope of the potential for subtle bugs which this simple rule removes. In a program with mutable variables, at any point in its execution, the value in a variable X is (in a mathematical sense) a function of the inputs to the program and every single decision made in the program's execution up till that point. With immutable variables, the value of X is a function of its definition, and any input values which are implicated in that.
We've effectively removed the homunculus' ability to shuffle objects between boxes: the shelves of the cabinet have been replaced with something more like museum display cases with glass lids.
Immutability exists outside the functional paradigm - strings, for instance, are immutable in Python - and is a fairly important idea right now in database design, where instead of keeping records which are regularly updated, we can simply keep an immutable list of transactions for each object, automatically giving an audit trail of when and where it changed.
The second restriction, which is only imposed on a small number of functional programming languages, is easy to state in terms of our workshop, but may be more difficult for non-programmers to understand. It's really two restrictions placed on the tools in the workshop, which are:
- a tool should always produce the same output objects if it's used on the same inputs, and
- a tool's only effects can be the value of its outputs.
The problem with our analogy is that actual physical tools pretty much obey these rules anyway. If you drill a hole in a blue block of wood with a drill, and then take a second blue block of wood and drill a hole in the same place, you'll get two identical blue blocks of wood with holes in them. In our analogy, tools represent functions, and, unfortunately, in software, functions can very often break both of these rules:
- functions can refer to variables other than their inputs (but which are still within scope), so their behaviour is not solely determined by those inputs;
- functions can have side-effects: they can modify variables other than their outputs, or maintain internal state variables which will affect the results of later usage on the same inputs
Side-effects are a rich source of bugs, as they are a way in which the overall state of an application can be changed in ways which are not particularly obvious to the coder. The conceptual problem with removing side-effects is that in one sense, especially if you've learned to program in the imperative paradigm, side-effects are how programs work. Your code accepts some inputs, performs calculations and shuffles data, and then does something (sets a variable, writes text to the command-line, makes a picture appear on the monitor); in other words, it has a side-effect. Understanding how programs can even make sense without side-effects is a big hurdle to using pure functional languages, and one of the reasons I'm writing this is an attempt to communicate what a radical idea it is, without having to go into too much technical detail (although I'm aware I might have gotten well past that point about a thousand words ago).
As we've placed more and more restrictions upon the actions which the poor homunculus can perform in the workshop, something interesting has happened, which should remind us that the homunculus itself has always been a bad analogy. Implicit in the metaphor is the idea of a decision-point, however fictional: picture the homunculus looking at the next line in its recipe book, reading it, and thinking: what comes next? Do I need to look up one of the countless boxes containing an object of type "wooden"? Which tool should I use? When I use it, what boxes should I change as it operates? And where should I put the results?
The effect of the restrictions has been to reduce the scope of the imaginary decisions which the homunculus has to make. It can't change an object and replace it in its box, because variables are immutable. It can't interrupt the operation of a tool to change anything, and it can't do anything with the output other than put them in new boxes. Steadily, what started out as feeling like a kind of puppet show or drama has become more like fitting together components or wiring up a machine.
Welcome to the machine
I can distinctly remember the first time I had this feeling, which was when I'd developed an intuition for the particular mathematical formalism which Haskell uses to represent operations with side-effects while remaining formally pure. (They're called monads, and are sufficiently hard to grasp that monad tutorials are their own mini-genre, and I don't want this essay to turn into one.)
The feeling was that in coding, I was no longer writing instructions, but describing the components of a machine: these components were well-defined, connected together properly, and, as a whole, were a sort of pipeline which accepted requests in at one end (the piece of software was a web server) and would provide the correct answers at the other.
This feeling is extremely good for a reason which is so simple and straightfoward that it sounds kind of dumb to express it: when you write a computer program, you are configuring a machine. You're not writing a script for an imaginary puppet.
This gets to a paradox which I am starting to feel is at the heart of coding, and what it is we are actually doing when we sit in front of a glowing rectangle for hours, swearing under our breath. The idea of a programming language which excludes state changes seems completely counter-intuitive to someone who's learned to program in the older paradings of imperative or object-oriented programming. The mathematical formalisms which languages like Haskell use in order to do stateful or unpredictable computations - random numbers, input from the external world, interaction with a human user, output to external data stores or equipment - are borrowed from some of the more abstract and recent fields in pure mathematics like category theory, and are non-trivial to understand, to the extent that one can do a lot of very productive programming in Haskell without being able to give a rigorous explanation of what a monad is, or even why it's necessary.
But, once you've reached the point I described at the start of this post, coding in a pure functional language starts to feel more intuitive and solid than any other form of coding, and when you then return to an imperative language, it feels incredibly wobbly and loose by comparison.
To take a simple example: the most elementary aspect of imperative programming is the idea that one statement is executed after another. In most languages, statements are separated by semicolons, like so:
do_first_thing(); do_second_thing(); do_third_thing();
Haskell's equivalent to this style of programming - which is obviously necessary in many contexts, especially when the application has to take something from a source, perform a computation on it, and then return it - is, in effect, to wrap the sequence of computations in a formalism which maps order-of-execution onto order-of-evaluation:
result = do_first_thing >>= do_second_thing >>= do_third thing
The key to understanding this is that instead of having a big, implicit, hard-to-visualise state which surrounds the sequence of statements, there's an actual piece of state value (in this case, a monadic wrapper) which gets passed from one statement to the next until the end result comes out.
Once you are used to a programming language in which a sequence of computations have to "plug into" one another (ie they have to pass a "state" token between them of the correct type), a language in which you can just do one thing after another and see what happens feels like driving with your seatbelt off.
One of the better analogies for monads, at least insofar as they are used to support sequential computations, is that they are "programmable semicolons". When I switched back into Perl after finishing a Haskell project, semicolons felt like broken, squishy monads.
(Goddamn it, I did end up writing a monad tutorial, after all.)
The paradox I'm talking my way around here is this: pretending that there's a little guy in the computer who does stuff for you, based on a set of instructions which you have written for him, is a natural way to try to understand what coding is about. But computers are intricate machines, and the metaphor of the homunculus is not only an imperfect fit for them, but can actively get in the way of configuring them to do things.
The homunculus seems like a slave, but I think my second reading of the metaphor is more truthful: the homunculus is a metaphor for the coder, lost in a maze of implicit state, having to decide (from all the possibilities, with whatever restrictions that make this set of possibilities more tractable) what the hell is going on here. It's better if we abandon it altogether, drop the pretense that we are writing a script for a puppet, and learn better ways to make our machines work.
Keep reading: Part 4 - The Algorithm