Small is Beautiful, Large is Useful, and Scheme is Both

They say, Scheme is small and this is good.

Have you heard of X? No? It is the smallest computational basis. It is a single function that can compute everything a Turing machine can compute; a Church lambda calculus; a Post model; a RAM; a what-have-you model of computation. Indeed, X is so simple that two equations suffice to specify it completely [Barendregt, page 166]. Imagine that: a complete language report in two lines; a compiler that fits in a few K instead of Ms; no more arguments about smallness.

Small alone can't be any good. If you used X alone, your programs would be the size of the universe or something like that. That's what the theory of computability teaches us [Church and Turing]. Adding LAMBDA and a few primitives to get a pure functional language isn't good enough either. That's what the theory of expressiveness shows [Felleisen; Mitchell; Riecke]. And, using an R5RS Scheme to build large systems with many people at a dozen sites isn't doable either. That's what the PLT experience determined.

When we set out to construct DrScheme using MzScheme, we also conducted an experiment:

Could we really build a graphical system that manages (shared) resources and that provides excellent error feedback with just plain Scheme?

Could we just add enough libraries to do all this? Or would we have to change the kernel of the language? As much as we tried to keep MzScheme small, it became clear quickly that we needed exceptions, structures, module-like features, a mechanism for concurrency, a way to manage resources such as windows, tcp connections, and so on. The list isn't infinite but it is much longer than I expected. Our "Revenge of the Son of the LISP machine" paper is a good summary for the state of the art around 1999 [Flatt and Son].

As language designers we stepped back time and again to look at our monster. What could we remove? What would we have to add in response? For some five years, we had first-class modules (units) and first-class classes in the core of the language. We had almost banned macros. They were so ugly I stopped teaching about them because I did want to use our own dog food in my courses but I couldn't stomach the macro system. It was such a step back from Eugene's extend-syntax. But then Matthew figured out the next big step in macro and module technology [Flatt, You Want It When?]. And with that out went units and classes from the core and many other things. So we learned lessons, and we need to keep building systems to learn more.

I have no question that the idea of Scheme is beautiful. At the same time, I have also learned that if I wish to use this beautiful idea in practice, I need to add the ingredients that it takes to build large systems. R6RS reflects this insight, and I am happy about it.

R6RS is "perfect"

When I read the "side by side" and "head to head" descriptions of the alternatives facing the Scheme community (see Comp.Lang.Scheme and the R6RS mailing list), I am wondering which one is which and which one is better.

• Is it really good that Scheme (the spec) doesn't support a module system?
• Is it really good that almost all major implementations support their own version of a module system?
• Is it really good that programmers can't even leave the module structure intact when porting code?

Imagine your own similar questions and add them here. We have lived in a side-by-side universe for a long time, and there are quite a few programmers who have suffered from this not-really-the-same-language problem. Besides the module system, there are other not-quite-the-same-but-related features that implementations have and programmers wish to use.

The R6RS process has pushed several major implementors/implementations to agree on a design for module systems and other constructs. Their report declares that they are ready to put a large amount of work in to get from r5rs to r6rs. I believe that this step would help the community in several arenas, listed in increasing order of relevance:

1. the academic publishing business
2. the fund raising business
3. adapting each others innovations
4. supporting programmers who learn on one and switch to another implementation
5. supporting commercial programmers who need reassurance that there is more than one implementation and implementor [ever attended Commercial Uses of Functional Programming?]

Is the document perfect? Is every construct exactly the 'right thing'? Of course not! Guy and Gerry revised their first Scheme report because they didn't get it 'right'. R3RS and R4RS and R5RS revised flaws in R(n-1)RS because the authors/editors didn't get it 'right'. It is extremely difficult, and usually impossible, to get the design of a complex artifacts (such as a programming language) 'right' the first time. In these cases, it's all about the feedback loop and revising your design based on observations. (Remember the 'science' part in the name of our discipline?) Indeed, 'right' doesn't exist; what exists is 'most pragmatic and internally beautiful,' and nothing else.

Our choice is quite simple: move forward as a community with some amount of convergence (r6rs) or split into dozens of mutually incompatible sub-communities (status quo, including SRFIs).

Also posted as "R6RS is perfect" at the R6RS discussion list.