Comment by patterns | Hacker News Reader

Comment by patterns | Hacker News Reader

I wonder whether the migration to Chez Scheme will improve the performance of Dr. Racket which consumes a lot of memory and feels very sluggish.

I have been experimenting with Racket for quite a while and I appreciate the effort that went into making the language extensible. That being said, I wish that the community would have embraced "object-oriented" techniques for building the VM - Ian Piumarta's COLA (Combined Object-Lambda Architecture) comes to mind [1]. I think this would have saved them a lot of performance troubles (but this is mere speculation) and would make the system far more flexible and pleasant to use.

While it's comforting to have a very powerful macro system at your fingertips to change the semantics of the language when needed, many of the macros in the base language and libraries could be eliminated with a more convenient default notation for closures and message-passing (to objects simulating control structures like in Smalltalk).

Racket has one of the best module systems I have worked with but modules cannot be (easily) parameterized. Research on systems such as Self and more recently Newspeak [2] gives ample of demonstration of the benefits (conceptual integrity, security) of having modules as objects.

What Racket also lacks is a good meta-object system with highly reflective capabilities (of say a CLOS or Smalltalk system). This makes it difficult to build tooling such as inspectors or browsers.

I hope that in the future these issues will be addressed.

[1] https://www.piumarta.com/papers/colas-whitepaper.pdf

[2] https://newspeaklanguage.org/

I would absolutely love to see such a thing (e.g. VPRI's 'frank' system), but it's quite a different approach than Scheme, and I wouldn't want to see PLT/Racket take such a drastic course change while there's still so much interesting work to be done in the Scheme world (e.g. Kernel's F-expressions are powerful, but hard to optimise; work on reflective towers and multi-stage programming is pushing languages beyond the traditional AOT/JIT dichotomy; and so on).

From my understanding, Piumarta's main trick with COLA is fast dynamic dispatch through a uniform interface (putting pointers to vtables at offset -1), and implementing vtables via interface (so they can be replaced). It's a cool approach, but it's also a case of 'everything can be solved by adding a layer of indirection'; and I've not sat down and thought through how these ideas might apply to different paradigms, and what might be unique or common to them all.

If I remember correctly, Adele Goldberg once remarked that "in Smalltalk everything happens somewhere else". Although there has been ongoing work to collapse the many layers of indirection you find in message-passing systems, I am not convinced that throwing more tools at the problem will address the deeper underlying issues. At some level, you need more concise descriptions of your system and operators to shape its structure and organization; perhaps even compute different organizations. This is very difficult to accomplish in purely object-oriented systems/languages that have no direct and convenient "algebraic" correspondence for composing complex communication structures and specifying relationships such as inheritance. For example, I was very taken by the idea of class definitions as expressions, having seen it in Racket first many years ago.

On the other hand, I believe that when it comes to low-level "machine" work, objects are a good abstraction to model components such as activation records such that they can be uniformly explored and modified (on the fly). But this is perhaps a moot point.

Over the years, I have studied many object/component-oriented systems and come up with a sizable catalogue of message-exchange patterns, plumbing and machinery. My hope is that at some point, this can be crystallized into a language or calculus for specifying systems; and Scheme/Racket is certainly a good language to think about these issues from another perspective (which is worthwhile to preserve).

So I guess, I understand your point. Thanks for raising it, much appreciated.

Oberon and Maude I think make a good case for (parameterised) modules, both from being able to reason mathematically and their use at low level OS system.

I've been very slowly trying to combine the work in Maru with Composita to yield a modular and deterministic lisp (Composita uses managed memory without GC).

http://concurrency.ch/Content/publications/Blaeser_Component...

Naturally you'd want to layer something like Shen or Maude on top of this, to provide the equational proof-checking. The K language is a good example - it provides facilities to model new languages and semantics like racket.

http://fsl.cs.illinois.edu/index.php/K-Maude:_A_Rewriting_Ba...

I have no CL experience but doesn't a CLOS system require multiple dispatch?

With macros you can easily build your own multiple dispatch, it's really not a concern.

Well, there's at least the concern that multiple dispatch results in unsafe edge cases across module boundaries, apparently.

So, multiple dispatch as a library yes (and racket has that), but probably not ok in the core language?

It's multiple dispatch, but the meta-object protocol runs deeper than that (e.g. :before, :after, :around )

CLOS is inherently multiple dispatch, but Flavours that are the most prominent precursor of CLOS were single dispatch. Interesting consequence of that is that IIRC traditional implementation of CLOS/MOP does not really use multiple-dispatch internally.

Reply on news.ycombinator.com