h1 12/09/2010 08:42:00 AM

I recently found a good series of articles on LtU, on the Ghosts of UNIX past. The author reflects on some successes and failures of Unix, trying to find patterns. I enjoyed reading the articles, they teach you a few things, and are generally good food for thought.

The first article describes the "full exploitation" pattern, successfully achieved in UNIX with files. The next articles describe various failures of UNIX design, classifying them in "conflated", "unfixable" and "high-maintenance" designs. While I don't have the culture to do the proposed exercises, I tried to relate this to programming language design -- which suprisingly hasn't been done on LtU.
  • Full exploitation is when you manage to re-use a concept a lot. This is good because it avoids inventing new things: it's hard enough to get a few concepts right. In programming languages, a good first-class notion of functions can be fully exploited to model control, event handlers, lazy computation, etc.
  • Conflated designs mix up several concepts. This is bad only if the two concepts cannot be separated. Otherwise, it's just syntactic sugar. Object oriented programming comes to mind: Methods are a conflated design, mixing functions and record members. Also, inheritance mixes record extension and dynamic binding -- which makes it hard for people to think about the latter aspect, a very important source of bugs.
  • Unfixable designs are unfixable designs. This is a bad definition, but sometimes it is hard to say why something is unfixable. Often, this has to do with the wide adoption of the wrong design. Note that conflated designs may often be fixable, essentially by creating ways to access conflated notions in isolation. I'm not sure what's a good programming language example here. Perhaps dynamic binding, probably stuck forever in emacs lisp. (By the way, a fun fact according to Olivier Danvy: dynamic binding was born from exagerated exploitation of a single stack in the implementation of Lisp -- in those old days, people were psyched about stacks and got a bit carried away.)
  • High-maintenance designs have nothing wrong in themselves but interfere with other things in such a way that requires a lot of care and work in those areas. With programming, various difficult things come to mind: state, concurrency, dynamic binding... Those aspects make it very hard to reason locally about a piece of code. In themselves, state and concurrency are cool, but their interaction leads to many problems.

More than programming languages, liquidsoap was in the back of my mind. But there does not seem to be interesting examples of those patterns in liquidsoap. In fact, it is not an exceptionally good design. It's just a first shot at making an end-user streaming application that is not just one object/tool but rather a system/language for building tools. Liquidsoap introduces an abstract notion of source, which is good, but does not fully exploit it. For instance, blank detection operators are black boxes that encapsulate the computation of the average audio volume, but a more flexible and efficient design would be achieved by isolating the computation of the volume stream as a source, then write operators that act depending on the characteristics of the volume source. Liquidsoap does not have unfixable designs because we give ourselves the liberty to change things when needed. It does have high-maintenance designs: the source protocol is complex, and we've had many bugs because it wasn't properly followed.

Beyond those not-so-good design examples, liquidsoap is broken in several ways. But I'm comfortable to admit it because it works good enough for many people and we know how to fix it. The code name for that exciting new line of work is, quite straightforwardly, Liquidsoap 2.0. Until now, we've been writing sources by hand in OCaml. Liquidsoap users write scripts in our own scripting language, and can only compose sources, never write a new one. In other words, Liquidsoap is an interpreter for a dedicated script language that manipulates an abstract/opaque notion of source. But liquidsoap 2.0 will be a compiler that lets you write sources, and not only compose them, and will produce optimized code. It will be as simple as it is today for beginners, but will enable much more for power users. Importantly, it will allow us developers to not write as much bug-prone code. There is a lot to say about what could be in liquidsoap 2.0, but I'll start with a simple down-to-earth story... next post!

Libellés : , ,

0 commentaires:

Un commentaire ?

< Accueil