Working on open source projects with other Prolog enthusiasts is a bit different because that crowd self-selects for being good with Prolog.
It had numerous setbacks. The Japanese thought they could parallelize Prolog programs for their Fifth Generation Computing Project in the 1980s but found out quickly that you couldn't. They made a language called KL1 which was parallelizable, but it wasn't as good as Prolog in other respects.
The ability to implement simple parsers in the Prolog using its evaluation process impressed me a lot when I didn't know much about parsers, now that I know how to implement parsers it doesn't impress me much.
You can write mixed logical/imperative problems in Prolog but boy is it awkward.
When I got interested in RDF circa 2010 I was interested in Datalog (a pure logical language) but found nobody else seemed interested in it and it was hard to find literature on it. That situation has changed a lot as Datalog is a pretty clear way to make database query code composable.
Production rules engines (say Drools) are another "old A.I." technology that is largely forgotten even though they are heavily used in banks and a few other corners of the business world. These implement “forward chaining” inference distinct from the “backward chaining” inference implemented in Prolog. Between RETE engines and effective indexing structures it is easy to handle 1000x's more rules in your knowledge base in the 1980s but production rules never got standardized like C, FORTRAN or COBOL and no really general answers have come up for questions like controlling the order of execution when that matters.
One of the lessons I'm still trying to absorb is how convinced we all were (and many still are) that there must be a ton of implicit parallelism in the world, but once we went looking for it, it turned out there was hardly any.
I've been chewing on this for years, trying to figure out whether this is something true about the world, true about the problems we try to solve, or because of the pervasive use of the imperative paradigm which is highly ordered and makes it very easy to impose ordering constraints. Now, clearly, the latter is a non-trivial component... but I still struggle with whether it is a full answer, because when people sat down with clean sheets of paper and tried to solve problems with extensive parallelism, they've largely failed to get more than low single-digit speedup factors. Non-imperative paradigms like logic programming have been tried, especially in these projects.
It isn't exactly news that our intuitive beliefs about our code and the real characteristics it has is quite out of whack. This underlies the pervasive advice to just go grab a profiler whenever you're trying to accelerate some code because even experts in the field with decades of experience are routinely completely wrong about what is slow in some bit of code. This is one particular aspect I'm still struggling with. It still feels like there should be so much more parallelism available....
There are some cases (make and the Spring Framework) where the user specifies the dependencies between things and the framework does a topological sort to figure out what order to do them in.
Even though this could be a basis for a whole paradigm of programming, professional programmers usually don't find it hard to figure out how to order things so there is little motivation to address this problem systematically. However, I think it is on the list of problems that non-professional programmers (say the subject matter expert who wants to learn Python to put their skills on wheels) get hung up. Instead it's seen as a special case that makes high-complexity systems scalable (... Spring lets you describe the parts of a system and how they are related without writing structurally unstable initialization code.)
I think most "business logic" is fairly linear most of the time, and that even if it's highly branched, usually you don't have a large number of concurrently executing units of logic. You do have things like iterating/mapping over arrays, but that's precisely what I think is generally not "high impact" to parallelize. And when it is high impact, we do in fact have a rich space of tools available in the programming and software world, depending on the nature of the task at hand.
I know nothing why this particular Prolog parallelization project failed, but the above is generally my intuition about parallelism in computer programming, so I imagine it's at least somewhat relevant in this case.
The problem with logic programming is the 'logic' part.
Modern imperative and functional programming languages are constructive.
Logic programming is not, and the expressive power varies with the exact logic being used.
Elementary logic gives you the following intuition. Propositional logic -- easy. First order logic -- easy (with caveats). And (some kinds of) second order logic -- only easy if you get lucky with the problem you are trying to solve.
For logic based programming languages and systems, both the language implementer and the programmer have to be careful about supporting and using language constructs which boil down to tractable computation.
This is much more difficult than it seems like.
For example, when using SMT solvers you learn quickly that multiplication and division with constants is very fast while the same operations with variables can be intractable.
Though to be fair, both of these libraries are in Lisp languages, which inherently support the design pattern of slipping into/out-of DSLs.
[1] http://nikodemus.github.io/screamer/ [2] https://github.com/clojure/core.logic
Having these tools close at hand the same way we have regex and (more recently) PEGs for grammars definitely makes it more likely that people will reach for them when appropriate, and exposure and familiarity to the practices can grow.
I suspect part of the problem is just that minikanren is primarily a learning tool and doesn't prioritize practical ergonomics. I haven't used core.logic but I've used minikanren in this embedded way and it's difficult to do cleanly in ways that have nothing to do with logic programming per se. I may be off but I think it's something that having a high quality professional strength implementation in the stdlib could actually help with.
Exhaustively searching an exponential space with no heuristics for reducing the search space is obviously gonna be way easier in a familiar language than in a new language.
The one use case I do see for prolog and similar tools in the modern day is for Constraint Satisfaction programming. Obviously, exhaustively searching an exponential solution space won’t be better for that, but typically you can use concepts like Edge and Arc consistency (which are core to CSP solvers) to greatly reduce the actual search space. Implementing those algorithms from scratch in an imperative language is annoyingly hard, so reaching for optimized, generalized implementations can be a good choice.
For something like a Sudoku solver, edge/arc consistency can greatly improve compute times. If you’re good at sudoku, you are already implementing these techniques when solving puzzles, just like good chess players implement minimax search without having to be told what minimax search is.
https://www.sciencedirect.com/topics/computer-science/arc-co...
Huh? This isn't remotely true especially today, and for the type of programming most do, where people work with functions and classes from libraries, with each line doing all kinds of O(?) stuff under the hood, and applying those to collections of objects, and so on.
Certainly, people manage to make poorly performing things even so, but at least at the base level, your primitives may be stupid, but they are generally fast.
The logic programming world works in a default space of O(n) operations, that stack together more freely than the imperative world, and that gives easy access to O(2^n). Since this is essentially impossible, a great deal of work is done to try to get that down, but you're always intrinsically starting behind the eight-ball. It is easier to work up from O(1) operations than to write an exponential or super-exponential algorithm and then try to trim it back down to a decent complexity for a normal programmer.
I think this is the root cause as to why logic programming is generally not something we see a lot of. It's like a wild stallion; it may be powerful but the amount of effort you pour into just keeping it under control may exceed any benefit you could get.
It isn't useless, of course. Amazing work has been done in the field of SAT solvers, and there's certainly a niche for it. The problems that are intrinsically higher-order polynomials or (technically) exponential, well, they are what they are and if you're going to be stuck in that world, logic programming may offer you a much better toolset than conventional programming on its own. But there was a hope a long time ago, in the Prolog era, that it could become part of the normal toolkit of general purpose programming, and I don't think that will ever happen, because of this line of logic.
This is a bit tangential to the article, it's just what I happened to read that finally crystallized this in my mind.
[1]: https://accidentallyquadratic.tumblr.com/