Announcing Canrun: A logic programming library inspired by the *Kanren family

I keep bumping into logic programming, especially in the form of the *Kanren family. In keeping with tradition, this is my attempt to implement it in Rust, my current hobby language of choice.

After a few false starts, I got a not completely horrific proof of concept running. Then I rewrote it once or twice to improve usability, make it internally type safe and add a form of constraints. The sordid history is buried in git.

The end result? I'm pretty satisfied with the learning experience I got out of it, so any value in the actual artifact is just gravy.

An Introduction

I'm still a novice, so take everything I say with a grain of salt.

Logic programming is weird. You write statements in (somewhat) arbitrary order and query the zero or more possible results if everything you said before holds true.

Things can be much fancier and terser in core.logic or Prolog, but this is what I've got so far.

use canrun::{Goal, both, unify, var, example::I32};

let x = var();
let y = var();
let goal: Goal<I32> = both(unify(x, y), unify(1, x));
let result: Vec<_> = goal.query(y).collect();
assert_eq!(result, vec![1])

Breaking it down...

Imports

use canrun::{Goal, both, unify, var, example::I32};

First are some imports, including an example domain that represents all of the types that will be valid in our little logic program. In this case, only i32.

Logic Variables

let x = var();
let y = var();

These are typed "logic variables" (specifically LVar<i32>), which can be unified with other values in our program.

Goals

let goal: Goal<I32> = both(unify(x, y), unify(1, x));

The both and unify functions are Goal constructors. A Goal represents one or more relations between logical values. In other words: A Goal is a set of predicates that we wish to prove. They can be evaluated to generate zero or more possible results. Note that it is constrained by the I32 domain through its type parameter.

unify(x, y) essentially declares that x and y are equal to each other. Or, more precisely, that they must unify with each other. This distinction matters for more complicated values. We'll come back to that.

both combines the two unify goals into a goal that will only succeed if both sub goals succeed.

Querying

let result: Vec<_> = goal.query(y).collect();
assert_eq!(result, vec![1])

To get results from a Goal, we can query it with one or more logic variables to get an iterator of possible results. We can take the first result with .nth(0), .collect() them all, or use any of the other Iterator trait methods.

For example, querying the goal either(unify(x, 1), unify(2, x)) for the value of x would return two possible results: vec![1, 2].

Getting More Advanced

Unification is quite a bit more interesting than simple equality when applied to structures.

The library provides a few "logically superpowered" structures such as LVec and LMap. Note that these are experimental even by Canrun standards.

use canrun::{Goal, val, var, all, unify};
use canrun_collections::{lvec, example::Collections};

let x = var();
let goal: Goal<Collections> = unify(
    lvec![x, 2, 3],
    lvec![1, 2, 3],
);
let results: Vec<_> = goal.query(x).collect();
assert_eq!(results, vec![1]);

In this case, we actually plucked a value out of a structure by unifying a compatible structure containing our LVar.

Additional specialized collection types and goal constructors are available, with more coming as needs arise and ability allows.

Member

let goal: Goal<Collections> = member(x, lvec![1, 2, 3]);
let results: Vec<_> = goal.query(x).collect();
assert_eq!(results, vec![1, 2, 3]);

Subset

let goal: Goal<Collections> = subset(
    lvec![1, x],
    lvec![1, 2, 3],
);
let results: Vec<_> = goal.query(x).collect();
assert_eq!(results, vec![2]);

What's Next?

Apart from random needs based additions...

Ergonomics

The biggest obvious pain point I see at the moment is how annoying it is to implement many of traits such as UnifyIn and ReifyIn. I think 99% of the use cases can be handled with a derive macro.

Even Fancier DSL

The function based API is not bad at all, but I still dream about doing some sort of fancy custom parser macro magic overkill to unlock something like:

let goal = logic! {
    y = x
    x = 1
}
goal.query(y)

This is mostly so I can type logic!

General Correctness

I've approached this as an engineering problem, with very little theoretical rigor or deep understanding of the prior art. This has been fun and enlightening, but I have no doubt that there are holes.

Conclusion

Sometimes the types get a little scary, but overall I'm quite pleased with how fluent Rust is able to be with a little architectural care and feeding.

How much I'll actually work on this project is completely dependant on how applicable it actually ends up being to my other experiments. Or how much I keep nerd sniping myself.

Next: A deeper dive into how Canrun actually works and how it came to be.