Crate holmes [−] [src]

Holmes

Holmes is a Datalog inspired system for binding codependent analyses together.

Tutorial

Basic Datalog

If you are already familiar with logic languages, this section will likely be straightforwards for you, but it may still be useful to provide an overview of basic functions and syntax.

Datalog is a forward-chaining logic language. This means that a program written in Datalog consists of a set of rules which "fire" whenever their requirements are met which operate on a database of facts.

Predicates

A predicate represents a property on a list of typed values. For example, to express the distance between two cities in miles, we might write

predicate!(distance(string, string, uint64))

N.B. while this code is being built via doctests, there are a few lines of support code above and below being hidden for clarity. See the complete example at the end of the section for a template.

Facts

Facts are formed by the application of predicates to values. Continuing with the example from before, we can add a fact to the database for the predicate we defined

predicate!(distance(string, string, uint64));
fact!(distance("New York", "Albuquerque", 1810))

Rules

Rules are formed from a body clause and a head clause. When the rule body matches, variable assignments from the match are substituted into the head clause, which is then added to the database. Here, we might want to add the symmetry property to our previous example, e.g. "If the distance from A to B is N, then the distance from B to A is also N".

predicate!(distance(string, string, uint64));
fact!(distance("New York", "Albuquerque", 1810));
rule!(distance(B, A, N) <= distance(A, B, N))

In a rule or a query (in the next section), the possible restrictions on each slot are:

Unbound: [_]
Constant Equality: (value)
Variable unification var

Queries

Now that the database has more facts in it than we started with, it makes sense to be able to query the database and see what is inside.

holmes_exec!(holmes, {
  predicate!(distance(string, string, uint64));
  fact!(distance("New York", "Albuquerque", 1810));
  rule!(distance(B, A, N) <= distance(A, B, N))
});


let mut res = try!(query!(holmes, distance(A, [_], [_])));

assert_eq!(res,
           vec![vec!["Albuquerque".to_value()],
                vec!["New York".to_value()]]);

Recursive Rules

Let's go one step further, and use a rule to check connectivity between cities, based on the facts in the database. We want to express "If A connects to B, and B connects to C, then A connects to C".

holmes_exec!(holmes, {
  predicate!(distance(string, string, uint64));
  fact!(distance("New York", "Albuquerque", 1810));
  fact!(distance("New York", "Las Vegas", 2225));
  fact!(distance("Las Vegas", "Palo Alto", 542));
  fact!(distance("Rome", "Florence", 173));
  rule!(distance(B, A, N) <= distance(A, B, N));
  predicate!(connected(string, string));
  rule!(connected(A, B) <= distance(A, B, [_]));
  rule!(connected(A, C) <= connected(A, B) & connected(B, C))
});
assert_eq!(try!(query!(holmes, connected(("Rome"), ("Las Vegas")))).len(),
           0);
let mut res = try!(query!(holmes, connected(("Palo Alto"), x)));
assert_eq!(res,
           vec![vec!["Albuquerque".to_value()],
                vec!["Las Vegas".to_value()],
                vec!["New York".to_value()],
                vec!["Palo Alto".to_value()]]);

Complete Example

Finally, just for reference (so you can actually write your own program using this) here's the unredacted version of that last example:

#[macro_use]
extern crate holmes;
extern crate tokio_core;
use holmes::{Engine, MemDB, Result};
use tokio_core::reactor::Core;
use holmes::pg::dyn::values::ToValue;
fn f () -> Result<()> {
  // Holmes uses the `tokio_core` event loop in order to schedule work
  // amongst various rules and enable asynchronous processing.
  // Unless you specifically want to do something async, this just
  // means you need to pass in a handle as shown here, and call
  // `core.run(holmes.quiesce())` before any queries to wait for
  // the engine to finish running.
  let mut core = Core::new().unwrap();
  let mut holmes_own = Engine::new(MemDB::new(), core.handle());
  // For the moment, the `holmes_exec` macro needs a &mut ident. I'll
  // try to make this more flexible in the future.
  let holmes = &mut holmes_own;
  holmes_exec!(holmes, {
    predicate!(distance(string, string, uint64));
    fact!(distance("New York", "Albuquerque", 1810));
    fact!(distance("New York", "Las Vegas", 2225));
    fact!(distance("Las Vegas", "Palo Alto", 542));
    fact!(distance("Rome", "Florence", 173));
    rule!(distance(B, A, N) <= distance(A, B, N));
    predicate!(connected(string, string));
    rule!(connected(A, B) <= distance(A, B, [_]));
    rule!(connected(A, C) <= connected(A, B) & connected(B, C))
  });

  // Now we force the database to compute until quiescence. Otherwise,
  // any query results may be partial.
  core.run(holmes.quiesce()).unwrap();

  assert_eq!(try!(query!(holmes, connected(("Rome"), ("Las Vegas")))).len(),
             0);
  let mut res = try!(query!(holmes, connected(("Palo Alto"), x)));
  // Order is not gauranteed when it comes back from the query, so I
  // sort it in the example to get the doctest to pass. `Value` only has
  // `PartialOrd` implemented for it, since there isn't a clean comparison
  // between `Value`s of different types, so I just default to `Greater`.
  res.sort_by(|x, y| x.partial_cmp(y).unwrap_or(
    ::std::cmp::Ordering::Greater));
  assert_eq!(res,
             vec![vec!["Albuquerque".to_value()],
                  vec!["Las Vegas".to_value()],
                  vec!["New York".to_value()],
                  vec!["Palo Alto".to_value()]]);
  Ok(())
}
fn main () {f().unwrap()}

Extensions

While Datalog itself is interesting, writing yet-another-datalog engine is not the goal of this project. Next, we'll go over some of the new features of this system.

Functions

Normally, logic languages expect the computation to be encoded as rules only (or in special cases, as external predicates). In order to allow the user to write things which make more sense as traditional code, we allow the binding of functions:

func!(let f : uint64 -> uint64 = |x : &u64| {
  x * 3
})

In this case, we have declared a function called f, said that it takes as input a uint64, and should output a uint64. The type of the input to the function should be the output of the .get() call of the relevant value, which will usually be a reference to the rust equivalent of the type. The output should be a value which .to_value() will convert to a correctly typed Value.

Additionally, the type system allows for tuples and lists. Tuple types are denoted (t1, t2), and list types are denoted [t]. Lists and tuples will be unpacked through by the func! macro, so a function with a [uint64] input would expect to take a Vec<&u64>, and a function taking (string, uint64) would expect to take a (&String, &u64). For example:

func!(let replicate : (string, uint64) -> [string] =
  |(s, n) : (&String, &u64)| {
    let mut vec : Vec<String> = Vec::new();
    for i in 0..*n {
      vec.push(s.clone());
    };
    vec
  }
)

Where Clauses

Telling Holmes about functions isn't useful without a way to use them. Where clauses are a way to perform a transformation on the data after the map, but before the head clause is produced and sent to the database.

Extending the example from earlier, we might want to generate a distances for the connection paths we found.

holmes_exec!(holmes, {
  predicate!(distance(string, string, uint64));
  fact!(distance("New York", "Albuquerque", 1810));
  fact!(distance("New York", "Las Vegas", 2225));
  fact!(distance("Las Vegas", "Palo Alto", 542));
  fact!(distance("Rome", "Florence", 173));
//rule!(distance(B, A, N) <= distance(A, B, N));
  predicate!(path(string, string, uint64));
  rule!(path(A, B, N) <= distance(A, B, N));
  func!(let add : (uint64, uint64) -> uint64 = |(x, y) : (&u64, &u64)| {
    x + y
  });
  rule!(path(A, C, NSum) <= path(A, B, N1) & path(B, C, N2), {
    let NSum = {add([N1], [N2])}
  })
});

The astute reader will notice there is something wrong with this example. It builds, and it runs, I'm not trying to mess with you while teaching. However, the last rule we added (which does the sum of the distances) will loop forever if there is any cycle in the distance predicate. This is why I commented out the rule flipping the distance direction around, as this would cause this example to run infinitely.

Normally in Datalog, we have a termination property - no matter what rules or facts you add, the database will always eventually stop growing. This proof follows from the inability of a rule firing to introduce a new value, which means there are only a finite number of derivable facts. With the addition of where clauses, we lose this property, because new values can appear, as per the add function above.

However, we also add other kinds of binds to the where clause that can help the programmer control this kind of situation.

N.B. the postgres backend doesn't currently support list persistence, so if you wanted to use a list in a predicate, you'd actually need to make a custom Path type and value that knew how to store itself.

Binds

Variable binding

This is as in the inital example. They are written let x = expression, and simply bind the expression to the variable.

Destructuring

This kind of bind is basically just shorthand to prevent the need for functions like access_tuple_field_1, access_tuple_field_2. It is written let (x, y, z) = expression

Value binding

This is the first unusual kind of binding, and the one we can use to fix up the previous example. Value binds are written let (expr) = expr2. If expr and expr2 evaluate to the same value, this expression has no effect. However, if expr and expr2 differ, the variable assignment currently generated by the where clause will stop.

To fix the previous example, we can track the path we've gone through thus far, and store it in an additional slot in the path predicate. Then, in the where clause for adding a new step to the path, we can check for membership in the existing path. If it is present, we can use a value binding to stop pursuing this avenue. If it is not present, then we can proceed as before.

try!(holmes_exec!(holmes, {
  predicate!(distance(string, string, uint64));
  fact!(distance("New York", "Albuquerque", 1810));
  fact!(distance("New York", "Las Vegas", 2225));
  fact!(distance("Las Vegas", "Palo Alto", 542));
  fact!(distance("Rome", "Florence", 173));
  rule!(distance(B, A, N) <= distance(A, B, N));
  predicate!(path(string, string, [string], uint64));
  func!(let two_vec : (string, string) -> [string] =
    |(x, y) : (&String, &String)| { vec![x.clone(), y.clone()] });
  rule!(path(A, B, steps, N) <= distance(A, B, N), {
    let steps = {two_vec([A], [B])}});
  func!(let add : (uint64, uint64) -> uint64 = |(x, y) : (&u64, &u64)| {
    x + y
  });
  func!(let append : (string, [string]) -> [string] =
    |(x, y) : (&String, Vec<&String>)| {
      let mut out : Vec<String> = y.into_iter().cloned().collect();
      out.push(x.clone());
      out
    });
  func!(let mem : (string, [string]) -> bool =
    |(needle, haystack) : (&String, Vec<&String>)| {
      haystack.contains(&needle)
    });
  rule!(path(A, C, path2, NSum) <= path(A, B, path, N1)
                                 & distance(B, C, N2), {
    // If we've already walked over C, we aren't interested
    let (false) = {mem([C], [path])};
    let path2 = {append([C], [path])};
    let NSum = {add([N1], [N2])}
  })
}));
let mut res = query!(holmes, path(("New York"), dest, [_], dist))?;

assert_eq!(res,
           vec![
                vec!["Albuquerque".to_value(), 1810.to_value()],
                vec!["Las Vegas".to_value(), 2225.to_value()],
                vec!["Palo Alto".to_value(), 2767.to_value()],
               ]);

Iteration

The last kind of bind is the iterative bind. This works similarly to the List monad in Haskell if you are familiar with it, but you don't need to know anything about that to proceed.

An iterative bind is written let [x] = expr, where the expression should evaluate to a list-typed value. When this bind is run, the set of possible answers splits into a different instance for each value in the list. So, if we had

rule!(q(x, y) <= p(y), {
  let [x] = f(y)
})

it would first find all y such that p(y), and then for each of them, it would apply f and get a list. Imagine that f just returns a list of y and y + 1, and that p is only populated with p(1) and p(2).

The match would produce the possible assignment sets y = 1 and y = 2. After running the where clause, the first one would become x = 1, y = 1, x = 2, y = 1, and the secould would become x = 2, y = 2, x = 3, y = 2 . This ends with the database containing q(1, 1), q(2, 1), q(2, 2), q(3, 2).

That example is somewhat abstract, but hopefully it illustrates the multiplicative effect of the iteration bind. The iteration bind can also be used to terminate early a rule, similar to the value bind, by iterating over an empty list. If an iteration bind is used multiple times in a where clause, it will operate on each of the new answer sets from the previous iteration bind individually.

As a more concrete example, say we wanted to define a predicate which contained all sities that might be used on a path from New York to Palo Alto. We can take the example from earlier and add:

predicate!(on_the_road(string, string, string));
rule!(on_the_road(A, B, stop) <= path(A, B, path, [_]), {
  let [stop] = [path]
})
let mut res = query!(holmes, on_the_road(("New York"), ("Palo Alto"),
                                         stop))?;

assert_eq!(res,
           vec![
                vec!["Las Vegas".to_value()],
                vec!["New York".to_value()],
                vec!["Palo Alto".to_value()],
               ]);

Reexports

pub use engine::Engine;

pub use pg::PgDB;

pub use mem_db::MemDB;

Modules

edsl	Holmes EDSL
engine	Holmes/Datalog Execution Engine
fact_db	This module defines the interface which a fact database must present to be used as a backend by the Holmes engine.
mem_db	This is a memory mock for the fact database interface.
pg	Postgres-based Fact Database
simple	You likely don't want to use this module - its primary purpose is to make benchmarking and testing easier to do in practice.

Macros

bind_match	Constructs a bind match outer object.
clause
clause_match	Generates a `MatchExpr` from a representation
db_expr
fact	Stores a fact with the `Holmes` context.
field
func	Registers a native rust function with the `Holmes` object for use in rules.
hexpr	Generates an expression structure
holmes_exec	Shorthand notation for performing many actions with the same holmes context Analogous to a weaker version of the `Reader` monad which cannot return values.
htype	Converts an EDSL type specification into a Holmes type object Takes the name of a variable containing a holmes object as the first parameter, and a type description as the second.
predicate	Registers a predicate with the `Holmes` context.
query	Runs a datalog query against the `Holmes` context
rule	Adds a Holmes rule to the system
typed_unpack	Given a value and a type it is believed to be, unpack it to the greatest extent possible (e.g. unpack through tupling and lists)
typet_boiler
typet_inner
typet_inner_eq
valuet_boiler