#relation #feedback #dataset #single-threaded #shifting #input #standing

standing-relations

Standing relations over a shifting dataset optimized for 'feedback loop' scenarios

3 releases

0.1.2 Jul 23, 2022
0.1.1 Jul 23, 2022
0.1.0 Jul 15, 2022

#8 in #feedback

42 downloads per month

MIT license

105KB
3K SLoC

Standing Relations

Overview

This crate provides an interface vaguely similar to differential-dataflow for creating standing relations over a shifting dataset. Critically, unlike differential-dataflow the operators here are single-threaded and optimized for fast turnaround rather than high throughput. That is, this package is intended to be used in "feedback loop" scenarios where the calling code uses the output to determine which inputs to feed in next.

Getting Started

To get started, create a CreationContext:

use standing_relations::CreationContext;

let mut context = CreationContext::new();

And some inputs:

let (mut input1, relation1) = context.new_input::<(char, usize)>();
let (mut input2, relation2) = context.new_input::<(char, String)>();

Set up your relational operations:

let foo = relation2.save();
let bar = relation1.join(foo.get());
let baz = foo
    .get()
    .map(|(_, s)| (s.as_str().chars().next().unwrap_or('x'), s.len()));
let qux = bar.map(|(c, n, s)| (c, n + s.len())).concat(baz).distinct();
let arrangement: Output<(char, usize), _> = qux.into_output(&context);

Begin inserting data. To do this, you must first change your CreationContext into an ExecutionContext by calling CreationContext::begin. This tells the system that your relational graph is fully built and you won't be making any more changes to it:

let mut context = context.begin();

input1.add(&context, ('a', 5));
input1.add(&context, ('b', 6));
input2.add(&context, ('b', "Hello".to_string()));
input2.add(&context, ('b', "world".to_string()));

Commit your changes:

context.commit();

Read the output:

assert_eq!(
    &*arrangement.get(&context),
    &HashMap::from_iter(vec![(('H', 5), 1), (('b', 11), 1), (('w', 5), 1)])
);

Make some more changes (and commit them):

input1.remove(&context, ('b', 6));
input2.add(&context, ('a', "Goodbye".to_string()));
context.commit();

Read the new output:

assert_eq!(
    &*arrangement.get(&context),
    &HashMap::from_iter(vec![
        (('G', 7), 1),
        (('H', 5), 1),
        (('a', 12), 1),
        (('w', 5), 1)
    ])
);

Tips and Gotchas

  • If the compiler is complaining, running slowly, or using too much memory, consider using Relation::dynamic to simplify your type signatures. Relation::t is useful for keeping track of a complex relation's item type.

  • Changes are propagated lazily from inputs to outputs only when Output::get is called. Calling ExecutionContext::commit only marks pending changes as "ready" to be propagated and any downstream outputs as dirty.

  • If for some reason you have created multiple CreationContexts, any function call involving Inputs, Outputs, or Relations from different contexts should result in a runtime panic with the message "Context mismatch".

Feedback Operators

In addition to creating acyclic relational graphs, it is also possible to create cyclic graphs where relational outputs feed back into inputs until all collections stabilize. To create a feedback loop in this way, use one of the three CreationContext::feed, CreationContext::feed_ordered, or CreationContext::feed_while methods.

CreationContext::feed connects a Relation to an Input such that whenever ExecutionContext::commit is called, any changes to the collection represented by the Relation argument are fed back into the Input argument. This repeats until the collection stops changing.

CreationContext::feed_ordered is similar to feed except that the Relation argument additionally has an ordering key. Rather than feeding all changes back into the Input, only those with the minimum present ordering key are fed back in. If any later changes are cancelled out as a result of this (if their count goes to zero), then they will not be fed in at all. This can be handy in situations where using feed can cause an infinite loop.

CreationContext::feed_while takes an Output as an argument rather than a Relation and rather than propagating changes to it's argument through will instead send the entire contents of that Output on every visit. feed_while is intended to be used in circumstances where there exists a negative feedback loop between the arguments and the caller wants to retain any visited values rather than have them be immediately deleted.

CreationContext::interrupt takes an Output to monitor. Calls to commit upon discovering that the Output is non-empty will immediately stop running and apply the supplied continuation. If a call to commit returns a Some, that serves as an indicator that this has happened.

If there are multiple calls to feed, feed_ordered, feed_while, or interrupt, earlier calls take higher priority. Higher priority feedbacks will run to completion before any lower priority ones are touched.

For an example, see dijkstra.rs

Dependencies

~87KB