durable

A lightweight library for real-time, consistent and scalable coordination of events.

durable_rules

durable_rules is a polyglot micro-framework for real-time, consistent and scalable coordination of events. With durable_rules you can track and analyze information about things that happen (events) by combining data from multiple sources to infer more complicated circumstances.

A full forward chaining implementation (A.K.A. Rete) is used to evaluate facts and massive streams of events in real time. A simple, yet powerful meta-liguistic abstraction lets you define simple and complex rulesets as well as control flow structures such as flowcharts, statecharts, nested statecharts and time driven flows.

The durable_rules core engine is implemented in C, which enables ultra fast rule evaluation as well as muti-language support. durable_rules relies on state of the art technologies: Node.js, Werkzeug, Sinatra are used to host rulesets written in JavaScript, Python and Ruby respectively. Inference state is cached using Redis. This allows for fault tolerant execution and scale-out without giving up performance.

durable_rules is simple. Using your scripting language of choice, you only need to describe the event or fact pattern to match (antecedent) and the action to take (consequent).

In this example the rule can be triggered by posting {"id": 1, "subject": "World"} to url http://localhost:5000/test/1.

require "durable"
Durable.ruleset :test do
  # antecedent 
  when_all (m.subject == "World") do
    # consequent 
    puts "Hello #{m.subject}"
  end
end
Durable.run_all
from durable.lang import *
 
with ruleset('test'):
    # antecedent 
    @when_all(m.subject == 'World')
    def say_hello(c):
        # consequent 
        print ('Hello {0}'.format(c.m.subject))
 
run_all()
var d = require('durable');
 
with (d.ruleset('test')) {
    // antecedent 
    whenAll(m.subject.eq('World'), function (c) {
        // consequent 
        console.log('Hello ' + c.m.subject);
    });
} 
d.runAll();

Let’s consider a couple of fictitious fraud rules used in bank account management.
Note: I'm paraphrasing the example presented in this article.

  1. If there are two debit requests greater than 200% the average monthly withdrawal amount in a span of 2 minutes, flag the account as medium risk.
  2. If there are three consecutive increasing debit requests, withdrawing more than 70% the average monthly balance in a span of three minutes, flag the account as high risk.
require "durable"
 
Durable.ruleset :fraud_detection do
  # compute monthly averages 
  when_all span(86400), (m.== "debit_cleared") | (m.== "credit_cleared") do
    debit_total = 0
    credit_total = 0
    for tx in m do
      if tx.== "debit_cleared"
        debit_total += tx.amount
      else
        credit_total += tx.amount
      end
    end
 
    s.balance = s.balance - debit_total + credit_total
    s.avg_balance = (s.avg_balance * 29 + s.balance) / 30
    s.avg_withdraw = (s.avg_withdraw * 29 + debit_total) / 30
  end
 
  # medium risk rule 
  when_all c.first = (m.== "debit_request") & 
                     (m.amount > s.avg_withdraw * 2),
           c.second = (m.== "debit_request") & 
                      (m.amount > s.avg_withdraw * 2) & 
                      (m.stamp > first.stamp) &
                      (m.stamp < first.stamp + 120) do
    puts "Medium risk"
  end
 
  # high risk rule 
  when_all c.first = m.== "debit_request",
           c.second = (m.== "debit_request") &
                      (m.amount > first.amount) & 
                      (m.stamp < first.stamp + 180),
           c.third = (m.== "debit_request") & 
                     (m.amount > second.amount) & 
                     (m.stamp < first.stamp + 180),
           s.avg_balance < (first.amount + second.amount + third.amount) / 0.7 do
    puts "High risk"
  end
end
 
Durable.run_all
from durable.lang import *
 
with ruleset('fraud_detection'):
    # compute monthly averages 
    @when_all(span(86400)(m.t == 'debit_cleared') | (m.t == 'credit_cleared'))
    def handle_balance(c):
        debit_total = 0
        credit_total = 0
        for tx in c.m:
            if tx.t == 'debit_cleared':
                debit_total += tx.amount
            else:
                credit_total += tx.amount
 
        c.s.balance = c.s.balance - debit_total + credit_total
        c.s.avg_balance = (c.s.avg_balance * 29 + c.s.balance) / 30
        c.s.avg_withdraw = (c.s.avg_withdraw * 29 + debit_total) / 30
    
    # medium risk rule 
    @when_all(c.first << (m.t == 'debit_request') & 
                         (m.amount > c.s.avg_withdraw * 2),
              c.second << (m.t == 'debit_request') & 
                          (m.amount > c.s.avg_withdraw * 2) & 
                          (m.stamp > c.first.stamp) &
                          (m.stamp < c.first.stamp + 120))
    def first_rule(c):
        print('Medium Risk')
 
    # high risk rule 
    @when_all(c.first << m.t == 'debit_request',
              c.second << (m.t == 'debit_request') &
                          (m.amount > c.first.amount) & 
                          (m.stamp < c.first.stamp + 180),
              c.third << (m.t == 'debit_request') & 
                         (m.amount > c.second.amount) & 
                         (m.stamp < c.first.stamp + 180),
              s.avg_balance < (c.first.amount + c.second.amount + c.third.amount) / 0.7)
    def second_rule(c):
        print('High Risk')
 
run_all()
var d = require('durable');
 
with (d.ruleset('fraudDetection')) {
    // compute monthly averages 
    whenAll(span(86400), or(m.t.eq('debitCleared'), m.t.eq('creditCleared')), 
    function(c) {
        var debitTotal = 0;
        var creditTotal = 0;
        for (var i = 0; i < c.m.length; ++i) {
            if (c.m[i].=== 'debitCleared') {
                debitTotal += c.m[i].amount;
            } else {
                creditTotal += c.m[i].amount;
            }
        }
 
        c.s.balance = c.s.balance - debitTotal + creditTotal;
        c.s.avgBalance = (c.s.avgBalance * 29 + c.s.balance) / 30;
        c.s.avgWithdraw = (c.s.avgWithdraw * 29 + debitTotal) / 30;
    });
 
    // medium risk rule 
    whenAll(c.first = and(m.t.eq('debitRequest'), 
                          m.amount.gt(c.s.avgWithdraw.mul(2))),
            c.second = and(m.t.eq('debitRequest'),
                           m.amount.gt(c.s.avgWithdraw.mul(2)),
                           m.stamp.gt(c.first.stamp),
                           m.stamp.lt(c.first.stamp.add(120))),
    function(c) {
        console.log('Medium risk');
    });
 
    // high risk rule  
    whenAll(c.first = m.t.eq('debitRequest'),
            c.second = and(m.t.eq('debitRequest'),
                           m.amount.gt(c.first.amount),
                           m.stamp.lt(c.first.stamp.add(180))),
            c.third = and(m.t.eq('debitRequest'),
                          m.amount.gt(c.second.amount),
                          m.stamp.lt(c.first.stamp.add(180))),
            s.avgBalance.lt(add(c.first.amount, c.second.amount, c.third.amount).div(0.7)),
    function(c) {
        console.log('High risk');
    });
}
 
d.runAll();

durable_rules can also be used to solve traditional Production Bussiness Rules problems. This example is an industry benchmark. Miss Manners has decided to throw a party. She wants to seat her guests such that adjacent people are of opposite sex and share at least one hobby.

Note how the benchmark flow structure is defined using a statechart to improve code readability without sacrificing performance nor altering the combinatorics required by the benchmark. For 128 guests, 438 facts, the execution time is less than 2 seconds in JavaScript and Python slightly above 2 seconds in Ruby. More details documented in this blog post.

IMac, 4GHz i7, 32GB 1600MHz DDR3, 1.12 TB Fusion Drive

Ruby

Python

JavaScript

Waltzdb is a constraint propagation problem for image recognition: given a set of lines in a 2D space, the system needs to interpret the 3D depth of the image. The first part of the algorithm consists of identifying four types of junctions, then labeling the junctions following Huffman-Clowes notation. Pairs of adjacent junctions constraint each other’s edge labeling. So, after choosing the labeling for an initial junction, the second part of the algorithm iterates through the graph, propagating the labeling constraints by removing inconsistent labels.

In this case too, the benchmark flow structure is defined using a statechart to improve code readability. The benchmark requirements are not altered. Execution time is around 3 seconds for 4 the case of 4 regions and around 20 for the case of 50. More details documented in this blog post.

IMac, 4GHz i7, 32GB 1600MHz DDR3, 1.12 TB Fusion Drive

Ruby

Python

JavaScript

Reference Manual:

Blog: