/rues

Rule evaluation sidecar

Primary LanguageRustMIT LicenseMIT

RuES - Expression Evaluation as Service

RuES is a minimal JMES expression evaluation side-car, that uses JMESPath, and it can handle arbitrary JSON. Which effectively makes it general purpose logical expression evaluation engine, just like some Python libraries that used to evaluate logical expression. This in turn can allow you implement complex stuff like Rule engine, RBAC, or Policy engines etc.

Here is what makes RuES special:

  • Lean and Zippy - Checkout initial benchmarks below. Under 20 MB with single CPU one will easily do 10K RPS.
  • Zero restarts - Add/remove rules on fly by making changes in rules.hjson without restarting.
  • HTTP & JSON - Ubiquitous! No custom protocols, no shenanigans.
  • UNIX philosophy - Only evaluates rules, no fancy hooks or integrations. Dead simple!

Why?

A very obvious question to ask might be, why RuES and why not just use a library? RuES can be beneficial in large scale scenarios with following benefits:

  • Unified and consistent rules - No need to deal with library differences, specially in a polyglot stack you won't have to worry about any inconsistencies, performance issues, or library maintenances.
  • Isolated and scalable - While embedded libraries can have a broader attack surface the isolated process gives you sandbox, and due to being lightweight have it as a sidecar giving you sub-millisecond latencies. This not only allows developers to hand off security to the right team, but also allows you to scale with your system.
  • Centrally managed - Allowing you to have centrally managed deployments, and rules. Changing rules doesn't even require a new deployment. The rules are live reloaded. That means with 0 downtime you can add/modify rules on the fly.

Usage

Make sure you have rules.hjson in your current working directory when launching rues. Given following example rules:

{
  example_one: "value == `2`"
  example_two: "a.b"
}

Each rule is exposed under /eval/{rule_name} as POST endpoint, which in turn can be posted payload to evaluate the expression. Simple use curl to test:

> curl -X POST http://localhost:8080/eval/example_one -H 'Content-Type: application/json' -d '{"value": 2}'
{"Success":{"expression":"value == `2`","name":"example_one","is_truthy":true,"value":true}}
> curl -X POST http://localhost:8080/eval/example_two -H 'Content-Type: application/json' -d '{"a": {"b": "Hello"}}'
{"Success":{"expression":"a.b","name":"example_two","is_truthy":true,"value":"Hello"}}

Response object contains Success if evaluation was successful. e.g.

{
   "Success": {
      "name": "filter_active",
      "expression": "[?isActive] | length(@)",
      "is_truthy": true,
      "value": 2
   }
}

Response will have an Error if there was an error in expression or there was some violation while evaluating the expression (in which case reason will contain a reason):

{
   "Error": {
      "name": "filter_registered",
      "expression": "[?matched('^201\\d', registered)] | length(@)",
      "reason": "Runtime error: Call to undefined function matched (line 0, column 9)\n[?matched('^201\\d', registered)] | length(@)\n         ^\n"
   }
}

Response will have a NotFound if the specified rule is not found:

{
   "NotFound": {
      "name": "filter_register"
   }
}

Batch Rules API

Many times you need evaluate a set of rules against a payload. RuES supports evaluating a context against multiple rules using batch API. Given the rules file:

{
  example_one: "c == `2`"
  example_two: "a.b"
}

One can invoke batch api by simply invoking /eval with POST data of:

{
   "context": {
      "c": 3,
      "a": {
         "b": true
      } 
   },
   "rules": ["example_one", "example_two", "example_three"]
}

The rules will be evaluated in sequence of order they were passed in, and server will return an array response:

[
   {"Success":{"expression":"c == `2`","name":"example_one","is_truthy":false,"value":false}},
   {"Success":{"expression":"a.b","name":"example_two","is_truthy":true,"value":true}},
   {"NotFound":{"name":"example_three"}}
]

Additional functions

In addition to built-in functions of JMES, there additional are following additional functions:

  • ✅ string[] match(expref string $regex, string $element) - Returns an array of all groups of regex matching or a null if there is no match. Regex specs can be found here. Regexes are compiled and cached in LRU order. The given Regex has to be an expression with string literal e.g. &'\d+' this is required so that regexes are always string literal and never variables eliminating any possibility of regex injection via variables, preventing any exploits or accidental explosion of regex patterns. Examples: 
    [?match(&'^[a-z0-9_-]{3,16}$', username)]
[?match(&'^[a-z0-9_-]{3,16}$', 'user_123')]
[?match(&'([12]\d{3}-(0[1-9]|1[0-2])-(0[1-9]|[12]\d|3[01]))', date)]
  • ✅ bool valid_email(string $element) - Returns true or false based on email format. In addition to formatting it also excludes temporary email addresses. Examples: 
    [?valid_email('user123@gmail.com')]
[?!valid_email('janette@guerrillamailblock.com')]
[?valid_email(contact.email)]
  • ✅ number now() - Returns current time unix timestamp as 64-bit float with number of seconds since 1970-01-01 the fractional part of timestamp contains upto microseconds of the timestamp. Examples: 
    now()
> 1641257813.803243
  • ✅ number duration(string $element) - Parses a duration string with same specs as systemd.time and returns a unix timestamp just like now. This will allow anyone to perform addition/subtraction operations on timestamps, Examples: 
    duration('1min100ms')
> 60.1
duration('1h')
> 360.0
  • 🚧 number parse_datetime(string $element[, string $format = 'rfc3339']) (To be implemented yet) - Converts datetime in given format to a timestamp. The timestamp then in turn can be used to do comparisons or reformatting.
  • 🚧 string to_datetime(number $element[, string $format = 'rfc3339']) (To be implemented yet) - Converts timestamp to a given string format.
  • 🚧 bool in_geo_fence(number[] $center, number $radius, number[] $element) (To be implemented yet) - Returns true or false if the $element lies within the $radius of $center.
  • 🚧 number[][] filter_in_geo_fence(number[] $center, number $radius, number[][] $elements) (To be implemented yet) - Returns all elements that lie within geo fence of given radius and center.
  • 🚧 bool match_glob(string $pattern, string $element) (To be implemented yet) - Returns true or false if the $element is a glob match of the $pattern.

Configuration variables

  • CONFIG_PATH - path to rules file, file can be .json, .yaml, or .hjson. Default: rules.hjson
  • BIND_ADDRESS - service address to bind to. Default: 0.0.0.0:8080

Benchmarks

My brief stress testing shows with a single CPU core (single worker), 3 rules, and payload size of 1.6 KB. Server was easily able to handle 10K RPS (even with sustained load) under 19 MB of RSS memory footprint, and a p99 of 4ms.

$ cat vegeta_attack.txt | vegeta attack -duration=10s -rate=10000 | vegeta report 
Requests      [total, rate, throughput]         100000, 10000.20, 9999.80
Duration      [total, attack, wait]             10s, 10s, 394.927µs
Latencies     [min, mean, 50, 90, 95, 99, max]  107.266µs, 811.954µs, 285.329µs, 2.128ms, 2.654ms, 4.517ms, 12.373ms
Bytes In      [total, mean]                     9566673, 95.67
Bytes Out     [total, mean]                     166000000, 1660.00
Success       [ratio]                           100.00%
Status Codes  [code:count]                      200:100000  
Error Set:

With two CPU cores (two workers), the results were even better:

$ cat vegeta_attack.txt | vegeta attack -duration=10s -rate=10000 | vegeta report
Requests      [total, rate, throughput]         100000, 10000.30, 10000.08
Duration      [total, attack, wait]             10s, 10s, 217.653µs
Latencies     [min, mean, 50, 90, 95, 99, max]  111.479µs, 270.125µs, 219.274µs, 413.215µs, 564.181µs, 1.021ms, 8.184ms
Bytes In      [total, mean]                     9566673, 95.67
Bytes Out     [total, mean]                     166000000, 1660.00
Success       [ratio]                           100.00%
Status Codes  [code:count]                      200:100000  
Error Set:

All the rules, and data has been shipped under stress_test. Feel free to share your results, and I will be more than happy to include your results.