Event Ruler (called Ruler in rest of the doc for brevity) is a Java library that allows matching Rules to Events. An event is a list of fields, which may be given as name/value pairs or as a JSON object. A rule associates event field names with lists of possible values. There are two reasons to use Ruler:
- It's fast; the time it takes to match Events doesn't depend on the number of Rules.
- Customers like the JSON "query language" for expressing rules.
Contents:
- Ruler by Example
- And and Or With Ruler
- How to Use Ruler
- JSON Text Matching
- JSON Array Matching
- Compiling and Checking Rules
- Performance
It's easiest to explain by example.
An Event is a JSON object. Here's an example:
{
"version": "0",
"id": "ddddd4-aaaa-7777-4444-345dd43cc333",
"detail-type": "EC2 Instance State-change Notification",
"source": "aws.ec2",
"account": "012345679012",
"time": "2017-10-02T16:24:49Z",
"region": "us-east-1",
"resources": [
"arn:aws:ec2:us-east-1:123456789012:instance/i-000000aaaaaa00000"
],
"detail": {
"c-count": 5,
"d-count": 3,
"x-limit": 301.8,
"source-ip": "10.0.0.33",
"instance-id": "i-000000aaaaaa00000",
"state": "running"
}
}
You can also see this as a set of name/value pairs. For brevity, we present only a sampling. Ruler has APIs for providing events both in JSON form and as name/value pairs:
+--------------+------------------------------------------+
| name | value |
|--------------|------------------------------------------|
| source | "aws.ec2" |
| detail-type | "EC2 Instance State-change Notification" |
| detail.state | "running" |
+--------------+------------------------------------------+
Events in the JSON form may be provided in the form of a raw JSON String, or a parsed Jackson JsonNode.
The rules in this section all match the sample event above:
{
"detail-type": [ "EC2 Instance State-change Notification" ],
"resources": [ "arn:aws:ec2:us-east-1:123456789012:instance/i-000000aaaaaa00000" ],
"detail": {
"state": [ "initializing", "running" ]
}
}
This will match any event with the provided values for the resource
,
detail-type
, and detail.state
values, ignoring any other fields in the
event. It would also match if the value of detail.state
had been
"initializing"
.
Values in rules are always provided as arrays, and match if the value in the
event is one of the values provided in the array. The reference to resources
shows that if the value in the event is also an array, the rule matches if the
intersection between the event array and rule-array is non-empty.
{
"time": [ { "prefix": "2017-10-02" } ]
}
Prefix matches only work on string-valued fields.
{
"source": [ { "prefix": { "equals-ignore-case": "EC2" } } ]
}
Prefix equals-ignore-case matches only work on string-valued fields.
{
"source": [ { "suffix": "ec2" } ]
}
Suffix matches only work on string-valued fields.
{
"source": [ { "suffix": { "equals-ignore-case": "EC2" } } ]
}
Suffix equals-ignore-case matches only work on string-valued fields.
{
"source": [ { "equals-ignore-case": "EC2" } ]
}
Equals-ignore-case matches only work on string-valued fields.
{
"source": [ { "wildcard": "Simple*Service" } ]
}
Wildcard matches only work on string-valued fields. A single value can contain zero to many wildcard characters, but consecutive wildcard characters are not allowed. To match the asterisk character specifically, a wildcard character can be escaped with a backslash. Two consecutive backslashes (i.e. a backslash escaped with a backslash) represents the actual backslash character. A backslash escaping any character other than asterisk or backslash is not allowed.
Anything-but matching does what the name says: matches anything except what's provided in the rule.
Anything-but works with single string and numeric values or lists, which have to contain entirely strings or entirely numerics. It also may be applied to a prefix match.
Single anything-but (string, then numeric):
{
"detail": {
"state": [ { "anything-but": "initializing" } ]
}
}
{
"detail": {
"x-limit": [ { "anything-but": 123 } ]
}
}
Anything-but list (strings):
{
"detail": {
"state": [ { "anything-but": [ "stopped", "overloaded" ] } ]
}
}
Anything-but list (numbers):
{
"detail": {
"x-limit": [ { "anything-but": [ 100, 200, 300 ] } ]
}
}
Anything-but prefix:
{
"detail": {
"state": [ { "anything-but": { "prefix": "init" } } ]
}
}
Anything-but suffix:
{
"detail": {
"instance-id": [ { "anything-but": { "suffix": "1234" } } ]
}
}
Anything-but-ignore-case list (strings):
{
"detail": {
"state": [ { "anything-but": {"equals-ignore-case": [ "Stopped", "OverLoaded" ] } } ]
}
}
{
"detail": {
"c-count": [ { "numeric": [ ">", 0, "<=", 5 ] } ],
"d-count": [ { "numeric": [ "<", 10 ] } ],
"x-limit": [ { "numeric": [ "=", 3.018e2 ] } ]
}
}
Above, the references to c-count
, d-count
, and x-limit
illustrate numeric matching,
and only
work with values that are JSON numbers. Numeric matching is limited to value between
-5.0e9 and +5.0e9 inclusive, with 15 digits of precision, that is to say 6 digits
to the right of the decimal point.
{
"detail": {
"source-ip": [ { "cidr": "10.0.0.0/24" } ]
}
}
This also works with IPv6 addresses.
Exists matching works on the presence or absence of a field in the JSON event.
The rule below will match any event which has a detail.c-count field present.
{
"detail": {
"c-count": [ { "exists": true } ]
}
}
The rule below will match any event which has no detail.c-count field.
{
"detail": {
"c-count": [ { "exists": false } ]
}
}
Note Exists
match only works on the leaf nodes. It does not work on intermediate nodes.
As an example, the above example for exists : false
would match the event below:
{
"detail-type": [ "EC2 Instance State-change Notification" ],
"resources": [ "arn:aws:ec2:us-east-1:123456789012:instance/i-000000aaaaaa00000" ],
"detail": {
"state": [ "initializing", "running" ]
}
}
but would also match the event below because c-count
is not a leaf node:
{
"detail-type": [ "EC2 Instance State-change Notification" ],
"resources": [ "arn:aws:ec2:us-east-1:123456789012:instance/i-000000aaaaaa00000" ],
"detail": {
"state": [ "initializing", "running" ]
"c-count" : {
"c1" : 100
}
}
}
{
"time": [ { "prefix": "2017-10-02" } ],
"detail": {
"state": [ { "anything-but": "initializing" } ],
"c-count": [ { "numeric": [ ">", 0, "<=", 5 ] } ],
"d-count": [ { "numeric": [ "<", 10 ] } ],
"x-limit": [ { "anything-but": [ 100, 200, 300 ] } ],
"source-ip": [ { "cidr": "10.0.0.0/8" } ]
}
}
As the examples above show, Ruler considers a rule to match if all of the fields named in the rule match, and it considers a field to match if any of the provided field values match, that is to say Ruler has applied "And" logic to all fields by default without "And" primitive is required.
There are two ways to reach the "Or" effects:
- Add multiple rules with the same rule name and each individual rule will be treated as one of "Or" condition by Ruler. Refer to below under addRule() section on how to achieve an "Or" effect in that way.
- Use the "$or" primitive to express the "Or" relationship within the rule.
The "$or" primitive to allow the customer directly describe the "Or" relationship among fields in the rule.
Ruler recognizes "Or" relationship only when the rule has met all below conditions:
- There is "$or" on field attribute in the rule followed with an array – e.g. "$or": []
- There are 2+ objects in the "$or" array at least : "$or": [{}, {}]
- There is no filed name using Ruler keywords in Object of "$or" array, refer to RESERVED_FIELD_NAMES_IN_OR_RELATIONSHIP in
/src/main/software/amazon/event/ruler/Constants.java#L38
for example, below rule will be not parsed as "Or" relationship because "numeric" and "prefix" are Ruler reserved keywords.{ "$or": [ {"numeric" : 123}, {"prefix": "abc"} ] }
Otherwise, Ruler just treats the "$or" as normal filed name the same as other string in the rule.
Normal "Or":
// Effect of "source" && ("metricName" || "namespace")
{
"source": [ "aws.cloudwatch" ],
"$or": [
{ "metricName": [ "CPUUtilization", "ReadLatency" ] },
{ "namespace": [ "AWS/EC2", "AWS/ES" ] }
]
}
Parallel "Or":
// Effect of ("metricName" || "namespace") && ("detail.source" || "detail.detail-type")
{
"$or": [
{ "metricName": [ "CPUUtilization", "ReadLatency" ] },
{ "namespace": [ "AWS/EC2", "AWS/ES" ] }
],
"detail" : {
"$or": [
{ "source": [ "aws.cloudwatch" ] },
{ "detail-type": [ "CloudWatch Alarm State Change"] }
]
}
}
"Or" has an "And" inside
// Effect of ("source" && ("metricName" || ("metricType && "namespace") || "scope"))
{
"source": [ "aws.cloudwatch" ],
"$or": [
{ "metricName": [ "CPUUtilization", "ReadLatency" ] },
{
"metricType": [ "MetricType" ] ,
"namespace": [ "AWS/EC2", "AWS/ES" ]
},
{ "scope": [ "Service" ] }
]
}
Nested "Or" and "And"
// Effect of ("source" && ("metricName" || ("metricType && "namespace" && ("metricId" || "spaceId")) || "scope"))
{
"source": [ "aws.cloudwatch" ],
"$or": [
{ "metricName": [ "CPUUtilization", "ReadLatency" ] },
{
"metricType": [ "MetricType" ] ,
"namespace": [ "AWS/EC2", "AWS/ES" ],
"$or" : [
{ "metricId": [ 1234 ] },
{ "spaceId": [ 1000 ] }
]
},
{ "scope": [ "Service" ] }
]
}
"$or" is possibly already used as a normal key in some applications (though its likely rare). For these cases, Ruler tries its best to maintain the backward compatibility. Only when the 3 conditions mentioned above, will ruler change behaviour because it assumes your rule really wanted an OR and was mis-configured until today. For example, the rule below will keep working as normal rule with treating "$or" as normal field name in the rule and event:
{
"source": [ "aws.cloudwatch" ],
"$or": {
"metricType": [ "MetricType" ] ,
"namespace": [ "AWS/EC2", "AWS/ES" ]
}
}
Refer to /src/test/data/normalRulesWithOrWording.json
for more examples that "$or" is parsed as normal field name by Ruler.
The keyword "$or" as "Or" relationship primitive should not be designed as normal field in both Events and Rules. Ruler supports the legacy rules where "$or" is parsed as normal field name to keep backward compatibility and give time for team to migrate their legacy "$or" usage away from their events and rules as normal filed name. Mix usage of "$or" as "Or" primitive, and "$or" as normal field name is not supported intentionally by Ruler to avoid the super awkward ambiguities on "$or" from occurring.
There are two ways to use Ruler. You can compile multiple rules
into a "Machine", and then use either of its rulesForEvent()
method
or rulesForJSONEvent()
methods to check which of the rules match any Event.
The difference between these two methods is discussed below. This discussion
will use rulesForEvent()
generically except where the difference matters.
Alternatively, you can use a single static boolean method to determine whether an individual event matches a particular rule.
There is a single static boolean method Ruler.matchesRule(event, rule)
-
both arguments are provided as JSON strings.
NOTE: There is another deprecated method called Ruler.matches(event, rule)
which
should not be used as its results are inconsistent with rulesForJSONEvent()
and
rulesForEvent()
. See the documentation on Ruler.matches(event, rule)
for details.
The matching time does not depend on the number of rules. This is the best choice if you have multiple possible rules you want to select from, and especially if you have a way to store the compiled Machine.
The matching time is impacted by the degree of non-determinism introduced by wildcard rules. Performance deteriorates as an increasing number of the wildcard rule prefixes match a theoretical worst-case event. To avoid this, wildcard rules pertaining to the same event field should avoid common prefixes leading up to their first wildcard character. If a common prefix is required, then use the minimum number of wildcard characters and limit repeating character sequences that occur following a wildcard character. MachineComplexityEvaluator can be used to evaluate a machine and determine the degree of non-determinism, or "complexity" (i.e. how many wildcard rule prefixes match a theoretical worst-case event). Here are some data points showing a typical decrease in performance for increasing complexity scores.
- Complexity = 1, Events per Second = 140,000
- Complexity = 17, Events per Second = 12,500
- Complexity = 34, Events per Second = 3500
- Complexity = 50, Events per Second = 2500
- Complexity = 100, Events per Second = 1250
- Complexity = 275, Events per Second = 100 (extrapolated data point)
- Complexity = 650, Events per Second = 10 (extrapolated data point)
It is important to limit machine complexity to protect your application. There are at least two different strategies for limiting machine complexity. Which one makes more sense may depend on your application.
- Aggregate Complexity. Create a machine using all rules, evaluate the complexity, and reject the rule set (or the current rule being added) if it exceeds the threshold.
- Individual Complexity. Create a machine using just an individual rule (not all rules), evaluate the complexity, and reject the rule if it exceeds the threshold. In addition, limit the number of rules that can be created that contain a wildcard pattern.
Strategy #1 is more ideal in that it measures the actual complexity of the machine containing all the rules. When possible, this strategy should be used. The downside is, let's say you have a control plane that allows the creation of one rule at a time, up to a very large number. Then for each of these control plane operations, you must load all the existing rules to perform the validation. This could be very expensive. It is also prone to race conditions. Strategy #2 is a compromise. The threshold used by strategy #2 will be lower than strategy #1 since it is a per-rule threshold. Let's say you want a machine's complexity, with all rules added, to be no more than 300. Then with strategy #2, for example, you could limit each single-rule machine to complexity of 10, and allow for 30 rules containing wildcard patterns. In an absolute worst case where complexity is perfectly additive (unlikely), this would lead to a machine with complexity of 300. The downside is that it is unlikely that the complexity will be perfectly additive, and so the number of wildcard-containing rules will likely be limited unnecessarily.
For strategy #2, depending on how rules are stored, an additional attribute may need to be added to rules to indicate which ones are nondeterministic (i.e. contain wildcard patterns) in order to limit the number of wildcard-containing rules.
The following is a code snippet illustrating how to limit complexity for a given pattern, like for strategy #2.
public class Validate {
private void validate(String pattern, MachineComplexityEvaluator machineComplexityEvaluator) {
// If we cannot compile, then return exception.
List<Map<String, List<Patterns>>> compilationResult = Lists.newArrayList();
try {
compilationResult.addAll(JsonRuleCompiler.compile(pattern));
} catch (Exception e) {
InvalidPatternException internalException =
EXCEPTION_FACTORY.invalidPatternException(e.getLocalizedMessage());
throw ExceptionMapper.mapToModeledException(internalException);
}
// Validate wildcard patterns. Look for wildcard patterns out of all patterns that have been used.
Machine machine = new Machine();
int i = 0;
for (Map<String, List<Patterns>> rule : compilationResult) {
if (containsWildcard(rule)) {
// Add rule to machine for complexity evaluation.
machine.addPatternRule(Integer.toString(++i), rule);
}
}
// Machine has all rules containing wildcard match types. See if the complexity is under the limit.
int complexity = machine.evaluateComplexity(machineComplexityEvaluator);
if (complexity > MAX_MACHINE_COMPLEXITY) {
InvalidPatternException internalException = EXCEPTION_FACTORY.invalidPatternException("Rule is too complex");
throw ExceptionMapper.mapToModeledException(internalException);
}
}
private boolean containsWildcard(Map<String, List<Patterns>> rule) {
for (List<Patterns> fieldPatterns : rule.values()) {
for (Patterns fieldPattern : fieldPatterns) {
if (fieldPattern.type() == WILDCARD) {
return true;
}
}
}
return false;
}
}
The main class you'll interact with implements state-machine based rule matching. The interesting methods are:
addRule()
- adds a new rule to the machinedeleteRule()
- deletes a rule from the machinerulesForEvent()
/rulesForJSONEvent()
- finds the rules in the machine that match an event
There are two flavors: Machine
and GenericMachine<T>
. Machine is simply GenericMachine<String>
. The
API refers to the generic type as "name", which reflects history: The String version was built first and
the strings it stored and returned were thought of as rule names.
For safety, the type used to "name" rules should be immutable. If you change the content of an object while it's being used as a rule name, this may break the operation of Ruler.
All forms of this method have the same first argument, a String which provides
the name of the Rule and is returned by rulesForEvent()
. The rest of the
arguments provide the name/value pairs. They may be provided in JSON as in
the examples above (via a String, a Reader, an InputStream, or byte[]
), or as
a Map<String, List<String>>
, where the keys are the field names and the
values are the list of possible matches; using the example above, there would
be a key named detail.state
whose value would be the list containing
"initializing"
and "running"
.
Note: This method (and also deleteRule()
) is synchronized, so only one thread
may be updating the machine at any point in time.
You can call addRule()
multiple times with the same name but multiple different
name/value patterns, thus achieving an "or" relationship;
rulesForEvent()
will return that name if any of the patterns match.
For example, suppose you call addRule()
with rule name as "R1" and add
the following pattern:
{
"detail": {
"c-count": [ { "numeric": [ ">", 0, "<=", 5 ] } ]
}
}
Then you call it again with the same name but a different pattern:
{
"detail": {
"x-limit": [ { "numeric": [ "=", 3.018e2 ] } ]
}
}
After this, rulesForEvent()
will return "R1" for either a c-count
value of 2
or an x-limit
value of 301.8.
This is a mirror-image of addRule()
; in each case the first argument is the rule
name, given as a String. Subsequent arguments provide the names and values,
and may be given in any of the same ways as with addRule()
.
Note: This method (and also addRule()
) is synchronized, so only one thread may
be updating the machine at any point in time.
The operation of this API can be subtle. The Machine compiles the mapping
of name/value patterns to Rule names into a finite automaton, but does not
remember what patterns are mapped to a given Rule name. Thus, there is no
requirement that the pattern in a deleteRule()
exactly match that in the
corresponding addRule()
. Ruler will look for matches to the name/value patterns
and see if they give a match to a rule with the provided name, and if so
remove them. Bear in mind that while performing deleteRule()
calls that do not exactly
match the corresponding addRule()
calls will not fail and will not leave the
machine in an inconsistent state, they may cause "garbage" to build up in the
Machine.
A specific consequence is that if you have called addRule()
multiple times with
the same name but different patterns, as illustrated above in the Rules and rule
names section, you would have to call deleteRule()
the same number of times,
with the same associated patterns, to remove all references to that rule name
from the machine.
This method returns a List<String>
for Machine (and List<T>
for GenericMachine) which contains
the names of the rules that match the provided event. The event may be provided to either method
as a single String
representing its JSON form.
The event may also be provided to rulesForEvent()
as a collection of strings which alternate field
names and values, and must be sorted lexically by field-name. This may be a List<String>
or String[]
.
Providing the event in JSON is the recommended approach and has several advantages. First of all, populating the String list or array with alternating name/value quantities, in an order sorted by name, is tricky, and Ruler doesn't help, just fails to work correctly if the list is improperly structured. Adding to the difficulty, the representation of field values, provided as strings, must follow JSON-syntax rules - see below under JSON text matching.
Finally, the list/array version of an event makes it impossible for Ruler to recognize array
structures and provide array-consistent matching, described below in this document. The
rulesForEvent(String eventJSON)
API is deprecated in favor of rulesForJSONEvent()
specifically because it does not support array-consistent matching.
rulesForJSONEvent()
also has the advantage that the code which turns the JSON form
of the event into a sorted list has been extensively profiled and optimized.
The performance of rulesForEvent()
and rulesForJSONEvent()
do not depend on the number of rules added
with addRule()
. rulesForJSONEvent()
is generally faster because of the optimized
event processing. If you do your own event processing and call rulesForEvent()
with a pre-sorted list of name and values, that is faster still; but you may not
be able to do the field-list preparation as fast as rulesForJSONEvent()
does.
This method roughly the number of objects within the machine. It's value only varies as rule are added or
removed. This is useful to identify large machines that potentially require loads of memory.
As this method is dependent on number of internal objects, this counts may change when ruler library internals
are changed. The method performs all of its calculation at runtime to avoid taking up memory and making the
impact of large rule-machines worse. Its computation is intentionally NOT thread-safe to avoid blocking rule
evaluations and machine changes. It means that if a parallel process is adding or removing from the machine,
you may get a different results compared to when such parallel processes are complete. Also, as the library
makes optimizations to its internals for some patterns (see ShortcutTransition.java
for more details), you
may also get different results depending on the order in which rules were added or removed.
If you think of your events as name/value pairs rather than nested JSON-style
documents, the Patterns
class (and its Range
subclass) may be useful in constructing rules. The following
static methods are useful.
public static ValuePatterns exactMatch(final String value);
public static ValuePatterns prefixMatch(final String prefix);
public static ValuePatterns prefixEqualsIgnoreCaseMatch(final String prefix);
public static ValuePatterns suffixMatch(final String suffix);
public static ValuePatterns suffixEqualsIgnoreCaseMatch(final String suffix);
public static ValuePatterns equalsIgnoreCaseMatch(final String value);
public static ValuePatterns wildcardMatch(final String value);
public static AnythingBut anythingButMatch(final String anythingBut);
public static AnythingBut anythingButPrefix(final String prefix);
public static ValuePatterns numericEquals(final double val);
public static Range lessThan(final double val);
public static Range lessThanOrEqualTo(final double val);
public static Range greaterThan(final double val);
public static Range greaterThanOrEqualTo(final double val);
public static Range between(final double bottom, final boolean openBottom, final double top, final boolean openTop);
Once you have constructed appropriate Patterns
matchers with these methods, you can use the
following methods to add to or delete from your machine:
public void addPatternRule(final String name, final Map<String, List<Patterns>> namevals);
public void deletePatternRule(final String name, final Map<String, List<Patterns>> namevals);
NOTE: The cautions listed in deleteRule()
apply
to deletePatternRule()
as well.
The field values in rules must be provided in their JSON representations.
That is to say, string values must be enclosed in "quotes". Unquoted values
are allowed, such as numbers (-3.0e5
) and certain JSON-specific literals (true
,
false
, and null
).
This can be entirely ignored if rules are provided to addRule()
() in JSON form,
or if you are working with Patterns as opposed to literal strings.
But if you are providing rules as name/value pairs, and you want to specify
that the field "xyz" matches the string "true", that has to be expressed as
"xyz", "\"true\""
. On the other hand, "xyz", "true"
would match only the
JSON literal true
.
Ruler supports rule-matching for events containing arrays, but only when the event
is provided in JSON form - when it's a list of pre-sorted fields, the array structure
in the event is lost. The behavior also depends on whether you use rulesForEvent()
or rulesForJSONEvent
.
Consider the following Event.
{
"employees":[
{ "firstName":"John", "lastName":"Doe" },
{ "firstName":"Anna", "lastName":"Smith" },
{ "firstName":"Peter", "lastName":"Jones" }
]
}
Then this rule will match:
{ "employees": { "firstName": [ "Anna" ] } }
That is to say, the array structure is "crushed out" of the rule pattern, and any contained objects are treated as if they are the value of the parent field. This works for multi-level arrays too:
{
"employees":[
[
{ "firstName":"John", "lastName":"Doe" },
{ "firstName":"Anna", "lastName":"Smith" }
],
[
{ "firstName":"Peter", "lastName":"Jones" }
]
]
}
In earlier versions of Ruler, the only Machine-based matching method
was rulesForEvent()
which unfortunately will also match the following rule:
{ "employees": { "firstName": [ "Anna" ], "lastName": [ "Jones" ] } }
As a fix, Ruler introduced rulesForJSONEvent()
which, as the name suggests, only
matches events provided in JSON form. rulesForJsonEvent()
will not match the
"Anna"/"Jones" rule above.
Formally: rulesForJSONEvent()
will refuse to recognize any match in which
any two fields are within JSON objects that are in different elements of the same array.
In practice, this means that it does about what you would expect.
There is a supporting class com.amazon.fsm.ruler.RuleCompiler
. It contains a
method named check()
which accepts a JSON rule definition and returns a
String value which, if null, means that the rule was syntactically valid. If
the return value is non-Null it contains a human-readable error message
describing the problem.
For convenience, it also contains a method named compile()
which works just
like check()
but signals an error by throwing an IOException and, on
success, returns a Map<String>, List<String>>
in the form that Machine's
addRule()
method expects. Since the Machine class uses this internally,
this method may be a time-saver.
When Ruler compiles keys, it uses dot (.
) as the joining character. This means
it will compile the following two rules to the same internal representation
## has no dots in keys
{ "detail" : { "state": { "status": [ "running" ] } } }
## has dots in keys
{ "detail" : { "state.status": [ "running" ] } }
It also means that these rules will match against following two events :
## has no dots in keys
{ "detail" : { "state": { "status": "running" } } }
## has dots in keys
{ "detail" : { "state.status": "running" } }
This behaviour may change in future version (to avoid any confusions) and should not be relied upon.
We measure Ruler's performance by compiling multiple rules into a Machine and matching events provided as JSON strings.
A benchmark which processes 213,068 JSON events with average size about 900 bytes against 5 each exact-match, prefix-match, suffix-match, equals-ignore-case-match, wildcard-match, numeric-match, and anything-but-match rules and counts the matches, yields the following on a 2019 MacBook:
Events are processed at over 220K/second except for:
- equals-ignore-case matches, which are processed at over 200K/second.
- prefix/equals-ignore-case matches, which are processed at over 200K/second.
- suffix/equals-ignore-case matches, which are processed at over 200K/second.
- wildcard matches, which are processed at over 170K/second.
- anything-but matches, which are processed at over 150K/second.
- numeric matches, which are processed at over 120K/second.
- complex array matches, which are processed at over 2.5K/second.
Here are some suggestions on processing rules and events:
- If your team is still using old API -- rulesForEvent, switch to rulesForJSONEvent API. Due to limited resource, old API will not be maintained well thought contributions are always welcomed.
- If your team does event flattening by yourself, you are recommended to use Ruler to flatten the event, just pass Json string or Json node. We have many optimizations within Ruler parsing code.
- if your team does Rule Json parsing by yourself, you are recommended to just pass the Json described rule string directly to Ruler, in which will do some pre-processing, e.g. add “”.
- In order to well protect the system and prevent ruler from hitting worse condition, limit number of fields in event and rule, e.g. for big event, consider to split to multiple small event and call ruler multiple times. while number of rule is purely depending on your memory budget which is up to you to decide that, but number of fields described in the rule is most important and sensitive on performance, if possible, try to design it as small as possible.
From performance consideration, Ruler is sensitive on below items, so, when you design the schema of your event and rule, here are some suggestions:
- Try to make Key be diverse both in event and rules, the more heterogeneous fields in event and rule, the higher performance.
- Shorten number of fields inside rules, the less key in the rules, the short path to find them out.
- Shorten number of fields inside event, the less key inside event, the less attempts will be required to find out rules.
- Shorten number of possible value in […](e.g. “a”:[1,2,3 …100] ) both inside event and rules, the more value, the more branches produced in FSM to iterator, then the more time takes for matching.
See CONTRIBUTING for more information.
This project is licensed under the Apache-2.0 License. See LICENSE for more information.