Skip to main content

Rule Engine

The rule engine service is the core of the Forge platform. It is responsible for evaluating the rules and signals sent to agents. It is what actually simulates agents on a higher level of abstraction.

You can learn more about how to create agents and write rules in Creating Agents and Writing rules.

How it works​

When running forge run, one of the services that is started is the rule engine service. Its "main" function can be found in the Runner.java file.

During initialization, a few things happen:

  • All rules are loaded and a Drools session pool initialized
  • The whole classpath is scanned for:
    • All subclasses of Signal - a mapping is created from the signal name to the class itself. This is how the service knows what class to deserialize the raw JSON message received from the topic. That's also why the signal names must be unique across the whole project.
    • Event mapping strategies (eg. GetOrCreateAgent), which can be used in signals.yaml to map events to agents.
    • Agent types (subclasses of Agent) - a mapping is created from the agent type to the class itself
    • All classes annotated with @DataModel - similarly to signals, creates a mapping from the data model name to the actual class
  • Runner.initialize function is called where you can do any additional initialization (typically registering for events)

After initialization, the start method is called, which performs the following steps:

  • The service loads all registered events and figures out which topics it should listen on
  • If it's running locally, it unlocks all agents (to help with development)
  • It starts a message consumer listening indefinitely on the appropriate topics

In addition to the event topics, two other topics are used:

  • rule_engine-input - used for API calls to the rule engine itself
  • agents-input - where messages to agents are sent

When a message on the rule_engine-input topic is received, the service will run the specified Rule Engine's API, just like any other service. You can use RuleEngineAPI to send such messages.

When a message on the agents-input topic is received, the service will deserialize it into the appropriate class and evaluate the agent specified in the to field with the signal added to the signals entry point. If the message is of type MESSAGE, REPLY or ERROR, the field messageType is used to find the signal class. If the message is of type EVENT, the field eventType is used. If it's DATA_CHANGE, the field objectType is used.

When a message on any of the event topics is received, the service will deserialize the event into the appropriate class, determine which agent(s) it should be sent to, and call evaluateWithSignal for each of them. It determines that by using the event mapping strategies specified in signals.yaml or in configureSignals (Events.on or DataChanges.on). To learn more about configuring signals, see Event subscriptions.

Note: It sends evaluateWithSignal messages for each agent in order to distribute the load across multiple rule engine instances in the event of many agents subscribing to the same event, which would block the service until it processes all of them.

Agent evaluation​

In an ideal case, all agents would constantly be running in an infinite loop, listening to signals and evaluating rules. Unfortunately, that's not possible when you have millions of agents.

Instead, the rule engine listens to signals (including a periodic heartbeat, see Heartbeat), loads the agent from the database, and evaluates the rules.

Locking​

In order to prevent concurrency issues and allow for horizontal scalability, the rule engine uses a locking mechanism to ensure that only one instance of the rule engine is evaluating a specific agent at a time. You can read more about agent locking here.

When the rule engine receives a signal for an agent, it first tries to lock the agent. If it succeeds, it evaluates the agent and unlocks it. If it doesn't succeed, the signal is atomically appended to pendingSignals (or pendingUpdates if the signal is actually an updateAgent API call).

Before each evaluation, if the lock succeeded, the service checks if there are any pending signals or updates and adds them to the current evaluation. After each evaluation, it checks again, and if there's something pending it schedules another evaluation. This makes sure that all signals are always handled immediately instead of waiting for the next heartbeat.

Important ❗ If an exception occurs during agent evaluation, the agent stays locked to prevent infinite error loops! If an agent is locked for more than 2 minutes, you will get an alert. You will then have to manually fix the error, and unlock the agent. See this for instructions.

Evaluation​

Before any evaluation can take place, we first need to load the agent's data from the database (MongoDB). We call this data the agent's summary. It contains the agent itself, all facts inserted into the memory, any pending signals or updates, and the lock information.

It looks something like this:

{
  "agentId": "9jk23qns",
  "facts": {
    "agents#MyAgent": {  // the agent itself (of type `agents/MyAgent.java`)
      "9jk23qns": {      // primary key (agent's ID) and the agent's data
        "id": "9jk23qns",
        "connections": {
          "armory": "2gVbxtQp"
        },
        "type": "MyAgent",
        "currentState": null,
        "dataUpdatedAt": "2023-03-27T12:12:02.354928Z",
        "connected": false
      }
    },
    "models#MyFact": {  // a fact that was inserted (of type `models/MyFact.java`)
      "jf8S7Jko": {
        "id": "jf8S7Jko",
        "status": "INITIALIZED",
        "createdAt": "2023-03-01T07:59:08.598193Z",
        "finishedAt": null
      }
    }
  },
  "pendingSignals": {  // optional
    "signals": {       // entry-point
      "com#mindsmiths#armory#event#UserConnectedEvent": {},  // similar to facts
      ...
    }
  },
  "pendingUpdates": {},  // optional
  "lockedAt": "2023-03-27T12:12:02.354928Z"  // optional
}

It's normal for there to be many empty objects in the pendingSignals field.

Once the summary is loaded the following happens:

  1. The facts and signals from the summary are deserialized into the specified classes
  2. A new Drools session is created
  3. Facts (along with the agent) are inserted into the memory
  4. Signals that triggered the evaluation, along with any pending signals, are inserted into the appropriate entry-point (usually signals)
  5. A special signal called Heartbeat is inserted into the signals entry-point
  6. Agent's rules are evaluated
  7. Resulting facts are serialized and the summary is updated
  8. An Evaluation event is emitted

It's important to keep in mind that the agent and all inserted facts have to be serializable to JSON.

Infinite loop protection​

It's quite easy to find yourself in an infinite loop when writing rules. That's why we've added a mechanism to stop rule execution after a certain number of rules have fired. The default value is 100, but you can tweak it by changing the MAX_RULES_FIRED variable in Runner.initialize.

Event subscriptions​

There are two ways to subscribe to events - with a yaml file or with code.

YAML​

Some plugins will automatically generate a signals.yaml file in the resources/config directory. The structure is as follows:

com.mindsmiths.armory.event.UserConnected:  # event class
  - !GetAgentByConnection                   # subscription strategy
    connectionName: armory                  # strategy-specific parameters...
    connectionField: connectionId
com.mindsmiths.armory.event.Submit:         # another event class
  - !GetAgentByConnection                   # ...
    connectionName: armory
    connectionField: connectionId
...

You then have to call configureSignals in Runner.initialize to load this YAML file:

public class Runner extends RuleEngineService {
    @Override
    public void initialize() {
        configureSignals(getClass().getResourceAsStream("config/signals.yaml"));
        ...
    }   
    ...
}

Subscription strategies​

Possible strategies are:

  • AlwaysSendTo - always sends the event to a specific agent. Parameters:
    • agentId (required) - ID of the agent to send the events to
  • AlwaysSendToMultiple - always sends the event to a specific list of agents. Parameters:
    • agentIds (required) - list of agent IDs to send the events to
  • SendToAll - sends the event to all agents of a specific type. Parameters:
    • agentType (required) - type of the agent to send the event to
  • GetOrCreateAgent - sends the event to an agent that has a specific ID, or creates a new agent if it doesn't exist. Parameters:
    • agentId (required) - ID of the agent to send the event to
    • agentType (required) - type of agent to create if no agent with agentId exists (eg. Smith)
  • GetAgentByConnection - sends the event to agent(s) with a specific connection. Parameters:
    • connectionName (required) - name of the connection
    • connectionField (required) - name of the field in the event that contains the connection ID. For example, if the event is {..., "eventType": "SmsMessageReceived", "from": "38591234567", "text": "Hello!"} and you have a connection phoneNumber on the agent, you would use from as the connectionField and phoneNumber as the connectionName. You can learn more about agent connections in Agent connections.
  • GetOrCreateAgentByConnection - sends the event to agent(s) with a specific connection, or creates a new agent if it doesn't exist. Parameters:
    • connectionName (required) - name of the connection
    • connectionField (required) - name of the field in the event that contains the connection ID
    • agentType (required) - type of agent to create if no agent with the specified connection exists (with the connection automatically added)

All strategies accept an optional entryPoint parameter if you want to send the event to a different entry-point than signals. You can create your custom strategy by implementing the EventMapping interface or extending any of the existing strategies. It will be registered automatically, no need for additional setup.

Code​

For more flexibility, you can also subscribe to events in code (usually in Runner.initialize).

You can do that by calling configureSignals and passing in any number of Events.on or DataChanges.on calls:

public class Runner extends RuleEngineService {
    @Override
    public void initialize() {
        configureSignals(
            Events.on(UserCreated.class).sendTo((event) -> Agents.createAgent(new MyAgent())),
            Events.on(UserConnectedEvent.class).sendTo((event) -> Agents.getByConnection("armory", event.getConnectionId())),
            Events.on(OptionActivated.class).sendTo((event) -> event.getAgentId()),
            DataChanges.on(Order.class).sendTo((order, changeType) -> Agents.getByConnection("userId", event.getUserId())),
            DataChanges.on(Order.class).sendToAll(HITLAgent.class)
        );
    }
    ...
}

The sentTo function accepts either an agent ID, a list of agent IDs, or a function that returns an agent ID, agent instance, or a list of either. The sendToAll function accepts an agent class.

As you can see, you can subscribe multiple times to the same event, and use different strategies, all of which will be executed. If you have a complex mapping, or one that repeats multiple times, you can extract it into a function like this:

public class Runner extends RuleEngineService {
    @Override
    public void initialize() {
        configureSignals(
            Events.on(OrderCreated.class).sendTo(this::getOrderHandlingAgents)
        );
    }

    public List<Agent> getOrderHandlingAgents(OrderCreated orderCreated) {
        List<Agent> agents = new ArrayList<>();
        for (String handler : orderCreated.getHandlers().split(","))
            agents.addAll(Agents.getByConnection("orderHandler", handler));
        agents.add(Agents.get("ORDER_MANAGER"));
        return agents;
    }
    ...
}

Agent connections​

Every agent has a special field called connections. Connections are used to find specific agents, usually when processing incoming events.

For example, you can add an email connection with the value of the user's email address to an agent, and then search for an agent with the matching email address when processing an incoming email. Other examples of connections include phone, armoryId, telegramId, etc.

You can also use connections to map agents to users in your database or CRM.

API​

Calls​

Evaluate​

Evaluates the agent with the specified ID.

evaluate(agentId: str) -> None

Evaluate with signals​

Evaluates the agent with the specified ID, and sends the specified signals to it. The dictionary key is the entry-point to send the signals to.

evaluate_with_signals(agentId: str, signals: Dict[str, List[Signal]]) -> None

Create agent​

Create a new agent of a given type, with the given data. This call doesn't require a specific constructor, the params are set directly on the agent's fields.

create_agent(agentType: str, **params) -> None

Example:

RuleEngineAPI.create_agent("MyAgent", name="John", age=30)

Update agent​

Update an existing agent with the given data.

update_agent(agentId: str, **params) -> None

Example:

RuleEngineAPI.update_agent("agent-123", age=31)

Delete agent​

Delete an existing agent.

delete_agent(agentId: str) -> None

Remove pending signals​

Remove all pending signals for the specified agent.

remove_pending_signals(agentId: str) -> None

Lock agents​

Lock the specified agents. Locked agents cannot be evaluated.

lock_agents(*agentIds: str) -> None

Example:

RuleEngineAPI.lock_agents("agent-123", "agent-456")

Unlock agents​

Unlock the specified agents.

unlock_agents(*agentIds: str) -> None

Lock all agents​

lock_all_agents() -> None

Unlock all agents​

unlock_all_agents() -> None

Events​

Agent​

class Agent(EmittableDataModel):
    id: str
    type: str
    connections: Dict[str, str]

The Agent data change is emitted when an agent is created, updated or deleted.

Evaluation​

class Evaluation(Event):
    agentId: str
    evaluatedAt: datetime

The Evaluation event is emitted after an agent finished its evaluation.

Tips and tricks​

Initializing an agent​

You might be tempted to create a custom constructor for your agent, and do something like this:

Events.on(UserCreated.class).sendTo((userCreated) -> Agents.getOrCreate(userCreated.getUserId(), new MyAgent(userCreated)))  // DON'T DO THIS

However, this results in the constructor being called on every event, which is probably not what you want. Instead, use the empty constructor and process the event in a rule:

rule "User created"
    when
        userCreated: UserCreated() from entry-point "signals"
        agent: MyAgent()
    then
        modify(agent) { initializeUser(userCreated) };
        delete(userCreated);
end

Another option to initialize an agent is to utilize the Initialize signal. This is the first signal that is automatically sent to an agent when it is created.

rule "Initialize"
    when
        Initialize() from entry-point "signals"
        agent: MyAgent()
    then
        agent.initialize();
end

Debugging an agent​

Since version 5.0.1, all rule engine logs include the agent ID. This allows you to easily search or filter (eg. grep) logs to debug a specific agent.

Getting current datetime​

You might be tempted to use LocalDateTime.now() or new Date() to get the current datetime, but this is NOT recommended. Since the rules are evaluated at slightly different times, edge cases often arise when comparing times, causing infinite loops and non-triggered rules. Additionally, it breaks the ability to test your rules with a "faked" time.

The right way to do it is by using the Heartbeat signal:

rule "Heartbeat"
    when
        Heartbeat(now: timestamp) from entry-point "signals"
        agent: MyAgent(lastCheckInAt before[2h] now)
    then
        // send reminder
        modify(agent) { setReminderSentAt(now) };
end

If you need to get the current datetime in a method, you can use Utils.now() (Utils.getUtcDatetime in older versions).

Creating another agent from within an agent​

You can create another agent from within an agent by calling Agents.createAgent from within a rule or a method.

Agents.createAgent(new MyOtherAgent());

You can send the other agent a message like this:

rule "Some rule"
    when
        ...
    then
        Agent other = Agents.createAgent(new MyOtherAgent());
        agent.send(other.getId(), new MyMessage());
        
        modify(agent) { setOtherAgentId(other.getId()) };  // usefull to save if you need it later
end

Deleting the current agent​

An agent can delete itself by calling delete(agent) in a rule, or Agents.deleteAgent(agent) elsewhere. This stops the current evaluation immediately and deletes the agent and its summary.

Stopping an evaluation early​

You can call Agents.stopEvaluation() at any point during an evaluation to stop the evaluation immediately. This is useful if you have a specific condition that should effectively deactivate an agent.

Salience​

There is a special Salience class that allows you to categorize the priority of a rule. You can always use an integer for salience, but this class allows you to use a more descriptive approach.

rule "Some rule"
    salience Salience.HIGH
    when
        ...
    then
        ...
end

Temporary facts​

Sometimes it's useful to insert a fact into the working memory for use in other rules, but you don't want it to persist. You can annotate an object with @TemporaryFact to prevent it from being persisted into the database.

@Data
@TemporaryFact
public class MyFact {
    ...
}
rule "Some rule"
    when
        ...
    then
        insert(new MyFact());  // exists only for the duration of the current evaluation
end

Using cron expressions​

There are useful utilities for working with cron expressions built-in. You can evaluate a cron expression like this:

rule "Heartbeat"
    when
        agent: MyAgent()
        Heartbeat(now: timestamp,
                  DateUtil.evaluateCronExpression("* * 9-17 ? * * *", now, agent.getTimezone())) from entry-point "signals"
    then
        ...
end

There's also an "or" operation (//) to combine multiple cron expressions:

* * 9-17 ? * MON-FRI * // * * 9-15 ? * SAT *

You can find the next active time for a cron expression with DateUtil.getNextValidTimeForCron, or the next starting period with DateUtil.getNextActiveTimePeriodStartForCron. If the expression is already satisfied at the given time, getNextValidTimeForCron will return the very next second (if it still satisfies the cron), but getNextActiveTimePeriodStartForCron will return the start of the next period.

Working with timezones​

Generally, a good practice for working with time zones is to always store and use UTC internally, and convert to the user's timezone only when displaying the time.

In Java, we usually use the LocalDateTime object, which does not store the timezone. It is implicitly assumed to be in UTC by the platform. To add timezone information, we can create a ZonedDateTime object.

You can use Utils.datetimeInTimezone() (forge-sdk>=5.0.8) or the following snipped to convert a UTC LocalDateTime to a ZonedDateTime in the user's timezone:

Utils.datetimeInTimezone(localDateTime, timezone)
// or
localDateTime.atZone(ZoneId.of("UTC")).withZoneSameInstant(ZoneId.of(timezone))

You can convert the resulting ZonedDateTime to a LocalDateTime with zonedDateTime.toLocalDateTime(), which discards the timezone information, but retains the date and time.

Changelog​

[5.0.10] - 2023-04-25​

Added​

  • Agents.deleteAgent(agentId)
  • Agents.getCreateOrDelete
  • Agents.getCreateOrDeleteByConnection
  • utilities for agent tests (assertRuleFired, assertRuleNotFired, getFactsOfType, ...)
  • Agent.removeConnection

Changed​

  • using forge-sdk==5.0.10

[5.0.9] - 2023-03-30​

Fixed​

  • connection errors for Redis and Mongo when unit testing agents

[5.0.8] - 2023-03-28​

Added​

  • unit testing agents with AgentTest

Changed​

  • using forge-sdk==5.0.8

[5.0.7] - 2023-03-16​

Changed​

  • using forge-sdk==5.0.7 (which fixes PostHog)

[5.0.6] - 2023-02-24​

Changed​

  • using forge-sdk==5.0.6 (which fixes fake time)

[5.0.5] - 2023-02-23​

Added​

  • RuleEngineAPI.removePendingSignals to remove pending signals from an agent

[5.0.4] - 2023-02-14​

Fixed​

  • IllegalStateException when shutting down

[5.0.3] - 2023-01-23​

Fixed​

  • UnsupportedOperationException when there were pending signals
  • IllegalStateException when shutting down
  • emitting agent data changes only when there were changes

Changed​

  • using forge-sdk==5.0.3

[5.0.2] - 2023-01-16​

Fixed​

  • reverted to using exact dependency versions since intervals caused build issues

[5.0.1] - 2023-01-05​

Added​

  • logs now include the agent ID during agent evaluation
  • when running locally, all agents are now unlocked when the rule engine starts
  • Salience enum to allow for more readable salience values

Changed​

  • most dependencies are now intervals instead of exact versions
  • agent IDs have 16 characters by default (instead of 8)

Removed​

  • CutelogLayout

[5.0.0] - 2022-12-09​

Changed​

  • compatible with SDK 5.0

[5.0.0b5] - 2022-12-08​

Changed​

  • deprecated Agent.addConnection (use Agent.setConnection)
  • upgraded dependencies (forge-sdk==5.0.0b1, io.github.classgraph==4.8.151)

Removed​

  • json-simple dependency

[5.0.0b4] - 2022-11-15​

Added​

  • stopping agent evaluation with Agents.stopEvaluation()
  • unlockAllAgents and lockAllAgents APIs

[5.0.0b3] - 2022-11-15​

Added​

  • GetAgentByConnection for signals.yaml

[5.0.0b2] - 2022-11-14​

Fixed​

  • ConncurrentModificationExeption when clearing entry points

[5.0.0b1] - 2022-11-09​

Fixed​

  • getByConnection and getOrCreateByConnection now return the correct agent instance instead of a generic Agent

[5.0.0b0] - 2022-10-19​

Changed​

  • GetOrCreateByConnection can recursively find fields
  • Agents.getOrCreateByConnection automatically sets the connection when creating agent

Added​

  • GetOrCreate strategy

[5.0.0a11] - 2022-10-17​

Changed​

  • Agents have a random id by default
  • using forge-sdk==5.0.0a8

[5.0.0a10] - 2022-10-13​

Fixed​

  • data change events now also insert the object and change type into the session

[5.0.0a9] - 2022-10-12​

Fixed​

  • subscriptions now handle null correctly

Added​

  • data change events now also insert the object and change type into the session

[5.0.0a8] - 2022-10-11​

Changed​

  • project MUST include kmodule.xml (BREAKING)

Fixed​

  • signals sometimes wouldn't be inserted into the session correctly

[5.0.0a7] - 2022-10-06​

Changed​

  • using forge-sdk==5.0.0a6

[5.0.0a6] - 2022-10-05​

Fixed​

  • "empty object is not a valid value" error

[5.0.0a5] - 2022-10-03​

Changed​

  • using forge-sdk==5.0.0a5

[5.0.0a4] - 2022-09-21​

Changed​

  • using forge-sdk==5.0.0a4

[5.0.0a3] - 2022-09-19​

Changed​

  • using forge-sdk==5.0.0a3

[5.0.0a2] - 2022-09-15​

Changed​

  • using forge-sdk==5.0.0a2

[5.0.0a1] - 2022-09-07​

Changed​

  • using the "address" concept when messaging

[4.0.0a21] - 2022-09-13​

Changed​

  • using forge-sdk==4.0.0a11

[4.0.0a20] - 2022-06-22​

Added​

  • Agents.getOrCreate

Changed​

  • Agents.getOrCreateByConnection needs an Agent instance instead of class
  • util.Util removed (functionality moved to forge-sdk)
  • using forge-sdk==4.0.0a10

[4.0.0a19] - 2022-06-14​

Fixed​

  • updating the agent collection on changes to the agent

[4.0.0a18] - 2022-06-13​

Added​

  • agentId index for the summary collection
  • drools_processing metric

Changed​

  • (!) rule packages no longer ignore the "Agent" suffix - they need to be the same as the agent's class (ignoring case)
  • entryPoint in subscriptions is now required
  • deleteAgent doesn't throw an error if the agent doesn't exist
  • using forge-sdk==4.0.0a9

Removed​

  • facts annotated with DataModel are no longer saved/updated in their own collections

[4.0.0a17] - 2022-06-07​

Fixed​

  • Agents.get returns the agent in the correct class

Added​

  • APIs for creating, updating and deleting agents
  • API for sending signals to agents
  • Signals and agents are automatically registered and converted

Changed​

  • using forge-sdk==4.0.0a8

Removed​

  • className field on Agent

[4.0.0a16] - 2022-05-20​

Added​

  • Infinite loop guard (limits number of rule firings)

Changed​

  • New fact serialization method (fixes problems with JsonNode)
  • Updated forge-sdk to 4.0.0a7
  • Updated drools to 7.68.0.Final

[4.0.0a0] - 2022-02-04​

Added​

  • UserEvaluation API