Streaming Event Processors

The StreamingEventProcessor, or Streaming Processor for short, is a type of Event Processor. As any Event Processor, it serves as the technical aspect to handle events by invoking the event handlers written in an Axon application.

The Streaming Processor defines itself by receiving the events from a StreamableMessageSource. The StreamableMessageSource is an infrastructure component through which we can open a stream of events. The source can also specify positions on the event stream, so-called Tracking Tokens, used as start positions when opening an event stream. An example of a StreamableMessageSource is the EventStore, like for example Axon Server or an RDBMS.

Furthermore, Streaming Processors use separate threads to process the events retrieved from the StreamableMessageSource. Using separate threads decouples the StreamingEventProcessor from other operations (e.g., event publication or command handling), allowing for cleaner separation within any application. Using separate threads allows for parallelization of the event load, either within a single JVM or between several.

When starting a Streaming Processor, it will open an event stream through the configured StreamableMessageSource. The first time a stream has started, it, by default, will begin at the tail (the oldest/the very first token) of the stream. It keeps track of the event processing progress while traversing the stream. It does so by storing the Tracking Tokens, or tokens for short, accompanying the events. This solution works towards tracking the progress since the tokens specify the event's position on the stream.

Head or Tail?

The oldest (very first) token is located at the tail of the stream, and the latest (newest) token is positioned at the head of the stream.

Maintaining the progress through tokens makes a Streaming Processor

  1. able to deal with stopping and starting the processor,

  2. more resilient against unintended shutdowns, and

  3. the token provides a means to replay events by adjusting the position of tokens.

All combined, the Streaming Processor allows for decoupling, parallelization, resiliency, and replay-ability. It is these features that make the Streaming Processor the logical choice for the majority of applications. Due to this, the "Tracking Event Processor," a type of Streaming Processor, is the default Event Processor.

Default Event Processor

Which EventProcessor type becomes the default processor depends on the event message source available in your application. In the majority of use cases, an Event Store is present. As the Event Store is a type of StreamableMessageSource, the default will switch to the Tracking Event Processor.

If the application only has an Event Bus configured, the framework will lack a StreamableMessageSource. It will fall back to the Subscribing Event Processor as the default in these scenarios. This implementation will use the configured EventBus as its SubscribableMessageSource.

There are two implementations of Streaming Processor available in Axon Framework:

  1. the Tracking Event Processor (TEP for short), and

  2. the Pooled Streaming Event Processor (PSEP for short).

Both implementations support the same set of operations. Operations like replaying events through a reset, parallelism and tracking the progress with tokens. They diverge on their threading approach and work separation, as discussed in more detail in this section.

Configuring

The Streaming Processors have several additional components that you can configure, next to the base options. For other streaming processor features that are configurable, we refer to their respective sections for more details. This chapter will cover how to configure a Tracking or Pooled Streaming Processor respectively.

Configuring a Tracking Processor

Firstly, to specify that new event processors should default to a TrackingEventProcessor, you can invoke the usingTrackingEventProcessors method:

public class AxonConfig { 
    // omitting other configuration methods...
    public void configureProcessorDefault(EventProcessingConfigurer processingConfigurer) { 
        processingConfigurer.usingTrackingEventProcessors();  
    }
}

For a specific Event Processor to be a Tracking instance, registerTrackingEventProcessor is used:

public class AxonConfig {
    // omitting other configuration methods...
    public void configureTrackingProcessors(EventProcessingConfigurer processingConfigurer) {
        // This configuration object allows for fine-grained control over the Tracking Processor
        TrackingEventProcessorConfiguration tepConfig =
              TrackingEventProcessorConfiguration.forSingleThreadedProcessing();
        
        // To configure a processor to be tracking ...
        processingConfigurer.registerTrackingEventProcessor("my-processor")
                            // ... to define a specific StreamableMessageSource ... 
                            .registerTrackingEventProcessor(
                                    "my-processor", conf -> /* create/return StreamableMessageSource */
                            )
                            // ... to provide additional configuration ...
                            .registerTrackingEventProcessor(
                                    "my-processor", conf -> /* create/return StreamableMessageSource */,
                                    conf -> tepConfig
                            );
    }
}

For more fine-grained control when configuring a Tracking Processor, the TrackingEventProcessorConfiguration can be used. When invoking the registerTrackingEventProcessor method, you can provide a tracking processor configuration object, or you can register the configuration instance explicitly:

public class AxonConfig {
    // omitting other configuration methods...
    public void registerTrackingProcessorConfig(EventProcessingConfigurer processingConfigurer) {
        TrackingEventProcessorConfiguration tepConfig =
                TrackingEventProcessorConfiguration.forSingleThreadedProcessing();
            
        // To register a default tracking config ...
        processingConfigurer.registerTrackingEventProcessorConfiguration(config -> tepConfig)
                            // ... to register a config for a specific processor.
                            .registerTrackingEventProcessorConfiguration("my-processor", config -> tepConfig);
    }
}

Configuring a Pooled Streaming Processor

Firstly, to specify that every new processors should default to a PooledStreamingEventProcessor, you can invoke the usingPooledStreamingEventProcessors method:

public class AxonConfig { 
    // omitting other configuration methods...
    public void configureProcessorDefault(EventProcessingConfigurer processingConfigurer) { 
        processingConfigurer.usingPooledStreamingEventProcessors();  
    }
}

For a specific Event Processor to be a Pooled Streaming instance, registerPooledStreamingProcessor is used:

public class AxonConfig {
    // omitting other configuration methods...
    public void configurePooledStreamingProcessors(EventProcessingConfigurer processingConfigurer) {
          // This configuration object allows for fine-grained control over the Pooled Streaming Processor
        EventProcessingConfigurer.PooledStreamingProcessorConfiguration psepConfig = 
                (config, builder) -> builder/* ... */;
          
        // To configure a processor to be pooled streaming ...
        processingConfigurer.registerPooledStreamingEventProcessor("my-processor")
                            // ... to define a specific StreamableMessageSource ... 
                            .registerPooledStreamingEventProcessor(
                                    "my-processor", conf -> /* create/return StreamableMessageSource */
                            )
                            // ... to provide additional configuration ...
                            .registerPooledStreamingEventProcessor(
                                    "my-processor", conf -> /* create/return StreamableMessageSource */, psepConfig
                            );
    }
}

For more fine-grained control when configuring a Pooled Streaming Processor, the PooledStreamingProcessorConfiguration can be used. When invoking the registerPooledStreamingEventProcessor method, you can provide a pooled streaming processor configuration object, or you can register the configuration instance explicitly:

public class AxonConfig {
    // omitting other configuration methods...
    public void registerPooledStreamingProcessorConfig(EventProcessingConfigurer processingConfigurer) {
        EventProcessingConfigurer.PooledStreamingProcessorConfiguration psepConfig = 
                (config, builder) -> builder/* ... */;
          
        // To register a default pooled streaming config ...
        processingConfigurer.registerPooledStreamingEventProcessorConfiguration(psepConfig)
                            // ... to register a config for a specific processor.
                            .registerPooledStreamingEventProcessorConfiguration("my-processor", psepConfig);
    }
}

Error Mode

The error mode differs between the Tracking- and Pooled Streaming Event Processor.

Whenever the error handler rethrows an exception, a TrackingEventProcessor will retry processing the event using an incremental back-off period. It will start at 1 second and double after each attempt until it reaches the maximum wait time of 60 seconds per attempt. This back-off time ensures that in a distributed environment, when another node is able to process events, it will have the opportunity to claim the token required to process the event.

The PooledStreamingEventProcessor simply aborts the failed part of the process. The Pooled Streaming Processor can deal with this since the threading mode is different from the Tracking Processor. As such, the chance is high the failed process will be picked up quickly by another thread within the same JVM. This chance increases further whenever the PSEP instance is distributed over several application instances.

Tracking Tokens

A vital attribute of the Streaming Event Processor is its capability to keep and maintain the processing progress. It does so through the TrackingToken, the "token" for short. Such a token accompanies each message a streaming processor receives through its event stream. It's this token that:

  1. specifies the position of the event on the overall stream, and

  2. is used by the Streaming Processor to open the event stream at the desired position on start-up.

Using tokens gives the Streaming Event Processor several benefits, like:

  • Being able to reopen the stream at any later point, picking up where it left off with the last event.

  • Dealing with unintended shutdowns without losing track of the last events they've handled.

  • Collaboration over the event handling load from two perspectives. First, the tokens make sure only a single thread is actively processing specific events. Secondly, it allows parallelization of the load over several threads or nodes of a Streaming Processor.

  • Replaying events by adjusting the token position of that processor.

To be able to reopen the stream at a later point, we should keep the progress somewhere. The progress is kept by updating and saving the TrackingToken after handling batches of events. Keeping the progress requires CRUD operation, for which the Streaming Processor uses the TokenStore.

For a Streaming Processor to process any events, it needs "a claim" on a TrackingToken. The processor will update this claim every time it has finished handling a batch of events. This so-called "claim extension" is, just as updating and saving of tokens, delegated to the Token Store. Hence, the Streaming Processors achieves collaboration among instances/threads through token claims.

In the absence of a claim, a processor will actively try to retrieve one. If a token claim is not extended for a configurable amount of time, other processor threads can "steal" the claim. Token stealing can, for example, happen if event processing is slow or encountered some exceptions.

Retrieving the current token inside an event handler

When processing an event it may be beneficial to retrieve the token belonging to that event. First and foremost, this can be achieved by adding a parameter of type TrackingToken to the event handler. This support is mentioned in the Supported Parameters for Event Handlers section.

Additionally, you can retrieve the token from the resources collection of the Unit of Work. Both the Tracking and Pooled Streaming Event Processor add the current TrackingToken under the key "Processor[{processor-name}]/Token".

Initial Tracking Token

The Streaming Processor uses a StreamableMessageSource to retrieve a stream of events that will open on start-up. It requires a TrackingToken to open this stream, which it will fetch from the TokenStore. However, if a Streaming Processor starts for the first time, there is no TrackingToken present to open the stream with yet.

Whenever this situation occurs, a Streaming Processor will construct an "initial token." By default, the initial token will start at the tail of the event stream. Thus, the processor will begin at the start and handle every event present in the message source. This start position is configurable, as is described here.

A Saga's Streaming Processor initial position

A Streaming Processor dedicated to a Saga will default the initial token to the head of the stream. The default initial token position ensures that the Saga does not react to events from the past, as in most cases, this would introduce unwanted side effects.

Conceptually there are a couple of scenarios when a processor builds an initial token on application startup. The obvious one is already shared, namely when a processor starts for the first time. There are, however, also other situations when a token is built that might be unexpected, like:

  • The TokenStore has (accidentally) been cleared between application runs, thus losing the stored tokens.

  • The application running the processor starts in a new environment (e.g., test or acceptance) for the first time.

  • An InMemoryTokenStore was used, and hence the processor could never persist the token to begin with.

  • The application is (accidentally) pointing to another storage solution than expected.

Whenever a Streaming Processor's event handlers show unexpected behavior in the form of missed or reprocessed events, a new initial token might have been triggered. In those cases, we recommend to validate if any of the above situations occurred.

Token Configuration

There are a couple of things we can configure when it comes to tokens. We can separate these options in "initial token" and "token claim" configuration, as described in the following sections:

Initial Token

The initial token for a StreamingEventProcessor is configurable for every processor instance. When configuring the initial token builder function, the received input parameter is the StreamableMessageSource. The message source, in turn, gives three possibilities to build a token, namely:

  1. createHeadToken() - Creates a token from the head of the event stream.

  2. createTailToken() - Creates a token from the tail of the event stream. Creating tail tokens is the default value for most Streaming Processors.

  3. createTokenAt(Instant) / createTokenSince(Duration) - Creates a token that tracks all events after a given time. If there is an event precisely at that given moment in time, it will also be taken into account.

Of course, you can completely disregard the StreamableMessageSource input parameter and create a token by yourself. Consider the following snippets if you want to configure a different initial token:

public class AxonConfig {
    // omitting other configuration methods...
    public void configureInitialTrackingToken(EventProcessingConfigurer processingConfigurer) {
        TrackingEventProcessorConfiguration tepConfig = 
                TrackingEventProcessorConfiguration.forSingleThreadedProcessing()
                                                   .andInitialTrackingToken(StreamableMessageSource::createHeadToken);
        
        processingConfigurer.registerTrackingEventProcessorConfiguration("my-processor", config -> tepConfig);
    }
}

Token Claims

As described here, a streaming processor should claim a token before it is allowed to perform any processing work. There are several scenarios where a processor may keep the claim for too long. This can occur when, for example, the event handling process is slow or encountered an exception.

In those scenarios, another processor can steal a token claim to proceed with processing. There are a couple of configurable values that influence this process:

  • tokenClaimInterval - Defines how long to wait between attempts to claim a segment. A processor uses this value to steal token claims from other processor threads. This value defaults to 5000 milliseconds.

  • eventAvailabilityTimeout - Defines the time to "wait for events" before extending the claim. Only the Tracking Event Processor uses this. The value defaults to 1000 milliseconds.

  • claimExtensionThreshold - Threshold to extend the claim in the absence of events. Only the Pooled Streaming Event Processor uses this. The value defaults 5000 milliseconds.

Consider the following snippets if you want to configure any of these values:

public class AxonConfig {
    // omitting other configuration methods...
    public void configureTokenClaimValues(EventProcessingConfigurer processingConfigurer) {
        TrackingEventProcessorConfiguration tepConfig = 
                TrackingEventProcessorConfiguration.forSingleThreadedProcessing()
                                                   .andTokenClaimInterval(1000, TimeUnit.MILLISECONDS)
                                                   .andEventAvailabilityTimeout(2000, TimeUnit.MILLISECONDS);
        
        processingConfigurer.registerTrackingEventProcessorConfiguration("my-processor", config -> tepConfig);
    }
}

Token Stealing

As described at the start, streaming processor threads can "steal" tokens from one another. A token is "stolen" when a thread loses a token claim. Situations like this internally result in an UnableToClaimTokenException, caught by both streaming event processor implementations and translated into warn- or info-level log statements.

Where the framework uses token claims to ensure that a single thread is processing a sequence of events, it supports token stealing to guarantee event processing is not blocked forever. In short, the framework uses token stealing to unblock your streaming processor threads when processing takes too long. Examples may include literal slow processing, blocking exceptional scenarios, and deadlocks.

However, token stealing may occur as a surprise for some applications, making it an unwanted side effect. As such, it is good to be aware of why tokens get stolen (as described above), but also when this happens and what the consequences are.

When is a Token stolen?

In practical terms, a token is stolen whenever the claim timeout is exceeded.

This timeout is met whenever the token's timestamp (e.g., the timestamp column of your token_entry table) exceeds the claimTimeout of the TokenStore. By default, the claimTimeout value equals 10 seconds. To adjust it, you must configure a TokenStore instance through its builder, as shown in the Token Store section. If you use Spring Boot, you can alternatively set the axon.eventhandling.tokenstore.claim-timeout for example to 30s to increase it to 30 seconds.

The token's timestamp is equally crucial in deciding when the timeout is met. The streaming processor thread holding the claim is in charge of updating the token timestamp. This timestamp is updated whenever the thread finishes a batch of events or whenever the processor extends the claim. When to extend a claim differs between the Tracking and Pooled Streaming processor. You should check out the token claim section if you want to know how to configure these values.

To further clarify, a streaming processor's thread needs to be able to update the token claim and, by extension, the timestamp to ensure it won't get stolen. Hence, a staling processor thread will, one way or another, eventually lose the claim.

Examples of when a thread may get its token stolen are:

  • Overall slow event handling

  • Too large event batch size

  • Blocking operations inside event handlers

  • Blocking exceptions inside event handlers

What are the consequences of Token stealing?

The consequence of token stealing is that an event may be handled twice (or more).

When a thread steals a token, the original thread was already processing events from the token's position. To protect against doubling event handling, Axon Framework will combine committing the event handling task with updating the token. As the token claim is required to update the token, the original thread will fail the update. Following this, a rollback occurs on the Unit of Work, resolving most issues arising from token stealing.

The ability to rollback event handling tasks sheds light on the consequences of token stealing. Most event processors project events into a projection stored within a database. Furthermore, if you store the projection in the same database as the token, the rollback will ensure the change is not persisted. Thus, the consequence of token stealing is limited to wasting processor cycles. This scenario is why we recommend storing tokens and projections in the same database.

If a rollback is out of the question for an event handling task, we strongly recommend making the task idempotent. You may have this scenario when, for example, the projection and tokens do not reside in the same database. or when the event handler dispatches an operation (e.g., through the CommandGateway). In making the invoked operation idempotent, you ensure that whenever the thread stealing a token handles an event twice (or more), the outcome will be identical.

Without idempotency, the consequences of token stealing can be manyfold:

  • Your projection (stored in a different database than your tokens!) may incorrectly project the state.

  • An event handler putting messages on a queue will put a message on the queue again.

  • A Saga Event Handler invoking a third-party service will invoke that service again.

  • An event handler sending an email will send that email again.

In short, any operation introducing a side effect that isn't handled in an idempotent fashion will occur again when a token is stolen.

Concluding, we can separate the consequence of token stealing into roughly three scenarios:

  1. We can rollback the operation. In this case, the only consequence is wasted processor cycles.

  2. The operation is idempotent. In this case, the only consequence is wasted processor cycles.

  3. When the task cannot be rolled back nor performed in an idempotent fashion, compensating actions may be the way out.

Token Store

The TokenStore provides the CRUD operations for the StreamingEventProcessor to interact with TrackingTokens. The streaming processor will use the store to construct, fetch and claim tokens.

When no token store is explicitly defined, an InMemoryTokenStore is used. The InMemoryTokenStore is not recommended in most production scenarios since it cannot maintain the progress through application shutdowns. Unintentionally using the InMemoryTokenStore counts towards one of the unexpected scenarios where the framework creates an initial token on each application start-up.

The framework provides a couple of TokenStore implementations:

  • InMemoryTokenStore - A TokenStore implementation that keeps the tokens in memory. This implementation does not suffice as a production-ready store in most applications.

  • JpaTokenStore - A TokenStore implementation using JPA to store the tokens with. Expects that a table is constructed based on the org.axonframework.eventhandling.tokenstore.jpa.TokenEntry. It is easily auto-configurable with, for example, Spring Boot.

  • JdbcTokenStore - A TokenStore implementation using JDBC to store the tokens with. Expects that the schema is constructed through the JdbcTokenStore#createSchema(TokenTableFactory) method. Several TokenTableFactory can be chosen here, like the GenericTokenTableFactory, PostgresTokenTableFactory or Oracle11TokenTableFactory implementation.

  • MongoTokenStore- A TokenStore implementation using Mongo to store the tokens with.

Where to store Tokens?

Where possible, we recommend using a token store that stores tokens in the same database as to where the event handlers update the view models. This way, changes to the view model can be stored atomically with the changed tokens. Furthermore, it guarantees exactly-once processing semantics.

Note that you can configure the token store to use for a streaming processor in the EventProcessingConfigurer:

To configure a TokenStore for all processors:

public class AxonConfig { 
    // omitting other configuration methods...
    public void registerTokenStore(EventProcessingConfigurer processingConfigurer) {
        TokenStore tokenStore = JpaTokenStore.builder()
                                             // …
                                             .build();
    
        processingConfigurer.registerTokenStore(config -> tokenStore);
    }
}

Alternatively, to configure a TokenStore for a specific processor, use:

public class AxonConfig { 
    // omitting other configuration methods...
    public void registerTokenStore(EventProcessingConfigurer processingConfigurer, String processorName) {
        TokenStore tokenStore = JdbcTokenStore.builder()
                                              // …
                                              .build();
    
        processingConfigurer.registerTokenStore(processorName, config -> tokenStore);
    }
}

Retrieving the Token Store Identifier

Implementations of TokenStore might share state in the underlying storage. To ensure correct operation, a token store has a unique identifier that uniquely identifies the storage location of the tokens in that store. This identifier can be queried with the retrieveStorageIdentifier method of your event processor.

StreamingEventProcessor eventProcessor = // …
String tokenStoreId =  eventProcessor.getTokenStoreIdentifier();

Parallel Processing

Streaming processors can use multiple threads to process an event stream. Using multiple threads allows the StreamingEventProcessor to more efficiently process batches of events. As described here, a streaming processor's thread requires a claim on a tracking token to process events.

Thus, to be able to parallelize the load, we require several tokens per processor. To that end, each token instance represents a segment of the event stream, wherein each segment is identified through a number. The stream segmentation approach ensures events aren't handled twice (or more), as that would otherwise introduce unintentional duplication. Due to this, the Streaming Processor's API references segment claims instead of token claims throughout.

You can define the number of segments used by adjusting the initialSegmentCount property. Only when a streaming processor starts for the first time can it initialize the number of segments to use. This requirement follows from the fact each token represents a single segment. Tokens, in turn, can only be initialized if they are not present yet, as is explained in more detail here.

Whenever the number of segments should be adjusted during runtime, you can use the split and merge functionality. To adjust the number of initial segments, consider the following sample:

The default number of segments for a TrackingEventProcessor is one.

public class AxonConfig {
    // omitting other configuration methods...
    public void configureSegmentCount(EventProcessingConfigurer processingConfigurer) {
        TrackingEventProcessorConfiguration tepConfig = 
                TrackingEventProcessorConfiguration.forParallelProcessing(2)
                                                   .andInitialSegmentsCount(2);
        
        processingConfigurer.registerTrackingEventProcessorConfiguration("my-processor", config -> tepConfig);
    }
}

Parallel Processing and Subscribing Event Processors

Note that Subscribing Event Processor don't manage their own threads. Therefore, it is not possible to configure how they should receive their events. Effectively, they will always work on a sequential-per-aggregate basis, as that is generally the level of concurrency in the command handling component.

The Event Handling Components a processor is in charge of may have specific expectations on the event order. The ordering is guaranteed when only a single thread is processing events. Maintaining the ordering requires additional work when the stream is segmented for parallel processing, however. When this is the case, the processor must ensure it sends the events to these handlers in that specific order.

Axon uses the SequencingPolicy for this. The SequencingPolicy is a function that returns a value for any given message. If the return value of the SequencingPolicy function is equal for two distinct event messages, it means that those messages must be processed sequentially. By default, Axon components will use the SequentialPerAggregatePolicy, making it so that events published by the same aggregate instance will be handled sequentially. Check out this section to understand how to influence the sequencing policy.

Each node running a streaming processor will attempt to start its configured amount of threads to start processing events. The number of segments that a single thread can claim differ between the Tracking- and Pooled Streaming Event Processor. A tracking processor can only claim a single segment per thread, whereas the pooled streaming processor can claim any amount of segments per thread. These approaches provide different pros and cons for each implementation, which this section explains further.

Sequential Processing

Even though events are processed asynchronously from their publisher, it is often desirable to process certain events in their publishing order. In Axon, the SequencingPolicy controls this order. The SequencingPolicy defines whether events must be handled sequentially, in parallel, or a combination of both. Policies return a sequence identifier of a given event.

If the policy returns the same identifier for two events, they must be handled sequentially by the Event Handling Component. Thus, if the SequencingPolicy returns a different value for two events, they may be processed concurrently. Note that if the policy returns a null sequence identifier, the event may be processed in parallel with any other events.

** Parallel Processing and Sagas**

A saga instance is never invoked concurrently by multiple threads. Therefore, the SequencingPolicy is irrelevant for a saga. Axon will ensure each saga instance receives the events it needs to process in the order they have been published on the event bus.

Conceptually, the SequencingPolicy decides whether an event belongs to a given segment. Furthermore, Axon guarantees that Events that are part of the same segment are processed sequentially.

The framework provides several policies you can use out of the box:

  • SequentialPerAggregatePolicy - The default policy. It will force domain events that were raised from the same aggregate to be handled sequentially. Thus, events from different aggregates may be handled concurrently. This policy is typically suitable for Event Handling Components that update details from aggregates in databases.

  • FullConcurrencyPolicy - This policy will tell Axon that this Event Processor may handle all events concurrently. This means that there is no relationship between the events that require them to be processed in a particular order.

  • SequentialPolicy - This policy tells Axon that it can process all events sequentially. Handling of an event will start when the handling of a previous event has finished.

  • PropertySequencingPolicy - When configuring this policy, the user is required to provide a property name or property extractor function. This implementation provides a flexible solution to set up a custom sequencing policy based on a standard value present in your events. Note that this policy only reacts to properties present in the event class.

  • MetaDataSequencingPolicy - When configuring this policy, the user is required to provide a metaDataKey to be used. This implementation provides a flexible solution to set up a custom sequencing policy based on a standard value present in your events' metadata.

Consider the following snippets when configuring a (custom) SequencingPolicy:

public class AxonConfig {
    // omitting other configuration methods...
    public void configureSequencingPolicy(EventProcessingConfigurer processingConfigurer) {
          PropertySequencingPolicy<SomeEvent, String> mySequencingPolicy = 
                  PropertySequencingPolicy.builder(SomeEvent.class, String.class)
                                          .propertyName("myProperty")
                                          .build();
          
          processingConfigurer.registerDefaultSequencingPolicy(config -> mySequencingPolicy)
                              // or, to define one for a specific processor:
                              .registerSequencingPolicy("my-processor", config -> mySequencingPolicy);
    }
}

If the available policies do not suffice, you can define your own. To that end, we should implement the SequencingPolicy interface. This interface defines a single method, getSequenceIdentifierFor(T), that returns the sequence identifier for a given event:

public interface SequencingPolicy<T> {
    
    Object getSequenceIdentifierFor(T event);
}

Thread Configuration

A Streaming Processor cannot process events in parallel without multiple threads configured. We can process events in parallel by running several nodes of an application. Or by configuring a StreamingEventProcessor to use several threads. The following section describes the threading differences between the Tracking- and Pooled Streaming Event Processor. These sections are followed up with samples on configuring multiple threads for the TEP and PSEP, respectively.

Thread and Segment Count

Adjusting the number of threads will not automatically parallelize a Streaming Processor. A segment claim is required to let a thread process any events. Hence, increasing the thread count should be paired with adjusting the segment count.

Tracking Processor Threading

The TrackingEventProcessor uses a ThreadFactory to start the process of claiming segments. It will use a single thread per segment it is able to claim until the processor exhausts the configured amount of threads. Each thread will open a stream with the StreamableMessageSource and start processing events at their own speed. Other segment operations, like split and merge, are processed by the thread owning the segment operated on.

Since the tracking processor can only claim a single segment per thread, segments may go unprocessed if there are more segments than threads. Hence, we recommend setting the number of threads (on every node) higher than or equal to the total number of segments.

To increase event handling throughput, we recommend changing the number of threads. How to do this is shown in the following sample:

public class AxonConfig {
    // omitting other configuration methods...
    public void configureThreadCount(EventProcessingConfigurer processingConfigurer) {
        TrackingEventProcessorConfiguration tepConfig =
                TrackingEventProcessorConfiguration.forParallelProcessing(4)
                                                   .andInitialSegmentsCount(4);

        processingConfigurer.registerTrackingEventProcessorConfiguration("my-processor", config -> tepConfig);
    }
}

Pooled Streaming Processor Threading

The PooledStreamingEventProcessor uses two threads pools instead of the single fixed set of threads used by the TrackingEventProcessor. The first thread pool is in charge of opening a stream with the event source, claiming as many segments as possible, and delegating all the work.

The work it coordinates is foremost the events to handle. Next to event coordination, it deals with segment operations like split and merge. The component coordinating all the work is called the Coordinator. This coordinator defaults to using a ScheduledExecutorService with a single thread, which suffices in most scenarios.

The second thread pool deals with all the segments the Coordinator of the pooled streaming processor could claim. The Coordinator starts a WorkPackage for each segment and provides them the events to handle. The work package will, in turn, invoke the Event Handling Components to process the events. These packages run within the second thread pool, the so-called "worker executor" pool. The worker-pool also defaults to ScheduledExecutorService with a single thread.

When you want to increase event handling throughput, we recommend changing the number of threads for the worker thread pool. How to do this is shown in the following sample:

public class AxonConfig {
    // omitting other configuration methods...
    public void configureThreadCount(EventProcessingConfigurer processingConfigurer) {
        // the "name" is the name of the processor, which can be used to define the thread factory name
        Function<String, ScheduledExecutorService> coordinatorExecutorBuilder =
                name -> Executors.newScheduledThreadPool(1, new AxonThreadFactory("Coordinator - " + name));

        Function<String, ScheduledExecutorService> workerExecutorBuilder =
                name -> Executors.newScheduledThreadPool(16, new AxonThreadFactory("Worker - " + name));

        EventProcessingConfigurer.PooledStreamingProcessorConfiguration psepConfig =
                (config, builder) -> builder.coordinatorExecutor(coordinatorExecutorBuilder)
                                            .workerExecutor(workerExecutorBuilder)
                                            .initialSegmentCount(32);

        processingConfigurer.registerPooledStreamingEventProcessorConfiguration("my-processor", psepConfig);
    }
}

Differences between Tracking and Pooled Streaming

Based on the threading approaches of the tracking processor and pooled streaming processor, there are a couple of differences to note:

  • Open Event Streams - The tracking processor will open a stream per segment it claims. The pooled streaming processor will always open a single event stream and delegate the events to the segment workers. Due to this, the tracking processor will use more I/O resources than the pooled streaming processor. However, the TEP's segments can move at their own speed as they open a dedicated event stream. The PSEP's segments will at least process as fast as the slowest segment in the set.

  • Segment Claims per Thread - The tracking processor can only claim a single segment per thread. The pooled streaming processor can claim any amount of segments, regardless of the number of threads configured. The maxClaimedSegments is configurable if required (the defaults is Short.MAX). The fact the TEP can only claim a single segment per thread highlights a problem of that implementation. Events will go unprocessed if there are more segments than threads when using the tracking processor since events belong to a single segment. Furthermore, it makes dynamic scaling tougher since you cannot adjust the number of threads at runtime. Here we see significant benefits for using the PSEP instead of the TEP since it completely drops the "one segment per thread" policy. As such, partial processing is never a problem, the PooledStreamingEventProcessor would encounter.

  • Thread Pool Configuration - The tracking processor does not allow sharing a thread pool between different instances. For the pooled streaming processor, a ScheduledExecutorService is configurable, which allows sharing the executor between different processor instances. Thus, the PSEP provides a higher level of flexibility towards optimizing the total amount of threads used within an application. The freedom in thread pool configuration is helpful when, for example, the number of different Event Processors in a single application increases.

Which Streaming Processor should I use?

In most scenarios, the PooledStreamingEventProcessor is the recommended processor implementation. We conclude this based on the segment-to-thread-count ratio, its ability to share thread pools, and the lower amount of opened event streams.

The TrackingEventProcessor will still be ideal if you anticipate the processing speed between segments to differ significantly. Also, if the application does not have too many processor instances, the need to share thread pools is loosened.

Multi-Node Processing

For streaming processors, it doesn't matter whether the threads handling the events are all running on the same node or on different nodes hosting the same (logical) processor. When two (or more) instances of a streaming processor with the same name are active on different machines, they are considered two instances of the same logical processor. Hence, it is not just a processor's own threads that compete for segments but also the processors on different application instances.

Thus, in a multi-node setup, each processor instance will try to claim segments, preventing events assigned to that segment from being processed on other nodes. In this process, the processor updates the token by adding a node identifier when it claims a segment to enforce the claim. The node identifier is configurable on the TokenStore. By default, it will use the JVM's name (usually a combination of the hostname and process ID) as the nodeId.

In a multi-node scenario, a fair distribution of the segments is often desired. Otherwise, the event processing load could be distributed unequally over the active instances. There are roughly three approaches to balancing the number of segments claimed per node:

  1. Through the Axon Server Dashboard's load balancing feature.

  2. For Axon Server and Spring Boot users, you can use the axon.axonserver.eventhandling.processors.[processor-name].load-balancing-strategy application property.

  3. Directly on a StreamingEventProcessor, with the releaseSegment(int segmentId) or releaseSegment(int segmentId, long releaseDuration, TimeUnit unit) method.

When Axon Server is in place, we recommend using either option one or two. Where option one requires access to the dashboard before load balancing is activated, option two works from within your framework application's properties file.

For those looking to configure load balancing through option 2, please consider the following application.properties file example:

# Enables automatic balancing for event processor "my-processor."
# Setting automatic balancing to true causes Axon Server to periodically check whether the segments are balanced.
# Note that automatic balancing is an Enterprise feature of Axon Server. 
axon.axonserver.eventhandling.processors.my-processor.automatic-balancing=true
# Set the load balancing strategy to, for example, "threadNumber."
# Note that this task is executed only once, on the start up of the Axon Framework application.
axon.axonserver.eventhandling.processors.my-processor.load-balancing-strategy=threadNumber

Whenever Axon Server is not used, we can achieve load balancing by having a streaming processor release its segments. Releasing segments is done by calling the releaseSegment method. When invoking releaseSegment, the StreamingEventProcessor will "let go of" the segment for some time.

class StreamingProcessorService {
    
    // The EventProcessingConfiguration allows access to all the configured EventProcessors
    private EventProcessingConfiguration processingConfiguration;

    // ...
    void releaseSegmentFor(String processorName, int segmentId) {
        // EventProcessingConfiguration#eventProcessor(String, Class) returns an optional of the event processor
        processingConfiguration.eventProcessor(processorName, StreamingEventProcessor.class)
                               .ifPresent(streamingProcessor -> streamingProcessor.releaseSegment(segmentId));
    }
}

Splitting and Merging Segments

The Streaming Event Processor provides scalability by supporting parallel processing. Through this, it is possible to tune the processor's performance by adjusting the number of threads. However, only changing the number of threads is insufficient since the parallelization is dictated through the number of segments.

When there is a high event load, ideally, we increase the number of segments. In turn, we can reduce the number of segments again if the load on the streaming processor decreases. To change the number of segments at runtime, the split and merge operations should be used. Splitting and merging allow you to control the number of segments dynamically.

There are roughly two approaches to adjust the number of segments for a streaming processor:

  1. Through the Axon Server Dashboard with the split and merge buttons

  2. Directly on a StreamingEventProcessor, with the splitSegment(int segmentId) and mergeSegment(int segmentId) methods

When Axon Server is in place, we recommend using option one since it is easiest to use. Whenever Axon Server is not used, and you want to adjust the number of segments, the split and merge methods should be accessible from within your application. For those required to take the second approach, consider the following snippet as a form of guidance:

class StreamingProcessorService {
    
    // The EventProcessingConfiguration allows access to all the configured EventProcessors
    private EventProcessingConfiguration processingConfiguration;

    // ...
    void splitSegmentFor(String processorName, int segmentId) {
        // EventProcessingConfiguration#eventProcessor(String, Class) returns an optional of the event processor
        processingConfiguration.eventProcessor(processorName, StreamingEventProcessor.class)
                               .ifPresent(streamingProcessor -> {
                                   // Use the result to check whether the operation succeeded
                                   CompletableFuture<Boolean> result =
                                           streamingProcessor.splitSegment(segmentId);
                               });
    }

    void mergeSegmentFor(String processorName, int segmentId) {
        processingConfiguration.eventProcessor(processorName, StreamingEventProcessor.class)
                               .ifPresent(streamingProcessor -> {
                                   // Use the result to check whether the operation succeeded
                                   CompletableFuture<Boolean> result =
                                           streamingProcessor.mergeSegment(segmentId);
                               });
    }
}

Note that if you are moving towards a solution using the StreamingProcessorController, there are a couple of points to consider. When invoking the split/merge operation on a StreamingEventProcessor, that processor should be in charge of the segment you want to split or merge. Thus, either the streaming processor already has a claim on the segment(s) or can claim the segment(s). Without the claims, the processor will simply fail the split or merge operation.

It is advised to check which segments a streaming processor has a claim on. For that, status of the processor is used. The status information shows which segments a processor instance owns. This guides which processor to invoke the split or merge on.

When doing a merge, the streaming processor should be in charge of both the provided segmentId and the segment the framework will merge it with. We can calculate the segment identifier the provided segmentId will be merged with through theSegment#mergeableSegmentId` method.

Segment Selection Considerations

When splitting or merging through Axon Server, it chooses the most appropriate segment to split or merge for you. When using the Axon Framework API directly, the developer should deduce the segment to split or segments to merge by themselves:

  • Split: for fair balancing, a split is ideally performed on the largest segment

  • Merge: for fair balancing, a merge is ideally performed on the smallest segment

Replaying Events

A benefit of streaming events is that we can reopen the stream at any point in time. Whenever some event handling components misbehaved, and the view models they update or actions they triggered should happen again, starting anew can be very useful. Handling events again by adjusting the position on the stream is what's called "a replay," a feature supported by the StreamingEventProcessor. The following sections describe how to initiate a replay and what replay API the framework provides.

Triggering a Reset

The reset API revolves around the resetTokens() method and provides a couple of options:

  • resetTokens() - Simple reset, adjusting the TrackingToken to the configured initial tracking token

  • resetTokens(R resetContext) - Resets the TrackingToken to the configured initial tracking token, providing the resetContext to the ResetHandlers

  • resetTokens(Function<StreamableMessageSource<TrackedEventMessage<?>>, TrackingToken> initialTrackingTokenSupplier) - Resets the TrackingToken to the results of the initialTrackingTokenSupplier

  • resetTokens(Function<StreamableMessageSource<TrackedEventMessage<?>>, TrackingToken> initialTrackingTokenSupplier, R resetContext) - Resets the TrackingToken to the results of the initialTrackingTokenSupplier, providing the resetContext to the ResetHandlers

  • resetTokens(TrackingToken startPosition) - Resets the TrackingToken to the provided startPosition

  • resetTokens(TrackingToken startPosition, R resetContext) - Resets the TrackingToken to the provided startPosition, providing the resetContext to the ResetHandlers

Partial Replays

A replay does not always have to start "from the beginning of time." Partially replaying the event stream suffices for a lot of applications.

To perform a so-called "partial replay," you should provide the token at a specific point in time. The StreamableMessageSource's createTokenAt(Instant) and createTokenSince(Duration) can be used for this.

If creating tokens based on time is not sufficient, but creating tokens based on the exact position is something that is more convenient, you could create a TrackingToken providing the position and give it to resetTokens(TrackingToken startPosition) or resetTokens(TrackingToken startPosition, R resetContext) methods. The concrete implementation of TrackingToken to provide depends on the Token Store being used.

Be mindful that when initiating a partial replay, the event handlers may handle an event in the middle of model construction. Hence, event handlers need to be "aware" that some events might not have been handled at all. Making the event handlers lenient (e.g., deal with missing data) or performing ad-hoc manual replays would help in that area.

As the method name suggests, the reset adjusts the tracking token to a new position. When starting a reset, the streaming processor is required to claim all its segments. All claims are required since the processor needs to update all tokens to their new position to start the replay.

To achieve this, the streaming event processor must be inactive when starting a reset. Hence, it is required to be shut down first before invoking the resetTokens operation. Once the reset was successful, the processor can be started up again.

Consider the following sample on how to trigger a reset within an application:

class StreamingProcessorController {
  
    private EventProcessingConfiguration processingConfiguration;
  
    // ...
    void resetTokensFor(String processorName) {
        // EventProcessingConfiguration#eventProcessor(String, Class) returns an optional of the event processor
        processingConfiguration.eventProcessor(processorName, StreamingEventProcessor.class)
                               .ifPresent(streamingProcessor -> {
                                   // shutdown this streaming processor
                                   streamingProcessor.shutDown();
                                   // reset the tokens to prepare the processor
                                   streamingProcessor.resetTokens();
                                   // start the processor to initiate the replay
                                   streamingProcessor.start();
                               });
    }
}

Resets in multi-node environments

If you are in a multi-node scenario, that means all nodes should shut down the StreamingEventProcessor. Otherwise, another node will pick up the segments released by the inactive processor instance.

Being able to shut down or start up all streaming processor instances is most easily achieved through the Axon Server Dashboard. The application's dashboard provides a "start" and "stop" button, which will start/stop the processor on every node.

When Axon Server is not used, you should construct a custom endpoint in your application. The StreamingProcessorService sample shared above would be ideal for adding a start and stop method.

Replay API

Initiating a replay through the StreamingEventProcessor opens up an API to tap into the process of replaying. It is, for example, possible to define a @ResetHandler, which provides a hook to prepare an Event Handling Component before the replay begins. A processor will invoke ResetHandler annotated methods as a result of StreamingEventProcessor#resetTokens.

During a reset through the StreamingEventProcessor#resetTokens API, you can supply a resetContext parameter. This context is supplied to @ResetHandler annotated methods and saved in the ReplayToken. This context can, for the duration of the replay, be accessed using the ReplayToken.replayContext methods or can be injected into event handlers using the @ReplayContext annotation.

The following sample Event Handling Component shows the available replay API:

@AllowReplay // 1.
@ProcessingGroup("card-summary")
public class CardSummaryProjection {
    //...
    @EventHandler
    @DisallowReplay // 2. - It is possible to prevent some handlers from being replayed
    public void on(CardIssuedEvent event) {
        // This event handler performs a "side effect",
        //  like sending an e-mail or a sms.
        // Neither, is something we want to reoccur when a 
        //  replay happens, hence we disallow this method 
        //  to be replayed
    }

    @EventHandler
    public void on(CardRedeemedEvent event, ReplayStatus replayStatus /* 3. */) {
        // We can wire a ReplayStatus here so we can see whether this
        // event is delivered to our handler as a 'REGULAR' event or
        // a 'REPLAY' event
        // Perform event handling
    }    

    @ResetHandler // 4. - This method will be called before replay starts
    public void onReset(ResetContext resetContext) {
        // Do pre-reset logic, like clearing out the projection table for a
        // clean slate. The given resetContext is [optional], allowing the 
        // user to specify in what context a reset was executed.
    }
    
    @EventHandler
    public void on(CardCancelledEvent event, @ReplayContext CardReplayContext context /* 5. */) {
        // During replays, this method will get the CardReplayContext injected that was providing during the reset call.
        // If there is no replay, no context was supplied or the context type does not match, the parameter is null. 
    }
    //...
}

The CardSummaryProjection shows a couple of interesting things to take note of when it comes to "being aware" of a replay in progress:

  1. An @AllowReplay can be used, situated either on an entire class or an @EventHandler annotated method. It defines whether the processor should invoke the given class or method when a replay is in transit.

  2. In addition to allowing a replay, @DisallowReplay can also be used. Similar to @AllowReplay, you can place it on class level and methods. It serves to define whether a processor should not invoke the class or method when a replay is in transit.

  3. To have more fine-grained control on what (not) to do during a replay, we can use the ReplayStatus parameter. The ReplayStatus is an additional parameter that we can add to @EventHandler annotated methods. It allows conditional operations in the event handlers based on whether a replay is taking place.

  4. If it is necessary to perform certain pre-replay logic, such as clearing out a projection table, we can use the @ResetHandler annotation. It allows adding a "reset context" to provide more information on why the reset is taking place. To include a resetContext the resetTokens(R resetContext) method (or other methods containing the resetContext parameter) should be invoked. The type of the resetContext is up to the user.

  5. If it is necessary to use information that was available at time of calling resetTokens(R resetContext) in your event handlers during a replay, you can use the @ReplayContext annotation to get access to this information. This information is stored in the ReplayToken and will be available until the end of the replay. The type of the resetContext is up to the user and is the same context as is used for the @ResetHandler (see 4.). The type of the resetContext has to match the parameter's, or it will be null.

Multiple Event Sources

You can configure a Streaming Event Processor to use multiple sources to process events from. When required to process events from several sources, we can configure a specific type of StreamableMessageSource: the MultiStreamableMessageSource. The MultiStreamableMessageSource is useful when a streaming processor should act on the events from:

  • several event stores,

  • from different storage types (e.g., an Event Store and a Kafka Stream)

Having multiple sources means that there might be a choice of multiple events that the processor could consume at any given instant. Therefore, you can specify a Comparator to choose between them. The default implementation chooses the event with the oldest timestamp (i.e., the event waiting for the longest).

Using multiple sources also means that the streaming processor's polling interval needs to be divided between sources. Some sources might use a strategy to optimize event discovery, thus minimizing overhead in establishing costly connections to the data sources. To that end, you can choose which source the processor does most of the polling on using the longPollingSource() method in the builder. This operation ensures one source consumes most of the polling interval while also checking intermittently for events on the other sources. The MultiStreamableMessageSource defaults the longPollingSource to the last configured source.

Consider the following sample when constructing a MultiStreamableMessageSource:

public class AxonConfig {
    // omitting other configuration methods...
    public MultiStreamableMessageSource buildMultiStreamableMessageSource(
            StreamableMessageSource<TrackedEventMessage<?>> eventSourceA,
            StreamableMessageSource<TrackedEventMessage<?>> eventSourceB,
            Comparator<Map.Entry<String, TrackedEventMessage<?>>> priorityA
    ) {
        return MultiStreamableMessageSource.builder()
                                           .addMessageSource("eventSourceA", eventSourceA)
                                           .addMessageSource("eventSourceB", eventSourceB)
                                           .longPollingSource("eventSourceA") // Overrides eventSourceB as the longPollingStream
                                           .trackedEventComparator(priorityA) // Where 'priorityA' is a comparator prioritizing events from eventSourceA
                                           .build();
    }
}

Assuming a buildMultiStreamableMessageSource(...) method is present, we can use the outcome to register a processor with the configuring EventProcessingConfigurer:

public class AxonConfig {
    // omitting other configuration methods...
    public void configureTrackingProcessor(EventProcessingConfigurer processingConfigurer) {
        processingConfigurer.registerTrackingEventProcessor(
                "my-processor", config -> buildMultiStreamableMessageSource(/*...*/)
        );
    }
}

Last updated