Spring Integration's MessageHandler interface is implemented by many of the components within the framework. In other words, this is not part of the public API, and a developer would not typically implement MessageHandler directly. Nevertheless, it is used by a Message Consumer for actually handling the consumed Messages, and so being aware of this strategy interface does help in terms of understanding the overall role of a consumer. The interface is defined as follows:

public interface MessageHandler {

    void handleMessage(Message<?> message);

}

Despite its simplicity, this provides the foundation for most of the components that will be covered in the following chapters (Routers, Transformers, Splitters, Aggregators, Service Activators, etc). Those components each perform very different functionality with the Messages they handle, but the requirements for actually receiving a Message are the same, and the choice between polling and event-driven behavior is also the same. Spring Integration provides two endpoint implementations that host these callback-based handlers and allow them to be connected to Message Channels.

Because it is the simpler of the two, we will cover the Event Driven Consumer endpoint first. You may recall that the SubscribableChannel interface provides a subscribe() method and that the method accepts a MessageHandler parameter (as shown in the section called “SubscribableChannel”):

subscribableChannel.subscribe(messageHandler);

Since a handler that is subscribed to a channel does not have to actively poll that channel, this is an Event Driven Consumer, and the implementation provided by Spring Integration accepts a a SubscribableChannel and a MessageHandler:

SubscribableChannel channel = context.getBean("subscribableChannel", SubscribableChannel.class);

EventDrivenConsumer consumer = new EventDrivenConsumer(channel, exampleHandler);

Spring Integration also provides a PollingConsumer, and it can be instantiated in the same way except that the channel must implement PollableChannel:

PollableChannel channel = context.getBean("pollableChannel", PollableChannel.class);

PollingConsumer consumer = new PollingConsumer(channel, exampleHandler);

	Note
	For more information regarding Polling Consumers, please also read Section 3.2, “Poller (Polling Consumer)” as well as Section 3.3, “Channel Adapter”.

There are many other configuration options for the Polling Consumer. For example, the trigger is a required property:

PollingConsumer consumer = new PollingConsumer(channel, handler);

consumer.setTrigger(new IntervalTrigger(30, TimeUnit.SECONDS));

Spring Integration currently provides two implementations of the Trigger interface: IntervalTrigger and CronTrigger. The IntervalTrigger is typically defined with a simple interval (in milliseconds), but also supports an initialDelay property and a boolean fixedRate property (the default is false, i.e. fixed delay):

IntervalTrigger trigger = new IntervalTrigger(1000);
trigger.setInitialDelay(5000);
trigger.setFixedRate(true);

The CronTrigger simply requires a valid cron expression (see the Javadoc for details):

CronTrigger trigger = new CronTrigger("*/10 * * * * MON-FRI");

In addition to the trigger, several other polling-related configuration properties may be specified:

PollingConsumer consumer = new PollingConsumer(channel, handler);

consumer.setMaxMessagesPerPoll(10);
consumer.setReceiveTimeout(5000);

The maxMessagesPerPoll property specifies the maximum number of messages to receive within a given poll operation. This means that the poller will continue calling receive() without waiting until either null is returned or that max is reached. For example, if a poller has a 10 second interval trigger and a maxMessagesPerPoll setting of 25, and it is polling a channel that has 100 messages in its queue, all 100 messages can be retrieved within 40 seconds. It grabs 25, waits 10 seconds, grabs the next 25, and so on.

The receiveTimeout property specifies the amount of time the poller should wait if no messages are available when it invokes the receive operation. For example, consider two options that seem similar on the surface but are actually quite different: the first has an interval trigger of 5 seconds and a receive timeout of 50 milliseconds while the second has an interval trigger of 50 milliseconds and a receive timeout of 5 seconds. The first one may receive a message up to 4950 milliseconds later than it arrived on the channel (if that message arrived immediately after one of its poll calls returned). On the other hand, the second configuration will never miss a message by more than 50 milliseconds. The difference is that the second option requires a thread to wait, but as a result it is able to respond much more quickly to arriving messages. This technique, known as long polling, can be used to emulate event-driven behavior on a polled source.

A Polling Consumer may also delegate to a Spring TaskExecutor, as illustrated in the following example:

PollingConsumer consumer = new PollingConsumer(channel, handler);

TaskExecutor taskExecutor = context.getBean("exampleExecutor", TaskExecutor.class);
consumer.setTaskExecutor(taskExecutor);

Furthermore, a PollingConsumer has a property called adviceChain. This property allows you to specify a List of AOP Advices for handling additional cross cutting concerns including transactions. These advices are applied around the doPoll() method. For more in-depth information, please see the sections AOP Advice chains and Transaction Support under Section 7.1.4, “Namespace Support”.

The examples above show dependency lookups, but keep in mind that these consumers will most often be configured as Spring bean definitions. In fact, Spring Integration also provides a FactoryBean called ConsumerEndpointFactoryBean that creates the appropriate consumer type based on the type of channel, and there is full XML namespace support to even further hide those details. The namespace-based configuration will be featured as each component type is introduced.

Note

Many of the MessageHandler implementations are also capable of generating reply Messages. As mentioned above, sending Messages is trivial when compared to the Message reception. Nevertheless, when and how many reply Messages are sent depends on the handler type. For example, an Aggregator waits for a number of Messages to arrive and is often configured as a downstream consumer for a Splitter which may generate multiple replies for each Message it handles. When using the namespace configuration, you do not strictly need to know all of the details, but it still might be worth knowing that several of these components share a common base class, the AbstractReplyProducingMessageHandler, and it provides a setOutputChannel(..) method.

Throughout the reference manual, you will see specific configuration examples for endpoint elements, such as router, transformer, service-activator, and so on. Most of these will support an input-channel attribute and many will support an output-channel attribute. After being parsed, these endpoint elements produce an instance of either the PollingConsumer or the EventDrivenConsumer depending on the type of the input-channel that is referenced: PollableChannel or SubscribableChannel respectively. When the channel is pollable, then the polling behavior is determined based on the endpoint element's poller sub-element and its attributes.

Below you find a poller with all available configuration options:

<int:poller cron=""                                 
            default="false"                         
            error-channel=""                        
            fixed-delay=""                          
            fixed-rate=""                           
            id=""                                   
            max-messages-per-poll=""                
            receive-timeout=""                      
            ref=""                                  
            task-executor=""                        
            time-unit="MILLISECONDS"                
            trigger="">                             
            <int:advice-chain />                    
            <int:transactional />                   
</int:poller>

Examples

For example, a simple interval-based poller with a 1-second interval would be configured like this:

 <int:transformer input-channel="pollable"
              ref="transformer"
              output-channel="output">
     <int:poller fixed-rate="1000"/>
</int:transformer>

As an alternative to fixed-rate you can also use the fixed-delay attribute.

For a poller based on a Cron expression, use the cron attribute instead:

 <int:transformer input-channel="pollable"
              ref="transformer"
              output-channel="output">
     <int:poller cron="*/10 * * * * MON-FRI"/>
 </int:transformer>

If the input channel is a PollableChannel, then the poller configuration is required. Specifically, as mentioned above, the trigger is a required property of the PollingConsumer class. Therefore, if you omit the poller sub-element for a Polling Consumer endpoint's configuration, an Exception may be thrown. The exception will also be thrown if you attempt to configure a poller on the element that is connected to a non-pollable channel.

It is also possible to create top-level pollers in which case only a ref is required:

 <int:poller id="weekdayPoller" cron="*/10 * * * * MON-FRI"/>

 <int:transformer input-channel="pollable"
              ref="transformer"
              output-channel="output">
    <int:poller ref="weekdayPoller"/>
 </int:transformer>

	Note
	The ref attribute is only allowed on the inner-poller definitions. Defining this attribute on a top-level poller will result in a configuration exception thrown during initialization of the Application Context.

Global Default Pollers

In fact, to simplify the configuration even further, you can define a global default poller. A single top-level poller within an ApplicationContext may have the default attribute with a value of true. In that case, any endpoint with a PollableChannel for its input-channel that is defined within the same ApplicationContext and has no explicitly configured poller sub-element will use that default.

 <int:poller id="defaultPoller" default="true" max-messages-per-poll="5" fixed-rate="3000"/>

 <!-- No <poller/> sub-element is necessary since there is a default -->
 <int:transformer input-channel="pollable"
                  ref="transformer"
                  output-channel="output"/>

Transaction Support

Spring Integration also provides transaction support for the pollers so that each receive-and-forward operation can be performed as an atomic unit-of-work. To configure transactions for a poller, simply add the <transactional/> sub-element. The attributes for this element should be familiar to anyone who has experience with Spring's Transaction management:

<int:poller fixed-delay="1000">
    <int:transactional transaction-manager="txManager"
                       propagation="REQUIRED"
                       isolation="REPEATABLE_READ"
                       timeout="10000"
                       read-only="false"/>
</int:poller>

For more information please refer to Section B.1.1, “Poller Transaction Support”.

AOP Advice chains

Since Spring transaction support depends on the Proxy mechanism  with TransactionInterceptor (AOP Advice) handling transactional behavior of the message flow initiated by the poller, some times there is a need to provide extra Advice(s) to handle other cross cutting behavior associated with the poller. For that poller defines an advice-chain element allowing you to add more advices - class that  implements MethodInterceptor interface.. 

<int:service-activator id="advicedSa" input-channel="goodInputWithAdvice" ref="testBean"
		method="good" output-channel="output">
	<int:poller max-messages-per-poll="1" fixed-rate="10000">
		<int:transactional transaction-manager="txManager" />
		 <int:advice-chain>
			<ref bean="adviceA" />
			<beans:bean class="org.bar.SampleAdvice"/>
		</int:advice-chain>
	</int:poller>
</int:service-activator>

For more information on how to implement MethodInterceptor please refer to AOP sections of Spring reference manual (section 8 and 9). Advice chain can also be applied on the poller that does not have any transaction configuration essentially allowing you to enhance the behavior of the message flow initiated by the poller.

TaskExecutor Support

The polling threads may be executed by any instance of Spring's TaskExecutor abstraction. This enables concurrency for an endpoint or group of endpoints. As of Spring 3.0, there is a task namespace in the core Spring Framework, and its <executor/> element supports the creation of a simple thread pool executor. That element accepts attributes for common concurrency settings such as pool-size and queue-capacity. Configuring a thread-pooling executor can make a substantial difference in how the endpoint performs under load. These settings are available per-endpoint since the performance of an endpoint is one of the major factors to consider (the other major factor being the expected volume on the channel to which the endpoint subscribes). To enable concurrency for a polling endpoint that is configured with the XML namespace support, provide the task-executor reference on its <poller/> element and then provide one or more of the properties shown below:

 <int:poller task-executor="pool" fixed-rate="1000"/>

 <task:executor id="pool"
                pool-size="5-25"
                queue-capacity="20"
                keep-alive="120"/>

If no task-executor is provided, the consumer's handler will be invoked in the caller's thread. Note that the caller is usually the default TaskScheduler (see Section E.3, “Configuring the Task Scheduler”). Also, keep in mind that the task-executor attribute can provide a reference to any implementation of Spring's TaskExecutor interface by specifying the bean name. The executor element above is simply provided for convenience.

As mentioned in the background section for Polling Consumers above, you can also configure a Polling Consumer in such a way as to emulate event-driven behavior. With a long receive-timeout and a short interval-trigger, you can ensure a very timely reaction to arriving messages even on a polled message source. Note that this will only apply to sources that have a blocking wait call with a timeout. For example, the File poller does not block, each receive() call returns immediately and either contains new files or not. Therefore, even if a poller contains a long receive-timeout, that value would never be usable in such a scenario. On the other hand when using Spring Integration's own queue-based channels, the timeout value does have a chance to participate. The following example demonstrates how a Polling Consumer will receive Messages nearly instantaneously.

 <int:service-activator input-channel="someQueueChannel"
        output-channel="output">
     <int:poller receive-timeout="30000" fixed-rate="10"/>

 </int:service-activator>

Using this approach does not carry much overhead since internally it is nothing more then a timed-wait thread which does not require nearly as much CPU resource usage as a thrashing, infinite while loop for example.

When configuring Pollers with a fixed-delay or fixed-rate attribute, the default implementation will use a PeriodicTrigger instance. The PeriodicTrigger is part of the Core Spring Framework and it accepts the interval as a constructor argument, only. Therefore it cannot be changed at runtime.

However, you can define your own implementation of the org.springframework.scheduling.Trigger interface. You could even use the PeriodicTrigger as a starting point. Then, you can add a setter for the interval (period), or you could even embed your own throttling logic within the trigger itself if desired. The period property will be used with each call to nextExecutionTime to schedule the next poll. To use this custom trigger within pollers, declare the bean defintion of the custom Trigger in your application context and inject the dependency into your Poller configuration using the trigger attribute, which references the custom Trigger bean instance. You can now obtain a reference to the Trigger bean and the polling interval can be changed between polls.

For an example, please see the Spring Integration Samples project. It contains a sample called dynamic-poller, which uses a custom Trigger and demonstrates the ability to change the polling interval at runtime.

The sample provides a custom Trigger which implements the org.springframework.scheduling.Trigger interface. The sample's Trigger is based on Spring's PeriodicTrigger implementation. However, the fields of the custom trigger are not final and the properties have explicit getters and setters, allowing to dynamically change the polling period at runtime.

	Note
	It is important to note, though, that because the Trigger method is nextExecutionTime(), any changes to a dynamic trigger will not take effect until the next poll, based on the existing configuration. It is not possible to force a trigger to fire before it's currently configured next execution time.

Throughout the reference manual, you will also see specific configuration and implementation examples of various endpoints which can accept a Message or any arbitrary Object as an input parameter. In the case of an Object, such a parameter will be mapped to a Message payload or part of the payload or header (when using the Spring Expression Language). However there are times when the type of input parameter of the endpoint method does not match the type of the payload or its part. In this scenario we need to perform type conversion. Spring Integration provides a convenient way for registering type converters (using the Spring 3.x ConversionService) within its own instance of a conversion service bean named integrationConversionService. That bean is automatically created as soon as the first converter is defined using the Spring Integration namespace support. To register a Converter all you need is to implement org.springframework.core.convert.converter.Converter and define it via convenient namespace support:

 <int:converter ref="sampleConverter"/>

	<bean id="sampleConverter" class="foo.bar.TestConverter"/>

or as an inner bean:

 <int:converter>
		<bean class="org.springframework.integration.config.xml.ConverterParserTests$TestConverter3"/>
	</int:converter>

If you want the polling to be asynchronous, a Poller can optionally specify a task-executor attribute pointing to an existing instance of any TaskExecutor bean (Spring 3.0 provides a convenient namespace configuration via the task namespace). However, there are certain things you must understand when configuring a Poller with a TaskExecutor. 

The problem is that there are two configurations in place. The Poller and the TaskExecutor, and they both have to be in tune with each other otherwise you might end up creating an artificial memory leak.

Let's look at the following configuration provided by one of the users on the Spring Integration forum (http://forum.springsource.org/showthread.php?t=94519):

<int:service-activator input-channel="publishChannel" ref="myService">
		<int:poller receive-timeout="5000" task-executor="taskExecutor" fixed-rate="50"/>
</int:service-activator>

<task:executor id="taskExecutor" pool-size="20" queue-capacity="20"/>

The above configuration demonstrates one of those out of tune configurations.

The poller keeps scheduling new tasks even though all the threads are blocked waiting for either a new message to arrive, or the timeout to expire. Given that there are 20 threads executing tasks with a 5 second timeout, they will be executed at a rate of 4 per second (5000/20 = 250ms). But, new tasks are being scheduled at a rate of 20 per second, so the internal queue in the task executor will grow at a rate of 16 per second (while the process is idle), so we essentially have a memory leak.

One of the ways to handle this is to set the queue-capacity attribute of the Task Executor to 0. You can also manage it by specifying what to do with messages that can not be queued by setting the rejection-policy attribute of the Task Executor (e.g., DISCARD). In other words there are certain details you must understand with regard to configuring the TaskExecutor. Please refer to - Section 25 - Task Execution and Scheduling of the Spring reference manual for more detail on the subject.

As mentioned above, it would be great to have no dependency on the Spring Integration API at all - including the gateway class. For that reason, Spring Integration provides the GatewayProxyFactoryBean that generates a proxy for any interface and internally invokes the gateway methods shown below. Using dependency injection you can then expose the interface to your business methods.

Here is an example of an interface that can be used to interact with Spring Integration:

package org.cafeteria;

public interface Cafe {

    void placeOrder(Order order);

}

Namespace support is also provided which allows you to configure such an interface as a service as demonstrated by the following example.

<int:gateway id="cafeService"
         service-interface="org.cafeteria.Cafe"
         default-request-channel="requestChannel"
         default-reply-channel="replyChannel"/>

With this configuration defined, the "cafeService" can now be injected into other beans, and the code that invokes the methods on that proxied instance of the Cafe interface has no awareness of the Spring Integration API. The general approach is similar to that of Spring Remoting (RMI, HttpInvoker, etc.). See the "Samples" Appendix for an example that uses this "gateway" element (in the Cafe demo).

Typically you don't have to specify the default-reply-channel, since a Gateway will auto-create a temporary, anonymous reply channel, where it will listen for the reply. However, there are some cases which may prompt you to define a default-reply-channel (or reply-channel with adapter gateways such as HTTP, JMS, etc.).

For some background, we'll quickly discuss some of the inner-workings of the Gateway. A Gateway will create a temporary point-to-point reply channel which is anonymous and is added to the Message Headers with the name replyChannel. When providing an explicit default-reply-channel (reply-channel with remote adapter gateways), you have the option to point to a publish-subscribe channel, which is so named because you can add more than one subscriber to it. Internally Spring Integration will create a Bridge between the temporary replyChannel and the explicitly defined default-reply-channel.

So let's say you want your reply to go not only to the gateway, but also to some other consumer. In this case you would want two things: a) a named channel you can subscribe to and b) that channel is a publish-subscribe-channel. The default strategy used by the gateway will not satisfy those needs, because the reply channel added to the header is anonymous and point-to-point. This means that no other subscriber can get a handle to it and even if it could, the channel has point-to-point behavior such that only one subscriber would get the Message. So by defining a default-reply-channel you can point to a channel of your choosing, which in this case would be a publish-subscribe-channel. The Gateway would create a bridge from it to the temporary, anonymous reply channel that is stored in the header.

Another case where you might want to provide a reply channel explicitly is for monitoring or auditing via an interceptor (e.g., wiretap). You need a named channel in order to configure a Channel Interceptor.

The reason that the attributes on the 'gateway' element are named 'default-request-channel' and 'default-reply-channel' is that you may also provide per-method channel references by using the @Gateway annotation.

 public interface Cafe {

     @Gateway(requestChannel="orders")
     void placeOrder(Order order);

 }

You may alternatively provide such content in method sub-elements if you prefer XML configuration (see the next paragraph).

It is also possible to pass values to be interpreted as Message headers on the Message that is created and sent to the request channel by using the @Header annotation:

 public interface FileWriter {

     @Gateway(requestChannel="filesOut")
     void write(byte[] content, @Header(FileHeaders.FILENAME) String filename);

 }

If you prefer the XML approach of configuring Gateway methods, you can provide method sub-elements to the gateway configuration.

<int:gateway id="myGateway" service-interface="org.foo.bar.TestGateway"
      default-request-channel="inputC">
  <int:method name="echo" request-channel="inputA" reply-timeout="2" request-timeout="200"/>
      <int:method name="echoUpperCase" request-channel="inputB"/>
      <int:method name="echoViaDefault"/>
</int:gateway>

You can also provide individual headers per method invocation via XML. This could be very useful if the headers you want to set are static in nature and you don't want to embed them in the gateway's method signature via @Header annotations. For example, in the Loan Broker example we want to influence how aggregation of the Loan quotes will be done based on what type of request was initiated (single quote or all quotes). Determining the type of the request by evaluating what gateway method was invoked, although possible, would violate the separation of concerns paradigm (the method is a java artifact),  but expressing your intention (meta information) via Message headers is natural in a Messaging architecture.

<int:gateway id="loanBrokerGateway"
         service-interface="org.springframework.integration.loanbroker.LoanBrokerGateway">
  <int:method name="getLoanQuote" request-channel="loanBrokerPreProcessingChannel">
    <int:header name="RESPONSE_TYPE" value="BEST"/>
  </int:method>
  <int:method name="getAllLoanQuotes" request-channel="loanBrokerPreProcessingChannel">
    <int:header name="RESPONSE_TYPE" value="ALL"/>
  </int:method>
</int:gateway>

In the above case you can clearly see how a different value will be set for the 'RESPONSE_TYPE' header based on the gateway's method.

When invoking methods on a Gateway interface that do not have any arguments, the default behavior is to receive a Message from a PollableChannel.

At times however, you may want to trigger no-argument methods so that you can in fact interact with other components downstream that do not require user-provided parameters, e.g. triggering no-argument SQL calls or Stored Procedures.

In order to achieve send-and-receive semantics, you must provide a payload. In order to generate a payload, method parameters on the interface are not necessary. You can either use the @Payload annotation or the payload-expression attribute in XML on the method sub-element. Below please find a few examples of what the payloads could be:

Here is an example using the @Payload annotation:

public interface Cafe {

    @Payload("new java.util.Date()")
    List<Order> retrieveOpenOrders();

}

If a method has no argument and no return value, but does contain a payload expression, it will be treated as a send-only operation.

Of course, the Gateway invocation might result in errors. By default any error that has occurred downstream will be re-thrown as a MessagingException (RuntimeException) upon the Gateway's method invocation. However there are times when you may want to simply log the error rather than propagating it, or you may want to treat an Exception as a valid reply, by mapping it to a Message that will conform to some "error message" contract that the caller understands. To accomplish this, our Gateway provides support for a Message Channel dedicated to the errors via the error-channel attribute. In the example below, you can see that a 'transformer' is used to create a reply Message from the Exception.

<int:gateway id="sampleGateway"
    default-request-channel="gatewayChannel"
    service-interface="foo.bar.SimpleGateway"
    error-channel="exceptionTransformationChannel"/>

<int:transformer input-channel="exceptionTransformationChannel"
        ref="exceptionTransformer" method="createErrorResponse"/>

The exceptionTransformer could be a simple POJO that knows how to create the expected error response objects. That would then be the payload that is sent back to the caller. Obviously, you could do many more elaborate things in such an "error flow" if necessary. It might involve routers (including Spring Integration's ErrorMessageExceptionTypeRouter), filters, and so on. Most of the time, a simple 'transformer' should be sufficient, however.

Alternatively, you might want to only log the Exception (or send it somewhere asynchronously). If you provide a one-way flow, then nothing would be sent back to the caller. In the case that you want to completely suppress Exceptions, you can provide a reference to the global "nullChannel" (essentially a /dev/null approach). Finally, as mentioned above, if no "error-channel" is defined at all, then the Exceptions will propagate as usual.

Important

Exposing the messaging system via simple POJI Gateways obviously provides benefits, but "hiding" the reality of the underlying messaging system does come at a price so there are certain things you should consider. We want our Java method to return as quickly as possible and not hang for an indefinite amount of time while the caller is waiting on it to return (void, return value, or a thrown Exception). When regular methods are used as a proxies in front of the Messaging system, we have to take into account the potentially asynchronous nature of the underlying messaging. This means that there might be a chance that a Message that was initiated by a Gateway could be dropped by a Filter, thus never reaching a component that is responsible for producing a reply. Some Service Activator method might result in an Exception, thus providing no reply (as we don't generate Null messages). So as you can see there are multiple scenarios where a reply message might not be coming. That is perfectly natural in messaging systems. However think about the implication on the gateway method. The Gateway's method input arguments  were incorporated into a Message and sent downstream. The reply Message would be converted to a return value of the Gateway's method. So you might want to ensure that for each Gateway call there will always be a reply Message. Otherwise, your Gateway method might never return and will hang indefinitely. One of the ways of handling this situation is via an Asynchronous Gateway (explained later in this section). Another way of handling it is to explicitly set the reply-timeout attribute. That way, the gateway will not hang any longer than the time specified by the reply-timeout and will return 'null' if that timeout does elapse. Finally, you might want to consider setting downstream flags such as 'requires-reply' on a service-activator or 'throw-exceptions-on-rejection' on a filter. These options will be discussed in more detail in the final section of this chapter.

As a pattern the Messaging Gateway is a very nice way to hide messaging-specific code while still exposing the full capabilities of the messaging system. As you've seen, the GatewayProxyFactoryBean provides a convenient way to expose a Proxy over a service-interface thus giving you POJO-based access to a messaging system (based on objects in your own domain, or primitives/Strings, etc).  But when a gateway is exposed via simple POJO methods which return values it does imply that for each Request message (generated when the method is invoked) there must be a Reply message (generated when the method has returned). Since Messaging systems naturally are asynchronous you may not always be able to guarantee the contract where "for each request there will always be be a reply".  With Spring Integration 2.0 we are introducing support for an Asynchronous Gateway which is a convenient way to initiate flows where you may not know if a reply is expected or how long will it take for replies to arrive.

A natural way to handle these types of scenarios in Java would be relying upon java.util.concurrent.Future instances, and that is exactly what Spring Integration uses to support an Asynchronous Gateway.

From the XML configuration, there is nothing different and you still define Asynchronous Gateway the same way as a regular Gateway.

<int:gateway id="mathService" 
     service-interface="org.springframework.integration.sample.gateway.futures.MathServiceGateway"
     default-request-channel="requestChannel"/>

However the Gateway Interface (service-interface) is a bit different.

public interface MathServiceGateway {
  Future<Integer> multiplyByTwo(int i);
}

As you can see from the example above the return type for the gateway method is a Future. When GatewayProxyFactoryBean sees that the return type of the gateway method is a Future, it immediately switches to the async mode by utilizing an AsyncTaskExecutor. That is all. The call to such a method always returns immediately with a Future instance. Then, you can interact with the Future at your own pace to get the result, cancel, etc. And, as with any other use of Future instances, calling get() may reveal a timeout, an execution exception, and so on.

MathServiceGateway mathService = ac.getBean("mathService", MathServiceGateway.class);
Future<Integer> result = mathService.multiplyByTwo(number);
// do something else here since the reply might take a moment
int finalResult =  result.get(1000, TimeUnit.SECONDS);

For a more detailed example, please refer to the async-gateway sample distributed within the Spring Integration samples.

Asynchronous Gateway and AsyncTaskExecutor

By default GatewayProxyFactoryBean uses org.springframework.core.task.SimpleAsyncTaskExecutor when submitting internal AsyncInvocationTask instances for any gateway method whose return type is Future.class. However the async-executor attribute in the <gateway/> element's configuration allows you to provide a reference to any implementation of java.util.concurrent.Executor available within the Spring application context.

As it was explained earlier, the Gateway provides a convenient way of interacting with a Messaging system via POJO method invocations, but realizing that a typical method invocation, which is generally expected to always return (even with an Exception), might not always map one-to-one to message exchanges (e.g., a reply message might not arrive - which is equivalent to a method not returning). It is important to go over several scenarios especially in the Sync Gateway case and understand the default behavior of the Gateway and how to deal with these scenarios to make the Sync Gateway behavior more predictable regardless of the outcome of the message flow that was initialed from such Gateway.

There are certain attributes that could be configured to make Sync Gateway behavior more predictable, but some of them might not always work as you might have expected. One of them is reply-timeout. So, lets look at the reply-timeout attribute and see how it can/can't influence the behavior of the Sync Gateway in various scenarios. We will look at single-threaded scenario (all components downstream are connected via Direct Channel) and multi-threaded scenarios (e.g., somewhere downstream you may have Pollable or Executor Channel which breaks single-thread boundary)

Long running process downstream

Sync Gateway - single-threaded. If a component downstream is still running (e.g., infinite loop or a very slow service), then setting a reply-timeout has no effect and the Gateway method call will not return until such downstream service exits (via return or exception). Sync Gateway - multi-threaded. If a component downstream is still running (e.g., infinite loop or a very slow service), in a multi-threaded message flow setting the reply-timeout will have an effect by allowing gateway method invocation to return once the timeout has been reached, since the GatewayProxyFactoryBean  will simply poll on the reply channel waiting for a message until the timeout expires. However it could result in a 'null' return from the Gateway method if the timeout has been reached before the actual reply was produced. It is also important to understand that the reply message (if produced) will be sent to a reply channel after the Gateway method invocation might have returned, so you must be aware of that and design your flow with this in mind.

Downstream component returns 'null'

Sync Gateway - single-threaded. If a component downstream returns 'null' and no reply-timeout has been configured, the Gateway method call will hang indefinitely unless: a) a reply-timeout has been configured or b) the requires-reply attribute has been set on the downstream component (e.g., service-activator) that might return 'null'. In this case, an Exception would be thrown and propagated to the Gateway. Sync Gateway - multi-threaded. Behavior is the same as above.

Downstream component return signature is 'void' while Gateway method signature is non-void

Sync Gateway - single-threaded. If a component downstream returns 'void' and no reply-timeout has been configured, the Gateway method call will hang indefinitely unless a reply-timeout has been configured  Sync Gateway - multi-threaded Behavior is the same as above.

Downstream component results in Runtime Exception (regardless of the method signature)

Sync Gateway - single-threaded. If a component downstream throws a Runtime Exception, such exception will be propagated via an Error Message back to the gateway and re-thrown. Sync Gateway - multi-threaded Behavior is the same as above.

Important

It is also important to understand that by default reply-timeout is unbounded* which means that if not explicitly set there are several scenarios (described above) where your Gateway method invocation might hang indefinitely. So, make sure you analyze your flow and if there is even a remote possibility of one of these scenarios to occur, set the reply-timeout attribute to a 'safe' value or, even better, set the requires-reply attribute of the downstream component to 'true' to ensure a timely response as produced by the throwing of an Exception as soon as that downstream component does return null internally. But also, realize that there are some scenarios (see the very first one) where reply-timeout will not help. That means it is also important to analyze your message flow and decide when to use a Sync Gateway vs an Async Gateway. As you've seen the latter case is simply a matter of defining Gateway methods that return Future instances. Then, you are guaranteed to receive that return value, and you will have more granular control over the results of the invocation. Also, when dealing with a Router you should remember that setting the resolution-required attribute to 'true' will result in an Exception thrown by the router if it can not resolve a particular channel. Likewise, when dealing with a Filter, you can set the throw-exception-on-rejection attribute. In both of these cases, the resulting flow will behave like that containing a service-activator with the 'requires-reply' attribute. In other words, it will help to ensure a timely response from the Gateway method invocation. Note * reply-timeout is unbounded for <gateway/> elements (created by the GatewayProxyFactoryBean). Inbound gateways for external integration (ws, http, etc.) share many characteristics and attributes with these gateways. However, for those inbound gateways, the default reply-timeout is 1000 milliseconds (1 second). If a downstream async handoff is made to another thread, you may need to increase this attribute to allow enough time for the flow to complete before the gateway times out.

The Service Activator is the endpoint type for connecting any Spring-managed Object to an input channel so that it may play the role of a service. If the service produces output, it may also be connected to an output channel. Alternatively, an output producing service may be located at the end of a processing pipeline or message flow in which case, the inbound Message's "replyChannel" header can be used. This is the default behavior if no output channel is defined, and as with most of the configuration options you'll see here, the same behavior actually applies for most of the other components we have seen.

To create a Service Activator, use the 'service-activator' element with the 'input-channel' and 'ref' attributes:

<int:service-activator input-channel="exampleChannel" ref="exampleHandler"/>

The configuration above assumes that "exampleHandler" either contains a single method annotated with the @ServiceActivator annotation or that it contains only one public method at all. To delegate to an explicitly defined method of any object, simply add the "method" attribute.

<int:service-activator input-channel="exampleChannel" ref="somePojo" method="someMethod"/>

In either case, when the service method returns a non-null value, the endpoint will attempt to send the reply message to an appropriate reply channel. To determine the reply channel, it will first check if an "output-channel" was provided in the endpoint configuration:

<int:service-activator input-channel="exampleChannel" output-channel="replyChannel"
                   ref="somePojo" method="someMethod"/>

If no "output-channel" is available, it will then check the Message's replyChannel header value. If that value is available, it will then check its type. If it is a MessageChannel, the reply message will be sent to that channel. If it is a String, then the endpoint will attempt to resolve the channel name to a channel instance. If the channel cannot be resolved, then a ChannelResolutionException will be thrown. It it can be resolved, the Message will be sent there. This is the technique used for Request Reply messaging in Spring Integration, and it is also an example of the Return Address pattern.

The argument in the service method could be either a Message or an arbitrary type. If the latter, then it will be assumed that it is a Message payload, which will be extracted from the message and injected into such service method. This is generally the recommended approach as it follows and promotes a POJO model when working with Spring Integration. Arguments may also have @Header or @Headers annotations as described in Section E.5, “Annotation Support”

	Note
	The service method is not required to have any arguments at all, which means you can implement event-style Service Activators, where all you care about is an invocation of the service method, not worrying about the contents of the message. Think of it as a NULL JMS message. An example use-case for such an implementation could be a simple counter/monitor of messages deposited on the input channel.

Using a "ref" attribute is generally recommended if the custom Service Activator handler implementation can be reused in other <service-activator> definitions. However if the custom Service Activator handler implementation is only used within a single definition of the <service-activator>, you can provide an inner bean definition:

<int:service-activator id="exampleServiceActivator" input-channel="inChannel"
            output-channel = "outChannel" method="foo">
    <beans:bean class="org.foo.ExampleServiceActivator"/>
</int:service-activator>

	Note
	Using both the "ref" attribute and an inner handler definition in the same <service-activator> configuration is not allowed, as it creates an ambiguous condition and will result in an Exception being thrown.

Service Activators and the Spring Expression Language (SpEL)

Since Spring Integration 2.0, Service Activators can also benefit from SpEL (http://static.springsource.org/spring/docs/3.0.x/spring-framework-reference/html/expressions.html).

For example, you may now invoke any bean method without pointing to the bean via a ref attribute or including it as an inner bean definition. For example:

<int:service-activator input-channel="in" output-channel="out" 
	expression="@accountService.processAccount(payload, headers.accountId)"/>
	
	<bean id="accountService" class="foo.bar.Account"/>

In the above configuration instead of injecting 'accountService' using a ref or as an inner bean, we are simply using SpEL's @beanId notation and invoking a method which takes a type compatible with Message payload. We are also passing a header value. As you can see, any valid SpEL expression can be evaluated against any content in the Message. For simple scenarios your Service Activators do not even have to reference a bean if all logic can be encapsulated by such an expression.

<int:service-activator input-channel="in" output-channel="out" expression="payload * 2"/>

In the above configuration our service logic is to simply multiply the payload value by 2, and SpEL lets us handle it relatively easy.

A Delayer is a simple endpoint that allows a Message flow to be delayed by a certain interval. When a Message is delayed, the original sender will not block. Instead, the delayed Messages will be scheduled with an instance of org.springframework.scheduling.TaskScheduler to be sent to the output channel after the delay has passed. This approach is scalable even for rather long delays, since it does not result in a large number of blocked sender Threads. On the contrary, in the typical case a thread pool will be used for the actual execution of releasing the Messages. Below you will find several examples of configuring a Delayer.

The <delayer> element is used to delay the Message flow between two Message Channels. As with the other endpoints, you can provide the 'input-channel' and 'output-channel' attributes, but the delayer also has 'default-delay' and 'delay-header-name' attributes that are used to determine the number of milliseconds that each Message should be delayed. The following delays all messages by 3 seconds:

 <int:delayer id="delayer" input-channel="input" default-delay="3000" output-channel="output"/>

If you need per-Message determination of the delay, then you can also provide the name of a header using the 'delay-header-name' attribute:

 <int:delayer id="delayer" input-channel="input" output-channel="output"
          default-delay="3000" delay-header-name="delay"/>

In the example above the 3 second delay would only apply in the case that the header value is not present for a given inbound Message. If you only want to apply a delay to Messages that have an explicit header value, then you can set the 'default-delay' to 0 or don't use it at all (by default it is 0). For any Message that has a delay of 0 (or less), the Message will be sent directly. In fact, if there is not a positive delay value for a Message, it will be sent to the output channel on the calling Thread.

Tip

The delay handler supports header values that represent an interval in milliseconds (any Object whose toString() method produces a value that can be parsed into a Long) as well as java.util.Date instances representing an absolute time. In the first case, the milliseconds will be counted from the current time (e.g. a value of 5000 would delay the Message for at least 5 seconds from the time it is received by the Delayer). With a Date instance, the Message will not be released until that Date occurs. In either case, a value that equates to a non-positive delay, or a Date in the past, will not result in any delay. Instead, it will be sent directly to the output channel on the original sender's Thread. If the header is not a Date, and can not be parsed as a Long, the default delay (if any) will be applied.

The delayer delegates to an instance of Spring's TaskScheduler abstraction. The default scheduler used by the delayer is the ThreadPoolTaskScheduler instance provided by Spring Integration on startup: Section E.3, “Configuring the Task Scheduler”. If you want to delegate to a different scheduler, you can provide a reference through the delayer element's 'scheduler' attribute:

 <int:delayer id="delayer" input-channel="input" output-channel="output"
          delay-header-name="delay"
          scheduler="exampleTaskScheduler"/>

 <task:scheduler id="exampleTaskScheduler" pool-size="3"/>

Tip

If you configure an external ThreadPoolTaskScheduler you can set on this scheduler property waitForTasksToCompleteOnShutdown = true. It allows successful completion of 'delay' tasks, which already in the execution state (releasing the Message), when the application is shutdown. Before Spring Integration 2.2 this property was available on the <delayer> element, because DelayHandler could create its own scheduler on the background. Since 2.2 delayer requires an external scheduler instance and waitForTasksToCompleteOnShutdown was deleted; you should use the scheduler's own configuration.

Tip

Also keep in mind ThreadPoolTaskScheduler has a property errorHandler which can be injected with some implementation of org.springframework.util.ErrorHandler. This handler allows to process an Exception from the thread of the scheduled task sending the delayed message. By default it uses an org.springframework.scheduling.support.TaskUtils$LoggingErrorHandler and you will see a stack trace in the logs. You might want to consider using an org.springframework.integration.channel.MessagePublishingErrorHandler, which sends an ErrorMessage into an error-channel, either from the failed Message's header or into the default error-channel.

The DelayHandler persists delayed Messages into the Message Group in the provided MessageStore. (The 'groupId' is based on required 'id' attribute of <delayer> element.) A delayed message is removed from the MessageStore by the scheduled task just before the DelayHandler sends the Message to the output-channel. If the provided MessageStore is persistent (e.g. JdbcMessageStore) it provides the ability to not lose Messages on the application shutdown. After application startup, the DelayHandler reads Messages from its Message Group in the MessageStore and reschedules them with a delay based on the original arrival time of the Message (if the delay is numeric). For messages where the delay header was a Date, that is used when rescheduling. If a delayed Message remained in the MessageStore more than its 'delay', it will be sent immediately after startup.

The <delayer> can be enriched with mutually exclusive sub-elements <transactional> or <advice-chain>. The List of these AOP Advices is applied to the proxied internal DelayHandler.ReleaseMessageHandler, which has the responsibility to release the Message, after the delay, on a Thread of the scheduled task. It might be used, for example, when the downstream message flow throws an Exception and the ReleaseMessageHandler's transaction will be rolled back. In this case the delayed Message will remain in the persistent MessageStore. You can use any custom org.aopalliance.aop.Advice implementation within the <advice-chain>. A sample configuration of the <delayer> may look like this:

 <int:delayer id="delayer" input-channel="input" output-channel="output"
        delay-header-name="delay"
        message-store="jdbcMessageStore">
     <int:advice-chain>
       <beans:ref bean="customAdviceBean"/>
       <tx:advice>
         <tx:attributes>
          <tx:method name="*" read-only="true"/>
         </tx:attributes>
       </tx:advice>
     </int:advice-chain>
 </int:delayer>

The DelayHandler can be exported as a JMX MBean with managed operations getDelayedMessageCount and reschedulePersistedMessages, which allows the rescheduling of delayed persisted Messages at runtime, for example, if the TaskScheduler has previously been stopped. These operations can be invoked via a Control Bus command:

 Message<String> delayerReschedulingMessage =
                MessageBuilder.withPayload("@'delayer.handler'.reschedulePersistedMessages()").build();
 controlBusChannel.send(delayerReschedulingMessage);

	Note
	For more information regarding the Message Store, JMX and the Control Bus, please read Chapter 8, System Management.

Depending on the complexity of your integration requirements scripts may be provided inline as CDATA in XML configuration or as a reference to a Spring resource containing the script. To enable scripting support Spring Integration defines a ScriptExecutingMessageProcessor which will bind the Message Payload to a variable named payload and the Message Headers to a headers variable, both accessible within the script execution context. All that is left for you to do is write a script that uses these variables. Below are a couple of sample configurations:

Filter

<int:filter input-channel="referencedScriptInput">
   <int-script:script lang="ruby" location="some/path/to/ruby/script/RubyFilterTests.rb"/>
</int:filter>

<int:filter input-channel="inlineScriptInput">
     <int-script:script lang="groovy"><![CDATA[
     return payload == 'good'
   ]]></int-script:script>
</int:filter>

Here, you see that the script can be included inline or can reference a resource location via the location attribute. Additionally the lang attribute corresponds to the language name (or JSR223 alias)

Other Spring Integration endpoint elements which support scripting include router, service-activator, transformer, and splitter. The scripting configuration in each case would be identical to the above (besides the endpoint element).

Another useful feature of Scripting support is the ability to update (reload) scripts without having to restart the Application Context. To accomplish this, specify the refresh-check-delay attribute on the script element:

<int-script:script location="..." refresh-check-delay="5000"/>

In the above example, the script location will be checked for updates every 5 seconds. If the script is updated, any invocation that occurs later than 5 seconds since the update will result in execution of the new script.

<int-script:script location="..." refresh-check-delay="0"/>

In the above example the context will be updated with any script modifications as soon as such modification occurs, providing a simple mechanism for 'real-time' configuration. Any negative number value means the script will not be reloaded after initialization of the application context. This is the default behavior.

	Important
	Inline scripts can not be reloaded.

<int-script:script location="..." refresh-check-delay="-1"/>

Script variable bindings

Variable bindings are required to enable the script to reference variables externally provided to the script's execution context. As we have seen, payload and headers are used as binding variables by default. You can bind additional variables to a script via <variable> sub-elements:

<script:script lang="js" location="foo/bar/MyScript.js">
	<script:variable name="foo" value="foo"/>
	<script:variable name="bar" value="bar"/>
	<script:variable name="date" ref="date"/>
</script:script>

As shown in the above example, you can bind a script variable either to a scalar value or a Spring bean reference. Note that payload and headers will still be included as binding variables.

If you need more control over how variables are generated, you can implement your own Java class using the ScriptVariableGenerator strategy:

public interface ScriptVariableGenerator {
  
	Map<String, Object> generateScriptVariables(Message<?> message);
	
}

This interface requires you to implement the method generateScriptVariables(Message). The Message argument allows you to access any data available in the Message payload and headers and the return value is the Map of bound variables. This method will be called every time the script is executed for a Message. All you need to do is provide an implementation of ScriptVariableGenerator and reference it with the script-variable-generator attribute:

<int-script:script location="foo/bar/MyScript.groovy"
				script-variable-generator="variableGenerator"/>

<bean id="variableGenerator" class="foo.bar.MyScriptVariableGenerator"/>

	Important
	You cannot provide both the script-variable-generator attribute and <variable> sub-element(s) as they are mutually exclusive. Also, custom variable bindings cannot be used with an inline script.

With Spring Integration 2.1, Groovy Support's configuration namespace is an extension of Spring Integration's Scripting Support and shares the core configuration and behavior described in detail in the Section 7.5, “Scripting support” section. Even though Groovy scripts are well supported by generic Scripting Support, Groovy Support provides the Groovy configuration namespace which is backed by the Spring Framework's org.springframework.scripting.groovy.GroovyScriptFactory and related components, offering extended capabilities for using Groovy. Below are a couple of sample configurations:

Filter

<int:filter input-channel="referencedScriptInput">
   <int-groovy:script location="some/path/to/groovy/file/GroovyFilterTests.groovy"/>
</int:filter>

<int:filter input-channel="inlineScriptInput">
     <int-groovy:script><![CDATA[
     return payload == 'good'
   ]]></int-groovy:script>
</int:filter>

As the above examples show, the configuration looks identical to the general Scripting Support configuration. The only difference is the use of the Groovy namespace as indicated in the examples by the int-groovy namespace prefix. Also note that the lang attribute on the <script> tag is not valid in this namespace.

Groovy object customization

If you need to customize the Groovy object itself, beyond setting variables, you can reference a bean that implements org.springframework.scripting.groovy.GroovyObjectCustomizer via the customizer attribute. For example, this might be useful if you want to implement a domain-specific language (DSL) by modifying the MetaClass and registering functions to be available within the script:

<int:service-activator input-channel="groovyChannel">
    <int-groovy:script location="foo/SomeScript.groovy" customizer="groovyCustomizer"/>
</int:service-activator>

<beans:bean id="groovyCustomizer" class="org.foo.MyGroovyObjectCustomizer"/>

Setting a custom GroovyObjectCustomizer is not mutually exclusive with <variable> sub-elements or the script-variable-generator attribute. It can also be provided when defining an inline script.

As described in (EIP), the idea behind the Control Bus is that the same messaging system can be used for monitoring and managing the components within the framework as is used for "application-level" messaging. In Spring Integration we build upon the adapters described above so that it's possible to send Messages as a means of invoking exposed operations. One option for those operations is Groovy scripts.

 <int-groovy:control-bus input-channel="operationChannel"/>

The Control Bus has an input channel that can be accessed for invoking operations on the beans in the application context.

The Groovy Control Bus executes messages on the input channel as Groovy scripts. It takes a message, compiles the body to a Script, customizes it with a GroovyObjectCustomizer, and then executes it. The Control Bus' MessageProcessor exposes all beans in the application context that are annotated with @ManagedResource, implement Spring's Lifecycle interface or extend Spring's CustomizableThreadCreator base class (e.g. several of the TaskExecutor and TaskScheduler implementations).

Important

Be careful about using managed beans with custom scopes (e.g. 'request') in the Control Bus' command scripts, especially inside an async message flow. If The Control Bus' MessageProcessor can't expose a bean from the application context, you may end up with some BeansException during command script's executing. For example, if a custom scope's context is not established, the attempt to get a bean within that scope will trigger a BeanCreationException.

If you need to further customize the Groovy objects, you can also provide a reference to a bean that implements org.springframework.scripting.groovy.GroovyObjectCustomizer via the customizer attribute.

<int-groovy:control-bus input-channel="input"
        output-channel="output"
        customizer="groovyCustomizer"/>

<beans:bean id="groovyCustomizer" class="org.foo.MyGroovyObjectCustomizer"/>

In addition to providing the general mechanism to apply AOP Advice classes in this way, three standard Advices are provided:

These are each described in detail in the following sections.

Retry Advice

The retry advice (org.springframework.integration.handler.advice.RequestHandlerRetryAdvice) leverages the rich retry mechanisms provided by the spring-retry project. The core component of spring-retry is the RetryTemplate, which allows configuration of sophisticated retry scenarios, including RetryPolicy and BackoffPolicy strategies, with a number of implementations, as well as a RecoveryCallback strategy to determine the action to take when retries are exhausted.

Stateless Retry

Stateless retry is the case where the retry activity is handled entirely within the advice, where the thread pauses (if so configured) and retries the action.

Stateful Retry

Stateful retry is the case where the retry state is managed within the advice, but where an exception is thrown and the caller resubmits the request. An example for stateful retry is when we want the message originator (e.g. JMS) to be responsible for resubmitting, rather than performing it on the current thread. Stateful retry needs some mechanism to detect a retried submission.

Further Information

For more information on spring-retry, refer to the project's javadocs, as well as the reference documentation for Spring Batch, where spring-retry originated.

	Caution
	The default back off behavior is no back off - retries are attempted immediately. Using a back off policy that causes threads to pause between attempts may cause performance issues, including excessive memory use and thread starvation. In high volume environments, back off policies should be used with caution.

Configuring the Retry Advice

The following examples use a simple <service-activator />> that always throws an exception:

public class FailingService {

	public void service(String message) {
		throw new RuntimeException("foo");
	}
}

Simple Stateless Retry

This example uses the default RetryTemplate which has a SimpleRetryPolicy which tries 3 times. There is no BackoffPolicy so the 3 attempts are made back-to-back-to-back with no delay between attempts. There is no RecoveryCallback so, the result is to throw the exception to the caller after the final failed retry occurs. In a Spring Integration environment, this final exception might be handled using an error-channel on the inbound endpoint.

<int:service-activator input-channel="input" ref="failer" method="service">
	<int:request-handler-advice-chain>
		<bean class="org.springframework.integration.handler.advice.RequestHandlerRetryAdvice" />
	</request-handler-advice-chain>
</int:service-activator>


DEBUG [task-scheduler-2]preSend on channel 'input', message: [Payload=...]
DEBUG [task-scheduler-2]Retry: count=0
DEBUG [task-scheduler-2]Checking for rethrow: count=1
DEBUG [task-scheduler-2]Retry: count=1
DEBUG [task-scheduler-2]Checking for rethrow: count=2
DEBUG [task-scheduler-2]Retry: count=2
DEBUG [task-scheduler-2]Checking for rethrow: count=3
DEBUG [task-scheduler-2]Retry failed last attempt: count=3

Simple Stateless Retry with Recovery

This example adds a RecoveryCallback to the above example; it uses a ErrorMessageSendingRecoverer to send an ErrorMessage to a channel.

<int:service-activator input-channel="input" ref="failer" method="service">
	<int:request-handler-advice-chain>
		<bean class="org.springframework.integration.handler.advice.RequestHandlerRetryAdvice">
			<property name="recoveryCallback">
				<bean class="org.springframework.integration.handler.advice.ErrorMessageSendingRecoverer">
					<constructor-arg ref="myErrorChannel" />
				</bean>
			</property>
		</bean>
	</request-handler-advice-chain>
</int:int:service-activator>


DEBUG [task-scheduler-2]preSend on channel 'input', message: [Payload=...]
DEBUG [task-scheduler-2]Retry: count=0
DEBUG [task-scheduler-2]Checking for rethrow: count=1
DEBUG [task-scheduler-2]Retry: count=1
DEBUG [task-scheduler-2]Checking for rethrow: count=2
DEBUG [task-scheduler-2]Retry: count=2
DEBUG [task-scheduler-2]Checking for rethrow: count=3
DEBUG [task-scheduler-2]Retry failed last attempt: count=3
DEBUG [task-scheduler-2]Sending ErrorMessage :failedMessage:[Payload=...]

Stateless Retry with Customized Policies, and Recovery

For more sophistication, we can provide the advice with a customized RetryTemplate. This example continues to use the SimpleRetryPolicy but it increases the attempts to 4. It also adds an ExponentialBackoffPolicy where the first retry waits 1 second, the second waits 5 seconds and the third waits 25 (for 4 attempts in all).

<int:service-activator input-channel="input" ref="failer" method="service">
	<int:request-handler-advice-chain>
		<bean class="org.springframework.integration.handler.advice.RequestHandlerRetryAdvice">
			<property name="recoveryCallback">
				<bean class="org.springframework.integration.handler.advice.ErrorMessageSendingRecoverer">
					<constructor-arg ref="myErrorChannel" />
				</bean>
			</property>
			<property name="retryTemplate" ref="retryTemplate" />
		</bean>
	</request-handler-advice-chain>
</int:service-activator>

<bean id="retryTemplate" class="org.springframework.retry.support.RetryTemplate">
	<property name="retryPolicy">
		<bean class="org.springframework.retry.policy.SimpleRetryPolicy">
			<property name="maxAttempts" value="4" />
		</bean>
	</property>
	<property name="backOffPolicy">
		<bean class="org.springframework.retry.backoff.ExponentialBackOffPolicy">
			<property name="initialInterval" value="1000" />
			<property name="multiplier" value="5" />
		</bean>
	</property>
</bean>


27.058 DEBUG [task-scheduler-1]preSend on channel 'input', message: [Payload=...]
27.071 DEBUG [task-scheduler-1]Retry: count=0
27.080 DEBUG [task-scheduler-1]Sleeping for 1000
28.081 DEBUG [task-scheduler-1]Checking for rethrow: count=1
28.081 DEBUG [task-scheduler-1]Retry: count=1
28.081 DEBUG [task-scheduler-1]Sleeping for 5000
33.082 DEBUG [task-scheduler-1]Checking for rethrow: count=2
33.082 DEBUG [task-scheduler-1]Retry: count=2
33.083 DEBUG [task-scheduler-1]Sleeping for 25000
58.083 DEBUG [task-scheduler-1]Checking for rethrow: count=3
58.083 DEBUG [task-scheduler-1]Retry: count=3
58.084 DEBUG [task-scheduler-1]Checking for rethrow: count=4
58.084 DEBUG [task-scheduler-1]Retry failed last attempt: count=4
58.086 DEBUG [task-scheduler-1]Sending ErrorMessage :failedMessage:[Payload=...]

Simple Stateful Retry with Recovery

To make retry stateful, we need to provide the Advice with a RetryStateGenerator implementation. This class is used to identify a message as being a resubmission so that the RetryTemplate can determine the current state of retry for this message. The framework provides a SpelExpressionRetryStateGenerator which determines the message identifier using a SpEL expression. This is shown below; this example again uses the default policies (3 attempts with no back off); of course, as with stateless retry, these policies can be customized.

<int:service-activator input-channel="input" ref="failer" method="service">
	<int:request-handler-advice-chain>
		<bean class="org.springframework.integration.handler.advice.RequestHandlerRetryAdvice">
			<property name="retryStateGenerator">
				<bean class="org.springframework.integration.handler.advice.SpelExpressionRetryStateGenerator">
					<constructor-arg value="headers['jms_messageId']" />
				</bean>
			</property>
			<property name="recoveryCallback">
				<bean class="org.springframework.integration.handler.advice.ErrorMessageSendingRecoverer">
					<constructor-arg ref="myErrorChannel" />
				</bean>
			</property>
		</bean>
	</int:request-handler-advice-chain>
</int:service-activator>



24.351 DEBUG [Container#0-1]preSend on channel 'input', message: [Payload=...]
24.368 DEBUG [Container#0-1]Retry: count=0
24.387 DEBUG [Container#0-1]Checking for rethrow: count=1
24.387 DEBUG [Container#0-1]Rethrow in retry for policy: count=1
24.387 WARN  [Container#0-1]failure occurred in gateway sendAndReceive
org.springframework.integration.MessagingException: Failed to invoke handler
...
Caused by: java.lang.RuntimeException: foo
...
24.391 DEBUG [Container#0-1]Initiating transaction rollback on application exception
...
25.412 DEBUG [Container#0-1]preSend on channel 'input', message: [Payload=...]
25.412 DEBUG [Container#0-1]Retry: count=1
25.413 DEBUG [Container#0-1]Checking for rethrow: count=2
25.413 DEBUG [Container#0-1]Rethrow in retry for policy: count=2
25.413 WARN  [Container#0-1]failure occurred in gateway sendAndReceive
org.springframework.integration.MessagingException: Failed to invoke handler
...
Caused by: java.lang.RuntimeException: foo
...
25.414 DEBUG [Container#0-1]Initiating transaction rollback on application exception
...
26.418 DEBUG [Container#0-1]preSend on channel 'input', message: [Payload=...]
26.418 DEBUG [Container#0-1]Retry: count=2
26.419 DEBUG [Container#0-1]Checking for rethrow: count=3
26.419 DEBUG [Container#0-1]Rethrow in retry for policy: count=3
26.419 WARN  [Container#0-1]failure occurred in gateway sendAndReceive
org.springframework.integration.MessagingException: Failed to invoke handler
...
Caused by: java.lang.RuntimeException: foo
...
26.420 DEBUG [Container#0-1]Initiating transaction rollback on application exception
...
27.425 DEBUG [Container#0-1]preSend on channel 'input', message: [Payload=...]
27.426 DEBUG [Container#0-1]Retry failed last attempt: count=3
27.426 DEBUG [Container#0-1]Sending ErrorMessage :failedMessage:[Payload=...]

Comparing with the stateless examples, you can see that with stateful retry, the exception is thrown to the caller on each failure.

<int:service-activator input-channel="input" ref="failer" method="service">
	<int:request-handler-advice-chain>
		<bean class="org.springframework.integration.handler.advice.RequestHandlerCircuitBreakerAdvice">
			<property name="threshold" value="2" />
			<property name="halfOpenAfter" value="12000" />
		</bean>
	</int:request-handler-advice-chain>
</int:service-activator>


05.617 DEBUG [task-scheduler-1]preSend on channel 'input', message: [Payload=...]
05.638 ERROR [task-scheduler-1]org.springframework.integration.MessageHandlingException: java.lang.RuntimeException: foo
...
10.598 DEBUG [task-scheduler-2]preSend on channel 'input', message: [Payload=...]
10.600 ERROR [task-scheduler-2]org.springframework.integration.MessageHandlingException: java.lang.RuntimeException: foo
...
15.598 DEBUG [task-scheduler-3]preSend on channel 'input', message: [Payload=...]
15.599 ERROR [task-scheduler-3]org.springframework.integration.MessagingException: Circuit Breaker is Open for ServiceActivator
...
20.598 DEBUG [task-scheduler-2]preSend on channel 'input', message: [Payload=...]
20.598 ERROR [task-scheduler-2]org.springframework.integration.MessagingException: Circuit Breaker is Open for ServiceActivator
...
25.598 DEBUG [task-scheduler-5]preSend on channel 'input', message: [Payload=...]
25.601 ERROR [task-scheduler-5]org.springframework.integration.MessageHandlingException: java.lang.RuntimeException: foo
...
30.598 DEBUG [task-scheduler-1]preSend on channel 'input', message: [Payload=foo...]
30.599 ERROR [task-scheduler-1]org.springframework.integration.MessagingException: Circuit Breaker is Open for ServiceActivator

In addition to the provided Advice classes above, you can implement your own Advice classes. While you can provide any implementation of org.aopalliance.aop.Advice, it is generally recommended that you subclass org.springframework.integration.handler.advice.AbstractRequestHandlerAdvice. This has the benefit of avoiding writing low-level Aspect Oriented Programming code as well as providing a starting point that is specifically tailored for use in this environment.

Subclasses need to implement the doInvoke() method:

	/**
	 * Subclasses implement this method to apply behavior to the {@link MessageHandler} callback.execute()
	 * invokes the handler method and returns its result, or null).
	 * @param callback Subclasses invoke the execute() method on this interface to invoke the handler method.
	 * @param target The target handler.
	 * @param message The message that will be sent to the handler.
	 * @return the result after invoking the {@link MessageHandler}.
	 * @throws Exception
	 */
	protected abstract Object doInvoke(ExecutionCallback callback, Object target, Message<?> message) throws Exception;

The callback parameter is simply a convenience to avoid subclasses dealing with AOP directly; invoking the callback.execute() method invokes the message handler.

The target parameter is provided for those subclasses that need to maintain state for a specific handler, perhaps by maintaining that state in a Map, keyed by the target. This allows the same advice to be applied to multiple handlers. The RequestHandlerCircuitBreakerAdvice uses this to keep circuit breaker state for each handler.

The message parameter is the message that will be sent to the handler. While the advice cannot modify the message before invoking the handler, it can modify the payload (if it has mutable properties). Typically, an advice would use the message for logging and/or to send a copy of the message somewhere before or after invoking the handler.

The return value would normally be the value returned by callback.execute(); but the advice does have the ability to modify the return value. Note that only AbstractReplyProducingMessageHandlers return a value.

public class MyAdvice extends AbstractRequestHandlerAdvice {

	@Override
	protected Object doInvoke(ExecutionCallback callback, Object target, Message<?> message) throws Exception {
		// add code before the invocation
		Object result = callback.execute();
		// add code after the invocation
		return result;
	}
}

Note

In addition to the execute() method, the ExecutionCallback provides an additional method cloneAndExecute(). This method must be used in cases where the invocation might be called multiple times within a single execution of doInvoke(), such as in the RequestHandlerRetryAdvice. This is required because the Spring AOP org.springframework.aop.framework.ReflectiveMethodInvocation object maintains state of which advice in a chain was last invoked; this state must be reset for each call. For more information, see the ReflectiveMethodInvocation JavaDocs.