Since content-based routing often requires some domain-specific logic, most use-cases will require Spring Integration's options for delegating to POJOs using the XML namespace support and/or Annotations. Both of these are discussed below, but first we present a couple implementations that are available out-of-the-box since they fulfill common requirements.

<bean id="payloadTypeRouter" class="org.springframework.integration.router.PayloadTypeRouter">
    <property name="channelIdentifierMap">
        <map>
            <entry key="java.lang.String" value-ref="stringChannel"/>
            <entry key="java.lang.Integer" value-ref="integerChannel"/>
        </map>
    </property>
</bean>

<payload-type-router input-channel="routingChannel">
  <mapping type="java.lang.String" channel="stringChannel" />
  <mapping type="java.lang.Integer" channel="integerChannel" />
</payload-type-router>

HeaderValueRouter

A HeaderValueRouter will send Messages to the channel based on the individual header value mappings. When a HeaderValueRouter is created it is initialized with the name of the header to be evaluated. The value of the header could be one of two things:

1. Arbitrary value

2. Channel name

If arbitrary then additional mappings for these header values to channel names is required, otherwise no additional configuration is needed.

Spring Integration provides a simple namespace-based XML configuration to configure a HeaderValueRouter. The example below demonstrates two types of namespace-based configuration for the HeaderValueRouter.

1. Configuration where mapping of header values to channels is required

<header-value-router input-channel="routingChannel" header-name="testHeader">
  <mapping value="someHeaderValue" channel="channelA" />
  <mapping value="someOtherHeaderValue" channel="channelB" />
</header-value-router>

During the resolution process this router may encounter channel resolution failures, causing an exception. If you want to suppress such exceptions and send unresolved messages to the default output channel (identified with the default-output-channel attribute) set ignore-channel-name-resolution-failures to true. Normally, messages for which the header value is not explicitly mapped to a channel will be sent to the default-output-channel. However, in cases where the header value is mapped to a channel name but the channel cannot be resolved, setting the ignore-channel-name-resolution-failures attribute to true will result in routing such messages to the default-output-channel.

2. Configuration where mapping of header values to channel names is not required since header values themselves represent channel names

<header-value-router input-channel="routingChannel" header-name="testHeader"/>

	Note
	The two router implementations shown above share some common attributes, such as default-output-channel and resolution-required. If resolution-required is set to true, and the router is unable to determine a target channel (e.g. there is no matching payload for a PayloadTypeRouter and no default-output-channel has been specified), then an Exception will be thrown.

<bean id="recipientListRouter" class="org.springframework.integration.router.RecipientListRouter">
    <property name="channels">
        <list>
            <ref bean="channel1"/>
            <ref bean="channel2"/>
            <ref bean="channel3"/>
        </list>
    </property>
</bean>

Spring Integration also provides namespace support for the RecipientListRouter configuration (see Section B.2, “Namespace Support”) as the example below demonstrates.

<recipient-list-router id="customRouter" input-channel="routingChannel"
        timeout="1234"
        ignore-send-failures="true"
        apply-sequence="true">
  <recipient channel="channel1"/>
  <recipient channel="channel2"/>
</recipient-list-router>

	Note
	The 'apply-sequence' flag here has the same effect as it does for a publish-subscribe-channel, and like a publish-subscribe-channel, it is disabled by default on the recipient-list-router. Refer to the section called “PublishSubscribeChannel Configuration” for more information.

Another convenient option when configuring a RecipientListRouter is to use Spring Expression Language (SpEL) support as selectors for individual recipient channels. This is similar to using a Filter at the beginning of 'chain' to act as a "Selective Consumer". However, in this case, it's all combined rather concisely into the router's configuration.

<int:recipient-list-router id="customRouter" input-channel="routingChannel">
	<int:recipient channel="channel1" selector-expression="payload.equals('foo')"/>
	<int:recipient channel="channel2" selector-expression="headers.contains('bar')"/>
</int:recipient-list-router>

In the above configuration a SpEL expression identified by the selector-expression attribute will be evaluated to determine if this recipient should be included in the recipient list for a given input Message. The evaluation result of the expression must be a boolean. If this attribute is not defined, the channel will always be among the list of recipients.

So as you can see, Spring Integration provides quite a few different router configurations for common content-based routing use cases as well as the option of implementing custom routers as POJOs. For example PayloadTypeRouter provides a simple way to configure a router which computes channels based on the payload type of the incoming Message while HeaderValueRouter provides the same convenience in configuring a router which computes channels by evaluating the value of a particular Message Header. There are also expression-based (SpEL) routers where the channel is determined based on evaluating an expression. Thus, these type of routers exhibit some dynamic characteristics.

However these routers all require static configuration. Even in the case of expression-based routers, the expression itself is defined as part of the router configuration which means that the same expression operating on the same value will always result in the computation of the same channel. This is acceptable in most cases since such routes are well defined and therefore predictable. But there are times when we need to change router configurations dynamically so message flows may be routed to a different channel.

Example:

You might want to bring down some part of your system for maintenance and temporarily re-reroute messages to a different message flow. Or you may want to introduce more granularity to your message flow by adding another route to handle a more concrete type of java.lang.Number (in the case of PayloadTypeRouter).

Unfortunately with static router configuration to accomplish this you would have to bring down your entire application, change the configuration of the router (change routes) and bring it back up. This is obviously not the solution.

The Dynamic Router pattern describes the mechanisms by which one can change/configure routers dynamically without bringing down the system or individual routers. 

Before we get into the specifics of how this is accomplished in Spring Integration let's quickly summarize the typical flow of the router, which consists of 3 simple steps:

There is not much that can be done with regard to dynamic routing if Step 1 results in the actual instance of the MessageChannel simply because the MessageChannel is the final product of any router's job. However, if Step 1 results in a channel identifier that is not an instance of MessageChannel, then there are quite a few possibilities to influence the process of deriving the Message Channel. Lets look at couple of the examples in the context of the 3 steps mentioned above: 

Payload Type Router

<payload-type-router input-channel="routingChannel">
  <mapping type="java.lang.String" channel="channel1" />
  <mapping type="java.lang.Integer" channel="channel2" />
</payload-type-router>

Within the context of the Payload Type Router the 3 steps mentioned above would be realized as:

In other words each step feeds the next step until the process completes.

Header Value Router

<header-value-router input-channel="inputChannel" header-name="testHeader">
  <mapping value="foo" channel="fooChannel" />
  <mapping value="bar" channel="barChannel" />
</header-value-router>

Similar to the PayloadTypeRouter:

The above two configurations of two different router types look almost identical. However if we look at the alternate configuration of the HeaderValueRouter we clearly see that there is no mapping sub element:

<header-value-router input-channel="inputChannel" header-name="testHeader">

But the configuration is still perfectly valid. So the natural question is what about the mapping in the Step 2?

What this means is that Step 2 is now an optional step. If mapping is not defined then the channel identifier value computed in Step 1 will automatically be treated as the channel name which will now be resolved to the actual MessageChannel in the Step 3. What it also means is that Step 2 is one of the key steps to provide dynamic characteristics to the routers, since it introduces a process which allows you to change the way 'channel identifier' resolves to 'channel name', thus influencing the process of determining the final instance of the MessageChannel from the initial channel identifier. 

For Example:

In the above configuration let's assume that the testHeader value is 'kermit' which is now a channel identifier (Step 1). Since there is no mapping in this router, resolving this channel identifier to a channel name (Step 2) is impossible and this channel identifier is now treated as channel name. However what if there was a mapping but for a different value? The end result would still be the same and that is: if new value cannot be determined through the process of resolving the 'channel identifier' to a 'channel name', such 'channel identifier' becomes 'channel name'.

So all that is left is for Step 3 to resolve the channel name ('kermit') to an actual instance of the MessageChannel identified by this name. That will be done via the default ChannelResolver implementation which is a BeanFactoryChannelResolver. It basically does a bean lookup for the name provided. So now all messages which contain the header/value pair as testHeader=kermit are going to be routed to a MessageChannel whose bean name (id) is 'kermit'.

But what if you want to route these messages to the 'simpson' channel? Obviously changing a static configuration will work, but will also require bringing your system down. However if you had access to the channel identifier map, then you could just introduce a new mapping where the header/value pair is now kermit=simpson, thus allowing Step 2 to treat 'kermit' as a channel identifier while resolving it to 'simpson' as the channel name .

The same obviously applies for PayloadTypeRouter where you can now remap or remove a particular payload type mapping. In fact, it applies to every other router including expression-based routers since their computed values will now have a chance to go through Step 2 to be additionally resolved to the actual channel name.

In Spring Integration 2.0 the routers hierarchy underwent significant refactoring so that now any router that is a subclass of the AbstractMessageRouter (which includes all framework defined routers) is a Dynamic Router simply because the channelIdentiferMap is defined at the AbstractMessageRouter level. That map's setter method is exposed as a public method along with 'setChannelMapping' and 'removeChannelMapping' methods. These allow you to change/add/remove router mappings at runtime as long as you have a reference to the router itself. It also means that you could expose these same configuration options via JMX (see Section 8.1, “JMX Support”) or the Spring Integration ControlBus (see Section 8.3, “Control Bus”) functionality. 

Control Bus

One way to manage the router mappings is through the Control Bus pattern which exposes a Control Channel where you can send control messages to manage and monitor Spring Integration components, including routers. For more information about the Control Bus see Section 8.3, “Control Bus”. Typically you would send a control message asking to invoke a particular operation on a particular managed component (e.g., router). The two managed operations (methods) that are specific to changing the router resolution process are:

You can also expose a router instance with Spring's JMS support and then use your favorite JMX client (e.g., JConsole) to manage those operations (methods) for changing the router's configuration. For more information on Spring Integration management and monitoring please visit Section 8.1, “JMX Support”.

Message Filters are used to decide whether a Message should be passed along or dropped based on some criteria such as a Message Header value or Message content itself. Therefore, a Message Filter is similar to a router, except that for each Message received from the filter's input channel, that same Message may or may not be sent to the filter's output channel. Unlike the router, it makes no decision regarding which Message Channel to send the Message to but only decides whether to send.

	Note
	As you will see momentarily, the Filter also supports a discard channel, so in certain cases it can play the role of a very simple router (or "switch") based on a boolean condition.

In Spring Integration, a Message Filter may be configured as a Message Endpoint that delegates to an implementation of the MessageSelector interface. That interface is itself quite simple:

 public interface MessageSelector {

     boolean accept(Message<?> message);

 }

The MessageFilter constructor accepts a selector instance:

 MessageFilter filter = new MessageFilter(someSelector);

The <filter> element is used to create a Message-selecting endpoint. In addition to "input-channel and output-channel attributes, it requires a ref. The ref may point to a MessageSelector implementation:

 <filter input-channel="input" ref="selector" output-channel="output"/>

 <bean id="selector" class="example.MessageSelectorImpl"/>

Alternatively, the method attribute can be added at which point the ref may refer to any object. The referenced method may expect either the Message type or the payload type of inbound Messages. The method must return a boolean value. If the method returns 'true', the Message will be sent to the output-channel.

 <filter input-channel="input" output-channel="output"
         ref="exampleObject" method="someBooleanReturningMethod"/>

 <bean id="exampleObject" class="example.SomeObject"/>

If the selector or adapted POJO method returns false, there are a few settings that control the handling of the rejected Message. By default (if configured like the example above), rejected Messages will be silently dropped. If rejection should instead result in an error condition, then set the throw-exception-on-rejection attribute to true:

 <filter input-channel="input" ref="selector"
      output-channel="output" throw-exception-on-rejection="true"/> 

If you want rejected messages to be routed to a specific channel, provide that reference as the discard-channel:

 <filter input-channel="input" ref="selector"
      output-channel="output" discard-channel="rejectedMessages"/> 

Note

Message Filters are commonly used in conjunction with a Publish Subscribe Channel. Many filter endpoints may be subscribed to the same channel, and they decide whether or not to pass the Message to the next endpoint which could be any of the supported types (e.g. Service Activator). This provides a reactive alternative to the more proactive approach of using a Message Router with a single Point-to-Point input channel and multiple output channels.

Using a ref attribute is generally recommended if the custom filter implementation is referenced in other <filter> definitions. However if the custom filter implementation is scoped to a single <filter> element, provide an inner bean definition:

<filter method="someMethod" input-channel="inChannel" output-channel="outChannel">
  <beans:bean class="org.foo.MyCustomFilter"/>
</filter>

	Note
	Using both the ref attribute and an inner handler definition in the same <filter> configuration is not allowed, as it creates an ambiguous condition, and an Exception will be thrown.

With the introduction of SpEL support, Spring Integration added the expression attribute to the filter element. It can be used to avoid Java entirely for simple filters.

        <filter input-channel="input" expression="payload.equals('nonsense')"/>
      

The string passed as the expression attribute will be evaluated as a SpEL expression with the Message available in the evaluation context. If it is necessary to include the result of an expression in the scope of the application context you can use the #{} notation as defined in the SpEL reference documentation .

        <filter input-channel="input" expression="payload.matches(#{filterPatterns.nonsensePattern})"/>
      

If the Expression itself needs to be dynamic, then an 'expression' sub-element may be used. That provides a level of indirection for resolving the Expression by its key from an ExpressionSource. That is a strategy interface that you can implement directly, or you can rely upon a version available in Spring Integration that loads Expressions from a "resource bundle" and can check for modifications after a given number of seconds. All of this is demonstrated in the following configuration sample where the Expression could be reloaded within one minute if the underlying file had been modified. If the ExpressionSource bean is named "expressionSource", then it is not necessary to provide the source attribute on the <expression> element, but in this case it's shown for completeness.

        
  <filter input-channel="input" output-channel="output">
      <expression key="filterPatterns.example" source="myExpressions"/>
  </filter>

  <beans:bean id="myExpressions" id="myExpressions"
          class="org.springframework.integration.expression.ReloadableResourceBundleExpressionSource">
      <beans:property name="basename" value="config/integration/expressions"/>
      <beans:property name="cacheSeconds" value="60"/>
  </beans:bean>

Then, the 'config/integration/expressions.properties' file (or any more specific version with a locale extension to be resolved in the typical way that resource-bundles are loaded) would contain a key/value pair:

          filterPatterns.example=payload > 100

Note

All of these examples that use expression as an attribute or sub-element can also be applied within transformer, router, splitter, service-activator, and header-enricher elements. Of course, the semantics/role of the given component type would affect the interpretation of the evaluation result in the same way that the return value of a method-invocation would be interpreted. For example, an expression can return Strings that are to be treated as Message Channel names by a router component. However, the underlying functionality of evaluating the expression against the Message as the root object, and resolving bean names if prefixed with '@' is consistent across all of the core EIP components within Spring Integration.

The Splitter is a component whose role is to partition a message in several parts, and send the resulting messages to be processed independently. Very often, they are upstream producers in a pipeline that includes an Aggregator.

The API for performing splitting consists of one base class, AbstractMessageSplitter, which is a MessageHandler implementation, encapsulating features which are common to splitters, such as filling in the appropriate message headers CORRELATION_ID, SEQUENCE_SIZE, and SEQUENCE_NUMBER on the messages that are produced. This enables tracking down the messages and the results of their processing (in a typical scenario, these headers would be copied over to the messages that are produced by the various transforming endpoints), and use them, for example, in a Composed Message Processor scenario.

An excerpt from AbstractMessageSplitter can be seen below:

public abstract class AbstractMessageSplitter
                extends AbstractReplyProducingMessageConsumer {
  ...
  protected abstract Object splitMessage(Message<?> message);

}

To implement a specific Splitter in an application, extend AbstractMessageSplitter and implement the splitMessage method, which contains logic for splitting the messages. The return value may be one of the following:

In Spring Integration, any POJO can implement the splitting algorithm, provided that it defines a method that accepts a single argument and has a return value. In this case, the return value of the method will be interpreted as described above. The input argument might either be a Message or a simple POJO. In the latter case, the splitter will receive the payload of the incoming message. Since this decouples the code from the Spring Integration API and will typically be easier to test, it is the recommended approach.

Basically a mirror-image of the Splitter, the Aggregator is a type of Message Handler that receives multiple Messages and combines them into a single Message. In fact, an Aggregator is often a downstream consumer in a pipeline that includes a Splitter.

Technically, the Aggregator is more complex than a Splitter, because it is stateful as it must hold the Messages to be aggregated and determine when the complete group of Messages is ready to be aggregated. In order to do this it requires a MessageStore.

The Aggregator combines a group of related messages, by correlating and storing them, until the group is deemed complete. At that point, the Aggregator will create a single message by processing the whole group, and will send the aggregated message as output.

Implementing an Aggregator requires providing the logic to perform the aggregation (i.e., the creation of a single message from many). Two related concepts are correlation and release.

Correlation determines how messages are grouped for aggregation. In Spring Integration correlation is done by default based on the CORRELATION_ID message header. Messages with the same CORRELATION_ID will be grouped together. However, the correlation strategy may be customized to allow other ways of specifying how the messages should be grouped together by implementing a CorrelationStrategy (see below).

To determine the point at which a group of messages is ready to be processed, a ReleaseStrategy is consulted. The default release strategy for the Aggregator will release a group when all messages included in a sequence are present, based on the SEQUENCE_SIZE header. This default strategy may be overridden by providing a reference to a custom ReleaseStrategy implementation.

The Aggregation API consists of a number of classes:

CorrelatingMessageHandler

The CorrelatingMessageHandler is a MessageHandler implementation, encapsulating the common functionalities of an Aggregator (and other correlating use cases), which are:

• correlating messages into a group to be aggregated

• maintaining those messages in a MessageStore until the group can be released

• deciding when the group can be released

• aggregating the released group into a single message

• recognizing and responding to an expired group

The responsibility of deciding how the messages should be grouped together is delegated to a CorrelationStrategy instance. The responsibility of deciding whether the message group can be released is delegated to a ReleaseStrategy instance.

Here is a brief highlight of the base AbstractAggregatingMessageGroupProcessor (the responsibility of implementing the aggregatePayloads method is left to the developer):

public abstract class AbstractAggregatingMessageGroupProcessor
              implements MessageGroupProcessor {

    protected Map<String, Object> aggregateHeaders(MessageGroup group) {
        // default implementation exists
    }

  	protected abstract Object aggregatePayloads(MessageGroup group, Map<String, Object> defaultHeaders);

}

The CorrelationStrategy is owned by the CorrelatingMessageHandler and it has a default value based on the CORRELATION_ID message header:

    public CorrelatingMessageHandler(MessageGroupProcessor processor, MessageGroupStore store,
                   CorrelationStrategy correlationStrategy, ReleaseStrategy releaseStrategy) {
     ...
      this.correlationStrategy = correlationStrategy == null ?
        new HeaderAttributeCorrelationStrategy(MessageHeaders.CORRELATION_ID) : correlationStrategy;
      this.releaseStrategy = releaseStrategy == null ? new SequenceSizeReleaseStrategy() : releaseStrategy;
      ...
  }

As for actual processing of the message group, the default implementation is the DefaultAggregatingMessageGroupProcessor. It creates a single Message whose payload is a List of the payloads received for a given group. This works well for simple Scatter Gather implementations with either a Splitter, Publish Subscribe Channel, or Recipient List Router upstream.

Note

When using a Publish Subscribe Channel or Recipient List Router in this type of scenario, be sure to enable the flag to apply-sequence. That will add the necessary headers (CORRELATION_ID, SEQUENCE_NUMBER and SEQUENCE_SIZE). That behavior is enabled by default for Splitters in Spring Integration, but it is not enabled for the Publish Subscribe Channel or Recipient List Router because those components may be used in a variety of contexts in which these headers are not necessary.

When implementing a specific aggregator strategy for an application, a developer can extend AbstractAggregatingMessageGroupProcessor and implement the aggregatePayloads method. However, there are better solutions, less coupled to the API, for implementing the aggregation logic which can be configured easily either through XML or through annotations.

In general, any POJO can implement the aggregation algorithm if it provides a method that accepts a single java.util.List as an argument (parameterized lists are supported as well). This method will be invoked for aggregating messages as follows:

• if the argument is a java.util.List<T>, and the parameter type T is assignable to Message, then the whole list of messages accumulated for aggregation will be sent to the aggregator

• if the argument is a non-parameterized java.util.List or the parameter type is not assignable to Message, then the method will receive the payloads of the accumulated messages

• if the return type is not assignable to Message, then it will be treated as the payload for a Message that will be created automatically by the framework.

	Note
	In the interest of code simplicity, and promoting best practices such as low coupling, testability, etc., the preferred way of implementing the aggregation logic is through a POJO, and using the XML or annotation support for configuring it in the application.

public interface ReleaseStrategy {

  boolean canRelease(MessageGroup group);

}

public interface CorrelationStrategy {

  Object getCorrelationKey(Message<?> message);

}

Aggregator (and some other patterns in Spring Integration) is a stateful pattern that requires decisions to be made based on a group of messages that have arrived over a period of time, all with the same correlation key. The design of the interfaces in the stateful patterns (e.g. ReleaseStrategy) is driven by the principle that the components (whether defined by the framework or a user) should be able to remain stateless. All state is carried by the MessageGroup and its management is delegated to the MessageGroupStore.

The MessageGroupStore accumulates state information in MessageGroups while waiting for a release strategy to be triggered, and that event might not ever happen. So to prevent stale messages from lingering, and for volatile stores to provide a hook for cleaning up when the application shuts down, the MessageGroupStore allows the user to register callbacks to apply to its MessageGroups when they expire. The interface is very straightforward:

public interface MessageGroupCallback {

  void execute(MessageGroupStore messageGroupStore, MessageGroup group);

}

The callback has direct access to the store and the message group so it can manage the persistent state (e.g. by removing the group from the store entirely).

The MessageGroupStore maintains a list of these callbacks which it applies, on demand, to all messages whose timestamp is earlier than a time supplied as a parameter:

public interface MessageGroupStore {
   void registerMessageGroupExpiryCallback(MessageGroupCallback callback);
   int expireMessageGroups(long timeout);
}

The expireMessageGroups method can be called with a timeout value: any message older than the current time minus this value will be expired, and have the callbacks applied. Thus it is the user of the store that defines what is meant by message group "expiry".

As a convenience for users, Spring Integration provides a wrapper for the message expiry in the form of a MessageGroupStoreReaper:

<bean id="reaper" class="org...MessageGroupStoreReaper">
  <property name="messageGroupStore" ref="messageStore"/>
  <property name="timeout" value="30000"/>
</bean>

<task:scheduled-tasks scheduler="scheduler">	
  <task:scheduled ref="reaper" method="run" fixed-rate="10000"/>
</task:scheduled-tasks>

The reaper is a Runnable, and all that is happening in the sample above is that the message group store's expire method is being called once every 10 seconds. The timeout itself is 30 seconds.

In addition to the reaper, the expiry callbacks are invoked when the application shuts down via a lifecycle callback in the CorrelatingMessageHandler.

The CorrelatingMessageHandler registers its own expiry callback, and this is the link with the boolean flag send-partial-result-on-expiry in the XML configuration of the aggregator. If the flag is set to true, then when the expiry callback is invoked, any unmarked messages in groups that are not yet released can be sent on to the output channel.

Related to the Aggregator, albeit different from a functional standpoint, is the Resequencer.

The Resequencer works in a similar way to the Aggregator, in the sense that it uses the CORRELATION_ID to store messages in groups, the difference being that the Resequencer does not process the messages in any way. It simply releases them in the order of their SEQUENCE_NUMBER header values.

With respect to that, the user might opt to release all messages at once (after the whole sequence, according to the SEQUENCE_SIZE, has been released), or as soon as a valid sequence is available.

Configuring a resequencer requires only including the appropriate element in XML.

A sample resequencer configuration is shown below.

<channel id="inputChannel"/>

<channel id="outputChannel"/>

<resequencer id="completelyDefinedResequencer" 
  input-channel="inputChannel" 
  output-channel="outputChannel" 
  discard-channel="discardChannel" 
  release-partial-sequences="true" 
  message-store="messageStore" 
  send-partial-result-on-expiry="true" 
  send-timeout="86420000"  /> 

	Note
	Since there is no custom behavior to be implemented in Java classes for resequencers, there is no annotation support for it.

The MessageHandlerChain is an implementation of MessageHandler that can be configured as a single Message Endpoint while actually delegating to a chain of other handlers, such as Filters, Transformers, Splitters, and so on. This can lead to a much simpler configuration when several handlers need to be connected in a fixed, linear progression. For example, it is fairly common to provide a Transformer before other components. Similarly, when providing a Filter before some other component in a chain, you are essentially creating a Selective Consumer. In either case, the chain only requires a single input-channel and a single output-channel eliminating the need to define channels for each individual component.

Tip

Spring Integration's Filter provides a boolean property throwExceptionOnRejection. When providing multiple Selective Consumers on the same point-to-point channel with different acceptance criteria, this value should be set to 'true' (the default is false) so that the dispatcher will know that the Message was rejected and as a result will attempt to pass the Message on to other subscribers. If the Exception were not thrown, then it would appear to the dispatcher as if the Message had been passed on successfully even though the Filter had dropped the Message to prevent further processing. If you do indeed want to "drop" the Messages, then the Filter's 'discard-channel' might be useful since it does give you a chance to perform some operation with the dropped message (e.g. send to a JMS queue or simply write to a log).

The handler chain simplifies configuration while internally maintaining the same degree of loose coupling between components, and it is trivial to modify the configuration if at some point a non-linear arrangement is required.

Internally, the chain will be expanded into a linear setup of the listed endpoints, separated by anonymous channels. The reply channel header will not be taken into account within the chain: only after the last handler is invoked will the resulting message be forwarded on to the reply channel or the chain's output channel. Because of this setup all handlers except the last required to implement the MessageProducer interface (which provides a 'setOutputChannel()' method). The last handler only needs an output channel if the outputChannel on the MessageHandlerChain is set.

	Note
	As with other endpoints, the output-channel is optional. If there is a reply Message at the end of the chain, the output-channel takes precedence, but if not available, the chain handler will check for a reply channel header on the inbound Message as a fallback.

In most cases there is no need to implement MessageHandlers yourself. The next section will focus on namespace support for the chain element. Most Spring Integration endpoints, like Service Activators and Transformers, are suitable for use within a MessageHandlerChain.

The <chain> element provides an input-channel attribute, and if the last element in the chain is capable of producing reply messages (optional), it also supports an output-channel attribute. The sub-elements are then filters, transformers, splitters, and service-activators. The last element may also be a router.

 <chain input-channel="input" output-channel="output">
     <filter ref="someSelector" throw-exception-on-rejection="true"/>
     <header-enricher>
         <header name="foo" value="bar"/>
     </header-enricher>
     <service-activator ref="someService" method="someMethod"/>
 </chain>

The <header-enricher> element used in the above example will set a message header named "foo" with a value of "bar" on the message. A header enricher is a specialization of Transformer that touches only header values. You could obtain the same result by implementing a MessageHandler that did the header modifications and wiring that as a bean, but the header-enricher is obviously a simpler option.

Sometimes you need to make a nested call to another chain from within a chain and then come back and continue execution within the original chain. To accomplish this you can utilize a Messaging Gateway by including a <gateway> element. For example:

  <si:chain id="main-chain" input-channel="in" output-channel="out">
    <si:header-enricher>
      <si:header name="name" value="Many" />
    </si:header-enricher>
    <si:service-activator>
      <bean class="org.foo.SampleService" />
    </si:service-activator>
    <si:gateway request-channel="inputA"/>  
  </si:chain>

  <si:chain id="nested-chain-a" input-channel="inputA">
    <si:header-enricher>
      <si:header name="name" value="Moe" />
    </si:header-enricher>
    <si:gateway request-channel="inputB"/> 
    <si:service-activator>
      <bean class="org.foo.SampleService" />
    </si:service-activator>
  </si:chain>

  <si:chain id="nested-chain-b" input-channel="inputB">
    <si:header-enricher>
      <si:header name="name" value="Jack" />
    </si:header-enricher>
    <si:service-activator>
      <bean class="org.foo.SampleService" />
    </si:service-activator>
  </si:chain>

In the above example the nested-chain-a will be called at the end of main-chain processing by the 'gateway' element configured there. While in nested-chain-a a call to a nested-chain-b will be made after header enrichment and then it will come back to finish execution in nested-chain-b. Finally the flow returns to the main-chain. When the nested version of a <gateway> element is defined in the chain, it does not require the service-interface attribute. Instead, it simple takes the message in its current state and places it on the channel defined via the request-channel attribute. When the downstream flow initiated by that gateway completes, a Message will be returned to the gateway and continue its journey within the current chain.