Error Handling

As described in the overview at the very beginning of this manual, one of the main motivations behind a message-oriented framework such as Spring Integration is to promote loose coupling between components. The message channel plays an important role, in that producers and consumers do not have to know about each other. However, the advantages also have some drawbacks. Some things become more complicated in a loosely coupled environment, and one example is error handling.

When sending a message to a channel, the component that ultimately handles that message may or may not be operating within the same thread as the sender. If using a simple default DirectChannel (when the <channel> element that has no <queue> child element and no 'task-executor' attribute), the message handling occurs in the same thread that sends the initial message. In that case, if an Exception is thrown, it can be caught by the sender, or it may propagate past the sender if it is an uncaught RuntimeException. This is the same behavior as an exception-throwing operation in a normal Java call stack.

A message flow that runs on a caller thread might be invoked through a messaging gateway (see Messaging Gateways) or a MessagingTemplate (see MessagingTemplate). In either case, the default behavior is to throw any exceptions to the caller. For the messaging gateway, see Error Handling for details about how the exception is thrown and how to configure the gateway to route the errors to an error channel instead. When using a MessagingTemplate or sending to a MessageChannel directly, exceptions are always thrown to the caller.

When adding asynchronous processing, things become rather more complicated. For instance, if the 'channel' element does provide a 'queue' child element (QueueChannel in Java & Annotations Configuration), the component that handles the message operates in a different thread than the sender. The same is true when an ExecutorChannel is used. The sender may have dropped the Message into the channel and moved on to other things. There is no way for the Exception to be thrown directly back to that sender by using standard Exception throwing techniques. Instead, handling errors for asynchronous processes requires that the error-handling mechanism also be asynchronous.

Spring Integration supports error handling for its components by publishing errors to a message channel. Specifically, the Exception becomes the payload of a Spring Integration ErrorMessage. That Message is then sent to a message channel that is resolved in a way that is similar to the 'replyChannel' resolution. First, if the request Message being handled at the time the Exception occurred contains an 'errorChannel' header (the header name is defined in the MessageHeaders.ERROR_CHANNEL constant), the ErrorMessage is sent to that channel. Otherwise, the error handler sends to a “global” channel whose bean name is errorChannel (this is also defined as a constant: IntegrationContextUtils.ERROR_CHANNEL_BEAN_NAME).

A default errorChannel bean is created internally by the Framework. However, you can define your own if you want to control the settings. The following example shows how to define an error channel in XML configuration backed by a queue with a capacity of 500:

Java
@Bean
QueueChannel errorChannel() {
    return new QueueChannel(500);
}
XML
<int:channel id="errorChannel">
    <int:queue capacity="500"/>
</int:channel>
The default error channel is a PublishSubscribeChannel. By default, it has a LoggingHandler as a subscriber with an ERROR logging level and subscription order as Ordered.LOWEST_PRECEDENCE - 100. If you subscribe additional consuming endpoints, that might throw an exception, and you don’t want to preempt the logging, ensure that the additional handlers have a higher order.

The most important thing to understand here is that the messaging-based error handling applies only to exceptions that are thrown by a Spring Integration task that is executing within a TaskExecutor. This does not apply to exceptions thrown by a handler that operates within the same thread as the sender (for example, through a DirectChannel as described earlier in this section).

When exceptions occur in a scheduled poller task’s execution, those exceptions are wrapped in ErrorMessage instances and sent to the 'errorChannel' as well. This is done via a MessagePublishingErrorHandler injected into the global taskScheduler bean. It is recommended to use that MessagePublishingErrorHandler for any custom taskScheduler if the error handling still has to be done using standard 'errorChannel' integration flow logic. A registered integrationMessagePublishingErrorHandler bean can be used in this case.

To enable global error handling, register a handler on that channel. For example, you can configure Spring Integration’s ErrorMessageExceptionTypeRouter as the handler of an endpoint that is subscribed to the errorChannel. That router can then spread the error messages across multiple channels, based on the Exception type.

Starting with version 4.3.10, Spring Integration provides the ErrorMessagePublisher and the ErrorMessageStrategy. You can use them as a general mechanism for publishing ErrorMessage instances. You can call or extend them in any error handling scenarios. The ErrorMessageSendingRecoverer extends this class as a RecoveryCallback implementation that can be used with retry, such as the RequestHandlerRetryAdvice. The ErrorMessageStrategy is used to build an ErrorMessage based on the provided exception and an AttributeAccessor context. It can be injected into any MessageProducerSupport or MessagingGatewaySupport. The requestMessage is stored under ErrorMessageUtils.INPUT_MESSAGE_CONTEXT_KEY in the AttributeAccessor context. The ErrorMessageStrategy can use that requestMessage as the originalMessage property of the ErrorMessage it creates. The DefaultErrorMessageStrategy does exactly that.

Starting with version 5.2, all the MessageHandlingException instances thrown by the framework components, includes a component BeanDefinition resource and source to determine a configuration point form the exception. In case of XML configuration, a resource is an XML file path and source an XML tag with its id attribute. With Java & Annotation configuration, a resource is a @Configuration class and source is a @Bean method. In most case the target integration flow solution is based on the out-of-the-box components and their configuration options. When an exception happens at runtime, there is no any end-user code involved in stack trace because an execution is against beans, not their configuration. Including a resource and source of the bean definition helps to determine possible configuration mistakes and provides better developer experience.

Starting with version 5.4.3, the default error channel is configured with the property requireSubscribers = true to not silently ignore messages when there are no subscribers on this channel (e.g. when application context is stopped). In this case a MessageDispatchingException is thrown which may lend on the client callback of the inbound channel adapter to negatively acknowledge (or roll back) an original message in the source system for redelivery or other future consideration. To restore the previous behavior (ignore non-dispatched error messages), the global integration property spring.integration.channels.error.requireSubscribers must be set to false. See Global Properties and PublishSubscribeChannel Configuration (if you configure a global errorChannel manually) for more information.

See also Error Handling Sample for more information.