Spring cloud stream unexpected shutdown is not covered by DLQ - spring-kafka

Question

We are using Spring Cloud Stream 2.2 with the Kafka binder. Something we have noticed is if the pod is killed in the middle of doing the job for whatever reason, then we will miss the message to be sent to DLQ.
We are managing exceptions by catching the failure to log it first, and then send the failure to another service to keep track of this situation, and finally throw exception again to be caught by error channel and captured by DLQ. This approach works seamlessly in normal failure, but if the failure has been triggered externally (like unexpected shutdown), then we miss the DLQ part as it seems the corresponding process is killed before reaching out to the error channel. I wonder if this is a known issue as it's impacting the at-least-once guarantee of this framework in our use case.
22:34:48.077 INFO Shutting down ExecutorService
22:34:48.135 INFO Consumer stopped
22:34:48.136 INFO stopped org.springframework.integration.kafka.inbound.KafkaMessageDrivenChannelAdapter#5174b135
22:34:48.155 INFO Registering MessageChannel outbox-usermgmt.event.job-creator-outbox-event-syncs.errors
22:34:48.241 INFO Channel 'application.outbox-usermgmt.event.job-creator-outbox-event-syncs.errors' has 1 subscriber(s).
22:34:48.241 INFO Channel 'application.outbox-usermgmt.event.job-creator-outbox-event-syncs.errors' has 0 subscriber(s).
22:34:48.246 INFO Registering MessageChannel progress-report.errors
22:34:48.258 INFO Channel 'application.progress-report.errors' has 0 subscriber(s).
22:34:48.262 INFO Registering MessageChannel job-created.errors
22:34:48.273 INFO Registering MessageChannel progress-report.errors
22:34:48.350 INFO Channel 'application.job-created.errors' has 0 subscriber(s).
22:34:48.366 INFO Registering MessageChannel job-created.errors
22:34:48.458 INFO Removing {logging-channel-adapter:_org.springframework.integration.errorLogger} as a subscriber to the 'errorChannel' channel
22:34:48.458 INFO Channel 'application.errorChannel' has 1 subscriber(s).
22:34:48.459 INFO stopped _org.springframework.integration.errorLogger
22:34:48.459 INFO Shutting down ExecutorService 'taskScheduler'
22:34:48.467 WARN Destroy method 'close' on bean with name 'genericSpecificFlexibleDeserializer' threw an exception: java.lang.NullPointerException
22:34:48.472 ERROR Job has failed, Fail to retrieve record's full tree, Connection closed unexpectedly
22:34:48.472 ERROR Fail to retrieve record's full tree
22:34:48.472 DEBUG Sending progress update of 0.0 with status of failed
22:34:48.474 ERROR Job has failed, Fail to retrieve record's full tree
22:34:48.538 INFO Closing JPA EntityManagerFactory for persistence unit 'default'
22:34:48.538 INFO Shutting down ExecutorService
22:34:48.541 INFO HikariPool-1 - Shutdown initiated...
22:34:48.543 INFO HikariPool-1 - Shutdown completed.
Code snippet:
try {
...
} catch (Exception ex) {
//capture the failure details in logs
//send failure progress update to another service
throw new JobProcessingException(ex);
}
It appears the framework commits the message before ensuring that the DLQ message is published to Kafka so the offset has moved but the message was skipped as nothing was published to DLQ.
P.S: This scenario happens for us whenever Kubernetes sends a restart signal to the pod for whatever reason like pod eviction, new release, etc. So I suppose if the kill signal was forced then we would not have the commit in the first place and the job was restarted.

Answer 1

This is a known problem - see https://github.com/spring-projects/spring-integration/issues/3450
The issue is that a PublishSubscribeChannel allows zero subscribers and no exception is thrown if there are none.
It has been resolved in Spring Integration (5.4.x) but is still a problem in the binder because it creates a pub/sub error channel by default.
See my comment there...
Yes; I think that solution makes sense; it shouldn't cause any real problems because the default errorChannel always gets one subscriber.
However, it won't solve the problem in the binder because the message producer gets a binding-specific error channel (which is bridged to the global error channel), so we'd need a similar change there.
It should be possible to work around it by declaring the binding's error channel as a DirectChannel #Bean in which case an exception will be thrown if the consumer has unsubscribed (during shutdown). However, this will mean errors will only go to the binding-specific error channel and won't be bridged to the global errorChannel.
https://github.com/spring-cloud/spring-cloud-stream/issues/2082