Cédric Schaller
Cédric Schaller

Reputation: 278

Error retrieving state store during graceful shutdown

Context and analysis

Our Spring Boot application relies on information stored in Kafka to answer REST requests. It does so by retrieving information from a global state store via the InteractiveQueryService.

Unfortunately, during a graceful shutdown, ongoing requests lead to the following error:

java.lang.IllegalStateException: Error retrieving state store: my-global-store-name-v0
    at org.springframework.cloud.stream.binder.kafka.streams.InteractiveQueryService.lambda$getQueryableStore$1(InteractiveQueryService.java:153)
    at org.springframework.retry.support.RetryTemplate.doExecute(RetryTemplate.java:344)
    at org.springframework.retry.support.RetryTemplate.execute(RetryTemplate.java:217)
    at org.springframework.cloud.stream.binder.kafka.streams.InteractiveQueryService.getQueryableStore(InteractiveQueryService.java:103)
    ...

Our analysis showed that this error is caused by the fact that WebServerGracefulShutdownLifecycle, which handles the graceful shutdown, has a phase with value Integer.MAX_VALUE - 1024 while StreamsBuilderFactoryManager has a phase with value Integer.MAX_VALUE - 100. This means that StreamsBuilderFactoryManager is shut down before the graceful shutdown is initiated, so that ongoing requests processed as part of the graceful shutdown no longer have access to the global state store.

This is indeed visible in the logs:

2025-01-08 17:04:16.225 DEBUG  org.springframework.context.support.DefaultLifecycleProcessor [Stopping beans in phase 2147483547]
...
2025-01-08 17:04:16.572 DEBUG  org.apache.kafka.streams.processor.internals.GlobalStateManagerImpl [Closing global storage engine my-global-store-name-v0]
...
2025-01-08 17:04:16.577 DEBUG  org.springframework.context.support.DefaultLifecycleProcessor [Bean 'streamsBuilderFactoryManager' completed its stop procedure]
...
2025-01-08 17:04:16.584 DEBUG  org.springframework.context.support.DefaultLifecycleProcessor [Stopping beans in phase 2147482623]  
2025-01-08 17:04:16.584 INFO   org.springframework.boot.web.embedded.tomcat.GracefulShutdown [Commencing graceful shutdown. Waiting for active requests to complete]  
2025-01-08 17:04:16.587 INFO   org.springframework.boot.web.embedded.tomcat.GracefulShutdown [Graceful shutdown complete]  

The Javadoc of StreamsBuilderFactoryManager highlights the fact that choosing a phase close to Integer.MAX_VALUE was a conscious decision:

 * This {@link SmartLifecycle} class ensures that the bean created from it is started very
 * late through the bootstrap process by setting the phase value closer to
 * Integer.MAX_VALUE. This is to guarantee that the {@link StreamsBuilderFactoryBean} on a
 * function with multiple bindings is only started after all the binding phases have completed successfully.

Also, the constant AbstractMessageListenerContainer.DEFAULT_PHASE, which is not used by the StreamsBuilderFactoryManager but which has the same value of Integer.MAX_VALUE - 100, makes the following claim (which unfortunately I was unable to verify):

// The default org.springframework.context.SmartLifecycle phase for listener containers 2147483547.

Question

Based on the analysis above, is it a bug that the value chosen for the phase of StreamsBuilderFactoryManager is above the one of WebServerGracefulShutdownLifecycle (in which case I am happy to create an issue on the github project and provide a fix)?

If not, how can we ensure that ongoing requests still have access to the global state store during the graceful shutdown so that they can be answered properly?

Upvotes: 1

Views: 51

Answers (0)

Related Questions