Why do I have to use .fluxTransform(f -> f) on an inbound webflux gateways when using Java DSL?

Question

I've run into failed replies when using the webflux gateway Java DSL in Spring Integration. It only works for the first few requests (<8 to be specific), I'm getting reply errors afterwards:

org.springframework.integration.MessageTimeoutException: failed to receive JMS response within timeout of: 5000ms
    at org.springframework.integration.jms.JmsOutboundGateway.handleRequestMessage(JmsOutboundGateway.java:741) ~[spring-integration-jms-5.3.2.RELEASE.jar:5.3.2.RELEASE]
    Suppressed: reactor.core.publisher.FluxOnAssembly$OnAssemblyException:

When I use .fluxTransform(f -> f) on the inbound gateway OR when I use the non-reactive http outbound gateway, I don't get the errors, even on a jmeter benchmark with thousands of requests.

• Why do I have to call fluxTransform(f -> f) in the first flow to make it work?
• Why does it work without fluxTransform(f -> f) when I use Http.outboundGatewayin the second flow?

Scenario
I've created a route using four gateways for a rather complex setup to make a web request on a remote machine, but I'm

Integration Flow 1:

inbound webflux gateway -> outbound jms gateway

  @Bean
  public IntegrationFlow step1() {
    // request-reply pattern using the jms outbound gateway
    var gateway = Jms.outboundGateway(jmsConnectionFactory)
        .requestDestination("inboundWebfluxQueue")
        .replyDestination("outboundWebfluxQueue")
        .correlationKey("JMSCorrelationID");

    // send a request to jms, wait for the reply and return message payload as response
    return IntegrationFlows.from(webfluxServer("/example/webflux"))
        // won't work consistently without the next line
        .fluxTransform(f -> f)
        .handle(gateway).get();
  }

Integration Flow 2:

inbound jms gateway -> outbound webflux gateway

  @Bean
  public IntegrationFlow step2_using_webflux() {
    var gateway = WebFlux.outboundGateway("http://localhost:8080/actuator/health")
        .httpMethod(HttpMethod.GET)
        .expectedResponseType(String.class)
        // ignore headers
        .mappedResponseHeaders();

    return IntegrationFlows.from(jmsInboundGateway())
        // use webflux outbound gateway to send the request to the TEST_URL
        .handle(gateway).get();
  }

The complete route looks like this:

client web request -> flow 1 -> (message broker) -> flow 2 -> server web request

Answer 1

Another way is to use a .channel(MessageChannels.flux()) instead of that .fluxTransform(f -> f). This way we really bring a back-pressure to the the WebFlux container making it waiting for available slot in the request event loop.

With that we just send to JMS queue not-honoring back-pressure and and your JMS consumer on the other side can't keep up. Plus we send a request to the same Netty server internally acquiring an event loop slot again for those internal requests.

If you are interested I wrote a unit test like this to see what is going on:

@SpringBootTest(webEnvironment = SpringBootTest.WebEnvironment.DEFINED_PORT)
public class IntegrationApplicationTests {


    @Autowired
    private TestRestTemplate template;

    @Test
    void testSpringIntegrationWebFlux() {
        var executor = new ThreadPoolTaskExecutor();
        executor.setCorePoolSize(100);
        executor.setWaitForTasksToCompleteOnShutdown(true);
        executor.setAwaitTerminationSeconds(10);
        executor.afterPropertiesSet();

        var numberOfExecutions = new AtomicInteger();

        for (var i = 0; i < 100; i++) {
            executor.execute(() -> {
                var responseEntity = this.template.getForEntity("/example/webflux", String.class);
                if (responseEntity.getStatusCode().is2xxSuccessful()) {
                    numberOfExecutions.getAndIncrement();
                }
            });
        }

        executor.shutdown();

        assertThat(numberOfExecutions.get()).isEqualTo(100);
    }

}