Improve performance of ReactiveAdapterRegistry [SPR-15747]

[spring-projects-issues]

Phil Webb opened SPR-15747 and commented

Whilst profiling startup times for Spring Boot I noticed that ReactiveAdapterRegistry is taking quite some time for a simple application.

For example, if I run ConfigurationClassProcessingTests.simplestPossibleConfig I get a test run time around 0.66 seconds. If I comment out the code from the ReactiveAdapterRegistry constructor it drops to 0.53 seconds.

Perhaps we can make the registry late binding, or unwind some of the existing lambda calls?

---

Affects: 5.0 RC2

Referenced from: commits dfcc9af

[spring-projects-issues]

Rossen Stoyanchev commented

I can't reproduce this reliably as described.

The following gives me the same results (between 1200-1350, sometimes down to 1100) whether the ReactiveAdapterRegistry constructor is commented out or not:

@Test
public void simplestPossibleConfig() {
     StopWatch watch = new StopWatch();
     watch.start();
     BeanFactory factory = initBeanFactory(SimplestPossibleConfig.class);
     String stringBean = factory.getBean("stringBean", String.class);
     watch.stop();
     System.out.println(watch.prettyPrint());
}

Narrowing down to just the constructor, I get about 100 milliseconds in either scenario:

@Test
public void simplestPossibleConfig() {
     StopWatch watch = new StopWatch();
     watch.start();
     new ReactiveAdapterRegistry();
     watch.stop();
     System.out.println(watch.prettyPrint());
}

I'm pretty sure this all comes from the classpath detection in the static variables. The following also gives 100 milliseconds:

@Test
public void simplestPossibleConfig() {
     StopWatch watch = new StopWatch();
     watch.start();
     boolean reactorPresent = ClassUtils.isPresent("reactor.core.publisher.Flux", ReactiveAdapterRegistry.class.getClassLoader());
     boolean rxJava1Present = ClassUtils.isPresent("rx.Observable", ReactiveAdapterRegistry.class.getClassLoader());
     boolean rxReactiveStreamsPresent = ClassUtils.isPresent("rx.RxReactiveStreams", ReactiveAdapterRegistry.class.getClassLoader());
     boolean rxJava2Present = ClassUtils.isPresent("io.reactivex.Flowable", ReactiveAdapterRegistry.class.getClassLoader());
     watch.stop();
     System.out.println(watch.prettyPrint());
}

[spring-projects-issues]

Rossen Stoyanchev commented

I've removed the static classpath checks in favor of lazily trying to register + ignore failures. Can you confirm if you see a difference?

[spring-projects-issues]

Ivan Pavlukhin commented

Hi, I am observing performance penalty on the path:

registerAdapter, ReactiveAdapterRegistry$ReactorJdkFlowAdapterRegistrar {org.springframework.core}
<init>, ReactiveAdapterRegistry {org.springframework.core}
<init>, InvocableHandlerMethod {org.springframework.web.reactive.result.method}
getRequestMappingMethod, ControllerMethodResolver {org.springframework.web.reactive.result.method.annotation}
handle, RequestMappingHandlerAdapter {org.springframework.web.reactive.result.method.annotation}

It is called for each incoming request and I observe a penalty around 2 ms for each request. Most likely it is caused by used reflection. Is there a need to create ReactiveAdapterRegistry for each request?

[spring-projects-issues]

Rossen Stoyanchev commented

What you're describing used to be an issue that was fixed in 5.0.2. Based onthe stack trace I wouldn't expect this to be an issue currently so I'm guessing you're on 5.0.1. In any this is an old ticket that will not re-open, so please create a new ticket if necessary.

[spring-projects-issues]

Ivan Pavlukhin commented

Ah, indeed I used 5.0.1, sorry for that.