Application Context initialized twice during test when exception thrown during initialization

[QwertGold]

I noticed that when I have a Spring Boot test, and the context fails to initialize, the Banner is printed twice, and I decided to look into why.

// spring-boot-test-autoconfigure 2.3.7.RELEASE
public class SpringBootDependencyInjectionTestExecutionListener extends DependencyInjectionTestExecutionListener {

	@Override
	public void prepareTestInstance(TestContext testContext) throws Exception {
		try {
			super.prepareTestInstance(testContext);
		}
		catch (Exception ex) {
			outputConditionEvaluationReport(testContext);
			throw ex;
		}
	}
.....
}

When an exception is thrown during initialization, outputConditionEvaluationReport(testContext) is called, which eventually leads to the context being initialized a second time inside DefaultCacheAwareContextLoaderDelegate::loadContext.

This is more a nuisance than a bug, as it only happens in test code; the cost is 'just' developer time and build resources.

I this particular case I want to fail fast, so I throw an exception because I know bad things will happen later. I'm trying to safeguard against less experienced developers, using resources outside the intended lifecycle. During tests we automatically create external resources in an early lifecycle phase, these resources may only be accessed during later phases, and after the application is started. If a developer violate this rule, and tries to access a resource during bean construction or in an earlier lifecycle, I can detect it, and throw an IlligealStateException. If I don't do this I will get an exception later from the resource Api, and it can be hard for junior developers to identify that they have violated the resource lifecycle, if they get a 404/500 error from a HttpClient.

One possible solution, is to have an AbortTestContextInitializationException, that you could throw, if you decide that the TestContext is in a state where it does not make sense to continue executing, and there is no point in trying to generate the ConditionEvaluationReport. Ideally the failure of the initialization would be cached, so other tests using the same context would fail immediately instead of using build resources , trying to create the context twice for every test.

[philwebb]

@QwertGold I think I understand the description, but it would be useful if you could provide a sample application that shows the problem.

[QwertGold]

Cool, let me build a small project to illustrate

[QwertGold]

I'm unable to reproduce this in a simple stand alone project. When I debug it I can see some differences in the call stack depth at the point where the ApplicationContext is created, so I will need some time to figure out where why my larger project behaves differently, maybe I'll learn something along the way - I usually do ;)

[QwertGold]

I figure out how to reproduce this, it seems to happen when the test has a real Web server environment @SpringBootTest(webEnvironment = SpringBootTest.WebEnvironment.RANDOM_PORT)

I started with JUnit 4 and boot 2.3.7 as this is what we use, but the behavior is the same for Junit5 and boot 2.4.2, the repo has commits for both cases, https://github.com/QwertGold/spring-24888

[rokii]

hit same issue , any update ?

I suggest adding below in the catch block{}

catch (Exception ex) {
     if(!(ex.getCause() instanceof UnsatisfiedDependencyException)){
       outputConditionEvaluationReport(testContext);
     }
     throw ex;
}

Or

catch (Exception ex) {
     if(testContext.hasApplicationContext()){
       outputConditionEvaluationReport(testContext);
     }
     throw ex;
}

[wilkinsona]

Thanks for the sample and for your patience while we found time to start looking at this, @QwertGold.

The problem doesn't occur with a mock web environment (plain @SpringBootTest) because the context refresh is triggered earlier (by ServletTestExecutionListener) which we don't replace so no attempt is made to output the condition evaluation report.

When there's a full-blown web environment, ServletTestExecutionListener does not run because the org.springframework.test.context.web.ServletTestExecutionListener.activateListener attribute in the test context has been set to false. This allows preparation of the test instance to proceed and to reach SpringBootDependencyInjectionTestExecutionListener which triggers the double initialization when trying to output the condition evaluation report after the refresh failure.

[wilkinsona]

The behaviour's the same back in 1.4.0-M2 when the printing of the condition evaluation report was first introduced (see https://github.com/spring-projects/spring-boot/issues/4901 and https://github.com/spring-projects/spring-boot/commit/e5f224118b7683faa14cc32b18cd3cbdae7664d7) and in 1.4.1 when it was refined (see https://github.com/spring-projects/spring-boot/issues/6874 and https://github.com/spring-projects/spring-boot/commit/7134586310a378557113e08090b18bcfc399dd0a).

When refresh fails the context won't be available so I can't see how the current approach will be able to access the context to generate the report. We could just stop using SpringBootDependencyInjectionTestExecutionListener as it doesn't appear to work as hoped or I think we need to rework the approach in a way that means we don't need to get the application context from the test context to trigger the generation of the report.

Flagging for an upcoming team meeting so that we can discuss what to do.

[mirceade]

What a nuisance, spent an hour debugging this...

[mikelhamer]

Just pulled my hair out trying to figure out why my context was loading twice as I was trying to figure out what was causing it to have an error during initialization in the first place.....not fun!

[sbrannen]

We could just stop using SpringBootDependencyInjectionTestExecutionListener as it doesn't appear to work as hoped or I think we need to rework the approach in a way that means we don't need to get the application context from the test context to trigger the generation of the report.

I'm favor of getting rid of SpringBootDependencyInjectionTestExecutionListener and replacing it with a dedicated mechanism.

What do you think about introducing an SPI (in Spring Framework 6.0) in the TestContext framework for "processing" ApplicationContext load failures -- basically a new interface that Boot could implement and register (potentially via the spring.factories mechanism)?

Related Issues:

• https://github.com/spring-projects/spring-framework/issues/14182

[wilkinsona]

That sounds great, Sam. Thanks. We'll be left with the problem described in this issue in 2.x, but with a much better path forward in 3.0.

[sbrannen]

That sounds great, Sam. Thanks. We'll be left with the problem described in this issue in 2.x, but with a much better path forward in 3.0.

• See https://github.com/spring-projects/spring-framework/issues/28826