Skip to content

Response stream flushed early for WebFlux applications wrapping a CompletableFuture #304

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
sapessi opened this issue Nov 14, 2019 · 3 comments · Fixed by #331
Closed

Response stream flushed early for WebFlux applications wrapping a CompletableFuture #304

sapessi opened this issue Nov 14, 2019 · 3 comments · Fixed by #331

Comments

@sapessi
Copy link
Contributor

sapessi commented Nov 14, 2019

  • Framework version: 1.4
  • Implementations: SpringBoot 2.x

Scenario

WebFlux method that returns a flux of multiple objects by wrapping a CompletableFuture. Issue originally reported by @cesardrk in #239

Expected behavior

Response output stream should remain open until the Flux is complete and has delivered all messages.

Actual behavior

When creating a flux from a CompletableFuture, Spring's ServletHttpHandlerAdapter calls flush right away, causing the framework to release the latch and return an empty response. Further, it seems that the adapter never calls the close method on the output stream making it pointless to move the latch release logic.

Steps to reproduce

This method will return and empty response right away:

    @RequestMapping(path = "/entity", method = RequestMethod.GET)
    ResponseEntity<Flux<String>> messageEntity() {
         return ResponseEntity.ok().body(Mono.fromFuture(CompletableFuture.supplyAsync(() -> {
                    try {
                        TimeUnit.SECONDS.sleep(5);
                    } catch (InterruptedException e) {
                        throw new IllegalStateException(e);
                    }
                    return MESSAGE;
                })
        ).flux());
    }

This method will succeed:

    @RequestMapping(path = "/asyncflux", method = RequestMethod.GET)
    Flux<String> futureFlux() {
        return Flux.create((sink) -> {
            System.out.println("Supply async");
            try {
                TimeUnit.SECONDS.sleep(5);
            } catch (InterruptedException e) {
                throw new IllegalStateException(e);
            }
            sink.next(MESSAGE);
        });
    }
@sapessi sapessi added this to the Release 1.5 milestone Nov 14, 2019
@cesardrk
Copy link

cesardrk commented Nov 14, 2019

@sapessi could it be something with the timing that the response.flushBuffer() is called in this case? While debugging i was tinkering and skipped the flushBuffer() call in the AwsLambdaServletContainerHandler.doFilter method and in the end i got the expected result just fine.

@sapessi
Copy link
Contributor Author

sapessi commented Nov 14, 2019

@cesardrk That seems to be the behavior according to my tests. When a Flux or Mono is initialized with a CompletableFuture they call flush right away on the output stream. They may write and call flush again after the future returns but we never get to find out because the framework, at the moment, releases the latch for the response at the first flush call.

I have tried to move the latch release logic to the close method of the output stream but it looks like they never call close. I'll have to dig deeper. I'll keep this issue updated as I discover exactly what is happening - the framework code changes at runtime rending debuggers pretty useless.

@sapessi
Copy link
Contributor Author

sapessi commented Mar 31, 2020

@cesardrk that was indeed the issue. I'm pushing a fix tonight to the v1.5 branch if you want to test. This was also the root cause of #306 and #279

sapessi added a commit that referenced this issue Mar 31, 2020
…uest asynchronously. This was causing race conditions in the SpringBoot 2 WebFlux implementation - requests that had to run through security or validation filters took longer and the library flushed an empty request, which caused the status code to default to 200. This fix addresses issues #279, #304, and #306
@sapessi sapessi linked a pull request Apr 7, 2020 that will close this issue
sapessi added a commit that referenced this issue Apr 8, 2020
* Bump spring.version in /aws-serverless-java-container-spring (#319)

Bumps `spring.version` from 5.1.9.RELEASE to 5.2.3.RELEASE.

Updates `spring-webmvc` from 5.1.9.RELEASE to 5.2.3.RELEASE
- [Release notes](https://github.com/spring-projects/spring-framework/releases)
- [Commits](spring-projects/spring-framework@v5.1.9.RELEASE...v5.2.3.RELEASE)

Updates `spring-test` from 5.1.9.RELEASE to 5.2.3.RELEASE
- [Release notes](https://github.com/spring-projects/spring-framework/releases)
- [Commits](spring-projects/spring-framework@v5.1.9.RELEASE...v5.2.3.RELEASE)

Signed-off-by: dependabot[bot] <support@github.com>

Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* Bump spring-webflux in /aws-serverless-java-container-springboot2 (#318)

Bumps [spring-webflux](https://github.com/spring-projects/spring-framework) from 5.1.9.RELEASE to 5.2.0.RELEASE.
- [Release notes](https://github.com/spring-projects/spring-framework/releases)
- [Commits](spring-projects/spring-framework@v5.1.9.RELEASE...v5.2.0.RELEASE)

Signed-off-by: dependabot[bot] <support@github.com>

Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* ci: Fixing Spring build to use 5.2 as latest

* chore(deps): Bump Spring 5.1 path release to address a security vulnerability

* chore(deps): Fixing usual spring dependency mess with exlusions out of the spring-security package used in the tests

* Fix for issue #317 (#323)

* fix issue 317 - use charset from request

* update dependencies

* update build dependencies, remove spring boot 2.0.x

* restoring ci config

Co-authored-by: Stefano Buliani <2996317+sapessi@users.noreply.github.com>

* test: Fixed Spring security tests for SpringBoot 2, added validation tests and updated servlet tests to use the new servletApplication option

* fix: Avoid flushing the response buffer if we are dispatching the request asynchronously. This was causing race conditions in the SpringBoot 2 WebFlux implementation - requests that had to run through security or validation filters took longer and the library flushed an empty request, which caused the status code to default to 200. This fix addresses issues #279, #304, and #306

* chore(deps): Bump spring dependency version and added webmvc optional dependency to truly support Servlet-only server

* feat: New application type parameter to SpringBootLambdaContainerHandler that tells the framework whether to start a reactive or servlet-based embedded server. Also added a new servletApplication method to the builder object.

* test: Fixed UTF-8 encoding test

* ci: Fixed dependencies for CI run on SpringBoot 2

* ci: More Spring dependency convergence issues during CI

* fix: Added null-check on getServerName in case the multi-value headers property is null. Unlikely outside of tests but better safe than sorry. This addresses #327

* fix: Changed servlet initialization mechanism so that servlet that requests load on startup are initialized right away, as part of the initialization() method call in LambdaServletContainerHandler. Also centralized the lazy Servlet initialization to the ServletExecutionFilter so that we don't have code scattered all around. This begins to address #287

* feat: Added new 0-parameter constructor for async initializer that uses the actual JVM start time to calculate the timeout milliseconds. Also added the new method to the builder object and deprecated the current method that receives a milliseconds epoch parameter. I'm not deprecating the constructor of the async initializer class that receives the parameter as it may still be useful for tests. This change was suggested in #287

* fix: Updated SpringBoot 1.x handler to use the new servlet initialization mechanism

* ci: switch SpringBoot slow integration test to use a custom async time since the JVM is reused for both tests in the and we cannot reuse the actual JVM init time

* feat: New models for HTTP API support for #329

* feat: First implementation of HTTP API servlet request, request reader, and security context writer - continuing to address #329

* test: Basic unit tests for the new HTTP API support in core library (#329)

* feat: Updated log formatter to support both versions (1 and 2) of the proxy request model (#329)

* feat: Further generified request readers to read to a generic HttpServletRequest rather than specific implementations of it. This makes it easier to create container handler implementations that support HTTP API, API Gateway, and ALB (#329)

* test: Fixed tests for new logged and generified request readers

* feat: Added HTTP API support to Jersey implementation with new getHttpApiV2ProxyHandler method (#329)

* feat: Added HTTP API support to Spark implementation (#329)

* feat: Added HTTP API support to Spring implementation (#329)

* feat: HTTP API support in SpringBoot 2 implementation. bug: Fixed an issue with the implementation of AsyncContext where it wasn't dispatching if the handler wasn't set

* feat: First pass of HTTP API support in struts 2 implementation (#329)

* fix: Added support for HTTP APIs to the request dispatcher

* chore(deps): Dependency bump all around. Rotated Jersey ci versions

* fix: Updated stream handling logic to work with reactive applications as suggested in #316

* test: Added unit test to replicate #333

* feat: New configuration parameter to skip exception mapping and allow exception to bubble up from #307

* fix: Fixed spotbugs issue in RuntimeException cast

* test: Added tests for more complex content types mentioned in issue #315

* docs: Updated samples to support SAM CLI operations out of the box to address #293 and switched to HTTP API by default

* feat: Updated archetypes to work out of the box with the SAM CLI, continuing to address #293

* chore: License header pass on the entire project

* fix: Set default value for setDisableException mapper in config to false

* fix: Updated default initialization timeout to 20 seconds

Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Eran Medan <eranmeda@amazon.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging a pull request may close this issue.

2 participants