Test: Inspect chunked responses during termination

[ennru]

Tests that a chunked HTTP response stream fails when the server terminates.

The Akka HTTP client reacts as it should with "EntityStreamException: Entity stream truncation. The HTTP parser was receiving an entity when the underlying connection was closed unexpectedly."

#1210 (comment) made me curious how a client will see server termination on a chunked HTTP response stream.

I expect this test to be flaky, so I'm not sure if it is worth merging.

References #1210

[akka-ci]

Test PASSed.

[jrudolph]

We are missing such a test indeed, mainly because we don't seem to verify yet that streams that are still ongoing are really aborted in time. What we should strive for is that the connection is aborted, otherwise, responses that use HttpEntity.CloseDelimited will see a truncated body without an error.

The test case seems a bit fragile since it uses the same ActorSystem for server and client. We should change that as well, because we might want to add Http().shutdownConnectionPools to the coordinated shutdown as well and that might interfere with the testing logic.

[akka-ci]

Test FAILed.

Pull request validation report

Failed Test Suites

Test result for 'akka-http-core / Pr-validation / ./ executeTests'

[info] ScalaTest
[info] Run completed in 4 minutes, 43 seconds.
[info] Total number of tests run: 1052
[info] Suites: completed 76, aborted 0  Scopes: pending 1
[info] Tests: succeeded 1048, failed 4, canceled 1, ignored 2, pending 55
[info] *** 4 TESTS FAILED ***
[error] Failed: Total 1052, Failed 4, Errors 0, Passed 1048, Ignored 2, Canceled 1, Pending 55
[error] Failed tests:
[error] 	akka.http.scaladsl.ClientServerSpec

[jrudolph]

Unrelated failure that will be fixed with #3145.

[jrudolph]

PLS BUILD

[akka-ci]

Test PASSed.

[raboof]

So this test shows the deadline isn't actually enforced for streams like this?

[raboof]

(I guess that might not be a big problem since the actorsystem shutdown will shut down the materializer and fail the stream anyway?)

[jrudolph]

I'm banging my head against this a bit... What I'm seeing is that the GracefulTerminationStage does something after those 50ms but that does not kill the connection immediately (especially not the response stream). My hypothesis is, since the termination stage does not monitor response entity streams, it does something which will only indirectly stop the connection at some point. In any case, it does not cancel connection but complete it, which will still keep proper semantics in this case (because chunked encoding can detect truncation) but not in the below one.

[ennru]

What we should strive for is that the connection is aborted, otherwise, responses that use HttpEntity.CloseDelimited will see a truncated body without an error.

Yes, I added a test for that and the client doesn't understand that it's over. The test is ignored for now.

[akka-ci]

Test PASSed.

[jrudolph]

LGTM, I'm fine with merging, I debugged the remaining issue a bit and will create a ticket to follow up.

[jrudolph]

I debugged the remaining issue a bit and will create a ticket to follow up.

#3209

[raboof]

Good to have this behavior covered with a test, indeed let's follow up on #3209 in a separate PR

[raboof]

Should we remove a couple of these as well?

[ennru]

Done.

[akka-ci]

Test PASSed.

[akka-ci]

Test PASSed.