Unrecognized VM option 'UseConcMarkSweepGC' on OpenJDK 15

[polyfractal]

Node failed to boot because UseConcMarkSweepGC isn't supported, presumably because it was removed in OpenJDK 15?

https://gradle-enterprise.elastic.co/s/kloq36r7j4e3w

ES_BUILD_JAVA=openjdk13,ES_RUNTIME_JAVA=openjdk15,nodes=general-purpose

OS | openSUSE Leap 15.1, Linux 4.12.14-lp151.28.36-default

Specs | 32 CPUs, 94.4GB RAM
java.version | 13.0.2
java.vm.name | OpenJDK 64-Bit Server VM
java.vm.version | 13.0.2+8
java.runtime.version | 13.0.2+8
java.home | /var/lib/jenkins/.java/openjdk-13.0.2-linux

12:31:34 » Exception in thread "main" java.lang.RuntimeException: starting java failed with [1]
12:31:34 »  output:
12:31:34 »  
12:31:34 »  error:
12:31:34 »  Unrecognized VM option 'UseConcMarkSweepGC'
12:31:34 »  Error: Could not create the Java Virtual Machine.
12:31:34 »  Error: A fatal exception has occurred. Program will exit.
12:31:34 »  	at org.elasticsearch.tools.launchers.JvmErgonomics.flagsFinal(JvmErgonomics.java:118)
12:31:34 »  	at org.elasticsearch.tools.launchers.JvmErgonomics.finalJvmOptions(JvmErgonomics.java:86)
12:31:34 »  	at org.elasticsearch.tools.launchers.JvmErgonomics.choose(JvmErgonomics.java:59)
12:31:34 »  	at org.elasticsearch.tools.launchers.JvmOptionsParser.main(JvmOptionsParser.java:92)

[elasticmachine]

Pinging @elastic/es-core-infra (:Core/Infra/Build)

[polyfractal]

Probably related to: #51335 :)

[jasontedor]

This is because:

• JDK 15 removed support for the CMS-related flags
• this is a build with a JDK 15 runtime Java
• we are trying to start the 7.5 BWC cluster using the same runtime Java (JDK 15)
• not using CMS-related flags (Restrict support for CMS to pre-JDK 14 #49123) was not backported to 7.5 since there was no expectation of supporting JDK 14 nor JDK 15 on 7.5

We use to have it in the build that when we started the BWC cluster, we would use a version of the JDK that the version that we were starting was known to be compatible with. I think that when starting older clusters, we should use the bundled JDK overriding the runtime JDK. That's what we need here, and in general, so that we are only starting clusters against versions of the JDK that the version being started supports. Can you investigate making this change @mark-vieira?

[mark-vieira]

I think that when starting older clusters, we should use the bundled JDK overriding the runtime JDK.

Agreed. In fact we have an open issue to use the configured bundled JDK version for all integration testing (except Java matrix tests).

That's what we need here, and in general, so that we are only starting clusters against versions of the JDK that the version being started supports.

We don't have a notion of maximum runtime Java version defined in an easily consumable way that I'm aware of. In any case, when we fetch BWC versions from a archive distribution we don't have any kind of metadata here either. I suppose we can hardcode this, which is probably fine given we only have to go back 1 major version and moving forward we always rely on using the bundled JDK. I'll work up a solution here.

[jasontedor]

That brings us to #43071 so it’s only that we never implemented this for testclusters.

We don't have a notion of maximum runtime Java version defined in an easily consumable way that I'm aware of.

We use to have this, but hard coded into the build as you suggest. It was never pretty to maintain. At some point we moved to using some content out of .ci (since we know the versions we test against there) but I’m having trouble tracking that down on mobile.

[mark-vieira]

Yeah, that PR you mentioned got blown away with the migration to test clusters and it seems we reintroduced the "just use whatever runtime java is" logic. I'm working up a PR for this. I'm not too worried about maintenance. We only have to consider all current active BWC versions < 7.2, which we know to work on Java 13. I don't suspect we'll ever have to change this as going forward we won't introduce any new versions that don't also include a bundled JDK.

[droberts195]

As well as the problem of 7.x being BWC tested against older versions that don't support Java 15, there is also the problem that we appear to be using Java 15 in CI builds for branches older than 7.x. For example, https://gradle-enterprise.elastic.co/s/mp33jqizng4l2 is a 7.5 branch build that tried to use Java 15 and failed because UseConcMarkSweepGC wasn't recognised.

[mark-vieira]

Good catch. Yes, as per Jason's comment we should not be testing 7.5 on Java 14 or 15. I've removed that with 24f0267.