JAVA_TOOL_OPTIONS environment expansion added in 0.112.0 crash loops java applications while running on gke

[ryanohnemus]

Component(s)

auto-instrumentation

What happened?

Description

After upgrading the otel-operator to version 0.112.0, java8 containers that use auto-instrumentation with inject-java attributes fail to start with the following error:

Picked up JAVA_TOOL_OPTIONS: $(JAVA_TOOL_OPTIONS) -javaagent:/otel-auto-instrumentation-java/javaagent.jar
Unrecognized option: $(JAVA_TOOL_OPTIONS)
Error: Could not create the Java Virtual Machine.
Error: A fatal exception has occurred. Program will exit.

I believe this is due to #1814 being patched in the 0.112.0 version.

Steps to Reproduce

Attempt to run a java8 application using autoinstrumentation after upgrading to otel-operator 0.112.0. The java container will crash loop with the above error.

Expected Result

Auto-instrumentation to load the javaagent without crashing the java application.

Actual Result

Java8 container crash loops.

Kubernetes Version

1.31.0

Operator version

0.112.0

Collector version

0.112.0

Environment information

Environment

Java Zulu JRE 8.0.432-1

Log output

Picked up JAVA_TOOL_OPTIONS: $(JAVA_TOOL_OPTIONS) -javaagent:/otel-auto-instrumentation-java/javaagent.jar
Unrecognized option: $(JAVA_TOOL_OPTIONS)
Error: Could not create the Java Virtual Machine.
Error: A fatal exception has occurred. Program will exit.

Additional context

Reverting to 0.107.0 of the otel-operator resolves the issue.

[pavolloffay]

hi @ryanohnemus this is not enough information to debug. Could you please share your deployment/pod specification?

[ryanohnemus]

hi @pavolloffay,

I ran through a few different versions of this setup and found this specifically is due to different handling of ENV vars within GKE vs normal k8s.

(As far as a deployment i was using the image: azul/zulu-openjdk:8-jre-latest with a command of ["java", "-version"] with autoinstrumentation enabled and an env var of JAVA_TOOL_OPTIONS=-Xdebug set. Java isn't the root issue in this case, it's actually the env variable handling differences in k8s version).

The change from #1814 appears to add a 2nd environment variable of the exact same name.
On a kind 1.31 cluster that results in the created pod existing with 2 env vars of the same name:

    - name: JAVA_TOOL_OPTIONS
      value: -Xdebug
    - name: JAVA_TOOL_OPTIONS
      value: $(JAVA_TOOL_OPTIONS) -javaagent:/otel-auto-instrumentation-java/javaagent.jar     

However when running on GKE's 1.31 version, it seems you are not allowed to have multiple env vars of the same name. These get collapsed when the pod is created and you have a pod that only has the last defined variable:

    - name: JAVA_TOOL_OPTIONS
      value: $(JAVA_TOOL_OPTIONS) -javaagent:/otel-auto-instrumentation-java/javaagent.jar     

This results in java attempting env expansion and failing with the Unrecognized option: $(JAVA_TOOL_OPTIONS) error and then exiting.

Can the original solution be updated to modify the JAVA_TOOL_OPTIONS env var in-place without this duplicate named entry being created?

[nickandreev]

Hi @pavolloffay @ryanohnemus ! Duplicate env vars don't work for us either (aws eks v1.28.13). Instead of two JAVA_TOOL_OPTIONS, there is only one JAVA_TOOL_OPTIONS in the pod resource with this value

    - name: JAVA_TOOL_OPTIONS
      value: $(JAVA_TOOL_OPTIONS) -javaagent:/otel-auto-instrumentation-java/javaagent.jar 

Which causes the pods to crash 😞

I think there is some kind of mutating webhook in our cluster that is causing such behaviour. I haven't identified which one, but I can investigate this later.