[SPARK-23449] Extra java options lose order in Docker context

• Type: Bug
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 2.3.0
• Fix Version/s: 2.3.1, 2.4.0
• Component/s: Kubernetes, Spark Core

Environment

Running Spark on K8S with supplied Docker image. Passing along extra java options.

• Assignee: Andrew Korzhuev
• Reporter: Andrew Korzhuev
• Shepherd: Marcelo Masiero Vanzin
• Created: 16/Feb/18 14:21
• Updated: 17/May/20 18:23
• Resolved: 26/Feb/18 18:29

Description

`spark.driver.extraJavaOptions` and `spark.executor.extraJavaOptions` when processed in `entrypoint.sh` does not preserve its ordering, which makes `-XX:+UnlockExperimentalVMOptions` unusable, as you have to pass it before any other experimental options.

 

Steps to reproduce:

1. Set `spark.driver.extraJavaOptions`, e.g. `-XX:+UnlockExperimentalVMOptions -XX:+UseG1GC -XX:+CMSClassUnloadingEnabled -XX:+UseCGroupMemoryLimitForHeap`
2. Submit application to k8s cluster.
3. Fetch logs and observe that on each run order of options is different and when `-XX:+UnlockExperimentalVMOptions` is not the first startup will fail.

 

Expected behaviour:

1. Order of `extraJavaOptions` should be preserved.

 

Cause:

`entrypoint.sh` fetches environment options with `env`, which doesn't guarantee ordering.

env | grep SPARK_JAVA_OPT_ | sed 's/[^=]*=\(.*\)/\1/g' > /tmp/java_opts.txt

Issue Links

• links to: [Github] Pull Request #20628 (andrusha)