[SPARK-28721] Failing to stop SparkSession in K8S cluster mode PySpark leaks Driver and Executors

• Type: Bug
• Status: Resolved
• Priority: Minor
• Resolution: Duplicate
• Affects Version/s: 2.4.1, 2.4.3
• Component/s: Kubernetes, PySpark, Spark Core
• Assignee: Unassigned
• Reporter: Patrick Clay
• Created: 13/Aug/19 23:30
• Updated: 17/May/20 18:25
• Resolved: 14/Aug/19 21:10

Description

This does not seem to affect 2.4.0.

To repro:

1. Download pristine Spark 2.4.3 binary
2. Edit pi.py to not call spark.stop()
3. ./bin/docker-image-tool.sh -r MY_IMAGE -t MY_TAG build push
4. spark-submit --master k8s://IP --deploy-mode cluster --conf spark.kubernetes.driver.pod.name=spark-driver --conf spark.kubernetes.container.image=MY_IMAGE:MY_TAG file:/opt/spark/examples/src/main/python/pi.py

The driver runs successfully and Python exits but the Driver and Executor JVMs and Pods remain up.

 

I realize that explicitly calling spark.stop() is always best practice, but since this does not repro in 2.4.0 it seems like a regression.

Issue Links

• duplicates: SPARK-27812 kubernetes client import non-daemon thread which block jvm exit.