How can I automatically capture the heap dump on the driver and executors in the event of an OOM error?

[User16752245312]

If you have a job that repeatedly run into Out-of-memory error (OOM) either on the driver or executors, automatically capture the heap dump on OOM event will help debugging the memory issue and identify the cause of the error.

Spark config:

spark.executor.extraJavaOptions -XX:+HeapDumpOnOutOfMemoryError -XX:HeapDumpPath=/dbfs/cluster-logs/heap-dumps/
 
spark.driver.extraJavaOptions -XX:+HeapDumpOnOutOfMemoryError -XX:HeapDumpPath=/dbfs/cluster-logs/heap-dumps/

Environment variables:

DBFS_FUSE_VERSION=1

[John_360]

Is it necessary to use exactly that HeapDumpPath? I find I'm unable to get driver heap dumps with a different path but otherwise the same configuration. I'm using spark_version 10.4.x-cpu-ml-scala2.12.

[User16752245312]

It's not necessary has to be exact as in the example, you can use any path. But make sure the path points to a directory that already exists. Also note that the path is on DBFS. We want a location where both the driver and executors can write to.