Jeffrey Charles created SPARK-24786:
---------------------------------------
Summary: Executors not being released after all cached data is
unpersisted
Key: SPARK-24786
URL: https://issues.apache.org/jira/browse/SPARK-24786
Project: Spark
Issue Type: Bug
Components: Scheduler
Affects Versions: 2.2.1
Environment: Zeppelin in EMR
Reporter: Jeffrey Charles
I'm persisting a dataframe in Zeppelin which has dynamic allocation enabled to
get a sense of how much memory the dataframe takes up. After I note the size, I
unpersist the dataframe. For some reason, Yarn is not releasing the executors
that were added to Zeppelin. If I don't run the persist and unpersist steps,
the executors that were added are removed about a minute after the paragraphs
complete. Looking at the storage tab in the Spark UI for the Zeppelin job, I
don't see anything cached. I do not want to set
spark.dynamicAllocation.cachedExecutorIdleTimeout to a lower value because I do
not want executors with cached data to be released, but I do want ones that had
cached data and no longer have cached data to be released.
Steps to reproduce:
# Enable dynamic allocation
# Set spark.dynamicAllocation.executorIdleTimeout to 60s
# Set spark.dynamicAllocation.cachedExecutorIdleTimeout to infinity
# Load a dataset, persist it, run a count on the persisted dataset, unpersist
the persisted dataset
# Wait a couple minutes
Expected behaviour:
All executors will be released as the executors are no longer caching any data
Observed behaviour:
No executors were released
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
