Karthik Palaniappan created SPARK-21960:
-------------------------------------------
Summary: Spark Streaming Dynamic Allocation should respect
spark.executor.instances
Key: SPARK-21960
URL: https://issues.apache.org/jira/browse/SPARK-21960
Project: Spark
Issue Type: Improvement
Components: DStreams
Affects Versions: 2.2.0
Reporter: Karthik Palaniappan
Priority: Minor
This check enforces that spark.executor.instances (aka --num-executors) is
either unset or explicitly set to 0.
https://github.com/apache/spark/blob/v2.2.0/streaming/src/main/scala/org/apache/spark/streaming/scheduler/ExecutorAllocationManager.scala#L207
If spark.executor.instances is unset, the check is fine, and the property
defaults to 2. Spark requests the cluster manager for 2 executors to start
with, then adds/removes executors appropriately.
However, if you explicitly set it to 0, the check also succeeds, but Spark
never asks the cluster manager for any executors. When running on YARN, I
repeatedly saw:
{code:java}
17/08/22 19:35:21 WARN org.apache.spark.scheduler.cluster.YarnScheduler:
Initial job has not accepted any resources; check your cluster UI to ensure
that workers are registered and have sufficient resources
17/08/22 19:35:36 WARN org.apache.spark.scheduler.cluster.YarnScheduler:
Initial job has not accepted any resources; check your cluster UI to ensure
that workers are registered and have sufficient resources
17/08/22 19:35:51 WARN org.apache.spark.scheduler.cluster.YarnScheduler:
Initial job has not accepted any resources; check your cluster UI to ensure
that workers are registered and have sufficient resources
{code}
I noticed that at least Google Dataproc and Ambari explicitly set
spark.executor.instances to a positive number, meaning that to use dynamic
allocation, you would have to edit spark-defaults.conf to remove the property.
That's obnoxious.
In addition, in Spark 2.3, spark-submit will refuse to accept "0" as a value
for --num-executors or --conf spark.executor.instances:
https://github.com/apache/spark/commit/0fd84b05dc9ac3de240791e2d4200d8bdffbb01a#diff-63a5d817d2d45ae24de577f6a1bd80f9
It is much more reasonable for Streaming DRA to use spark.executor.instances,
just like Core DRA. I'll open a pull request to remove the check if there are
no objections.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
