Github user jerryshao commented on a diff in the pull request:
https://github.com/apache/spark/pull/17480#discussion_r111292239
--- Diff:
core/src/main/scala/org/apache/spark/ExecutorAllocationManager.scala ---
@@ -249,7 +249,14 @@ private[spark] class ExecutorAllocationManager(
* yarn-client mode when AM re-registers after a failure.
*/
def reset(): Unit = synchronized {
- initializing = true
+ /**
+ * When some tasks need to be scheduled and initial executor = 0,
resetting the initializing
+ * field may cause it to not be set to false in yarn.
+ * SPARK-20079: https://issues.apache.org/jira/browse/SPARK-20079
+ */
+ if (maxNumExecutorsNeeded() == 0) {
+ initializing = true
--- End diff --
For the 2nd point `an executor idle timeout has elapsed.`, in the `reset`
we also clear the executor idle timeout tracking info, which means executor
idle timeout will be recalculated based on newly registered executors. I think
this point could be true.
Yes we could set this flag `initializing` to false in reset, but I think
the scenario I mentioned above still should be concerned.
>the purpose of "initializing" is to avoid unnecessary executors ramp down
before the stage submitted or executor timeout. For example if min executor
number is 0, initial number is 10. If "initializing" is set to false, executor
number will ramp down to 0 immediately, and during this time if stage is
submitted, then still requires unnecessary executor ramp up to meet this
stage's requirement.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
