Github user caneGuy commented on a diff in the pull request:
https://github.com/apache/spark/pull/18739#discussion_r129753610
--- Diff:
core/src/main/scala/org/apache/spark/scheduler/TaskSetManager.scala ---
@@ -665,10 +667,15 @@ private[spark] class TaskSetManager(
}
}
if (blacklistedEverywhere) {
- val partition = tasks(indexInTaskSet).partitionId
- abort(s"Aborting $taskSet because task $indexInTaskSet
(partition $partition) " +
- s"cannot run anywhere due to node and executor blacklist.
Blacklisting behavior " +
- s"can be configured via spark.blacklist.*.")
+ val dynamicAllocationEnabled =
conf.getBoolean("spark.dynamicAllocation.enabled", false)
+ val mayAllocateNewExecutor =
+ conf.getInt("spark.executor.instances", -1) >
currentExecutorNumber
+ if (!dynamicAllocationEnabled && !mayAllocateNewExecutor) {
--- End diff --
@squito yes,i agree with you.Since i can not think of any solution to
handle all cases well too.Fail-fast may be the right way.So i closed this pr.
What i want to discuss(may be confirm) is that should i report blacklist
nodes to yarn whenever it is blacklisted, but not wait for maybeFinishTaskSet
being called.Right now,yarn do not know some node which is blacklisted in
TaskSet because it must wait maybeFinishTaskSet being called to update.When the
node has more resource than other, yarn may launch task on the blacklisted node
when dynamic allocation is enabled.And job may fail by
abortIfCompletelyBlacklisted.If we report blacklist information
real-time,things may be improved a little?
Does this make sense?If do,i will open an other pr.Thanks for your time.
---
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]
