Github user kayousterhout commented on the pull request:
https://github.com/apache/spark/pull/8887#issuecomment-147796048
I'm wondering if it's possible to do this in a simpler way and without
changing the Pool interface by:
(1) creating a new "ReasonPending" ExecutorLostReason
(2) In TaskSchedulerImpl, move the "rootPool.executorLost" call out of
removeExecutor and into executorLost, to a place where it gets called even if
the executorId is no longer in activeExecutorIds (even if the executor is no
longer considered active) (this means that the implementation of
Pool.executorLost needs to be idempotent because it may be called multiple
times, but the implementation in TaskSetManager seems to be)
(3) if an executor lost reason is ReasonPending, don't call
rootPool.executorLost(executorId, host) call yet (so the tasks don't get
cancelled)
(4) YARN can then call executorLost a second time once it knows the correct
reason
It seems like that might be a less invasive change that reuses most of the
existing code for dealing with losses, and also avoids making the scheduling
interfaces more complex? Thoughts?
---
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 infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
