Github user kayousterhout commented on the pull request:
https://github.com/apache/spark/pull/214#issuecomment-38496480
Echoing what @CodingCat said, I think this solution has the same problem
that I mentioned in response to your design posted in the JIRA
(https://spark-project.atlassian.net/browse/SPARK-1141). Often, when the
cluster is busy, resourceOffer() gets called with just one machine at a time
(because all of the machines are busy, and then a single task finished, so
resourceOffer() is called for the slot that the task just freed up). With your
solution, in that case, the serialization won't be asynchronous: one of the
task set managers will accept the resource offer, and then the DAGScheduler
will block waiting for the serialization to finish.
I think ideally what we'd like here is something similar to
TaskResultGetter, which is also used by TaskSchedulerImpl.
---
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.
---
