Github user JoshRosen commented on the pull request:
https://github.com/apache/spark/pull/4066#issuecomment-72763703
There's one subtle, extremely-unlikely race condition that can still occur
here and which I would like to fix: currently, we seem to assume that messages
sent from the DAGScheduler to the local OutputCommitCoordinator will be
processed before commit requests for tasks. However, we send these as
fire-and-forget messages and do not wait for acknowledgements, so it would
technically be legal for OutputCommitCoordinator to receive a request to commit
for a task that belongs to a stage that OutputCommitCoordinator has not heard
about. This will cause the coordinator to deny the task attempt, since it will
think that it's from a completed stage. I think that this isn't a huge deal in
practice, since I think we'll just end up scheduling another task (plus this
particular race should be extremely unlikely).
Still, though, I'd like to consider whether there's a good way to fix this
in order to guarantee that requests to commit output for a stage occur after
the stage start event (from the OutputCommitter's POV).
---
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]
