Github user squito commented on the pull request:
https://github.com/apache/spark/pull/9214#issuecomment-155143926
@JoshRosen @rxin @mateiz so, what is the path forward here? Josh is
correct that this doesn't really correctly handle the case of missing files.
But meanwhile, spark is still not fault tolerant. The way I see it, our
options are:
1) Prevent executors from re-registering, as Josh has suggested above.
Almost certainly out of scope for 1.6, and I am skeptical that it would even
make 1.7.
2) use this approach despite the flaw when files go missing. It may work
even if a file disappears, or it may not. But in other cases, it is still a
huge improvement.
3) Don't even try to handle the case of missing shuffle files, the same way
we won't handle corrupted shuffle files. So the ShuffleOutputCoordinator logic
would change so that it would not commit outputs if *any* output file already
existed. This would be a behavior change, specifically a reversal of
SPARK-4085, but would be a consistent model of fault-tolerance.
4) Use one shuffle output per attempt.
https://github.com/apache/spark/pull/6648
---
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]
