Github user cmccabe commented on the pull request:
https://github.com/apache/spark/pull/1486#issuecomment-54901700
Thanks for the review, @pwendell. I agree that delay scheduling is the way
to go, the question was whether to do it in this patch or in a follow-up.
Perhaps we can get it into this one...
I think part of the issue is that right now, RDD#getPreferredLocations
returns a list of strings, *not* a list of TaskLocation objects. I wanted to
avoid changing the return type of that function for compatibility reasons. We
could certainly have HadoopRDD#getPreferredLocations and
NewHadoopRDD#getPreferredLocations check which locations were cached in HDFS,
via SparkContext, and only return those. But this would have the same drawback
you're criticizing in the current solution-- i.e., the non-cached locations
would be ignored.
I think I can implement delay scheduling without too big a change, provided
that the "race condition" I mentioned earlier (race between resourceOffer and
scheduling) is not an issue. But I don't see a good way to avoid adding the
PartitionLocation.scala class-- let me know your thoughts on this.
@sryza: SplitInfoReflections is fine with me
---
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]
