Github user varuvish commented on the issue:
https://github.com/apache/spark/pull/19456
Can you guarantee that your users share an even workload? With this setup
you could potentially have a user with a high workload using 1/#users
resources. Wouldn't using the queue feature
Github user varuvish commented on the issue:
https://github.com/apache/spark/pull/21333
Nice change! I tested this out as well and verified that the shuffle
doesn't happen. I did notice that this change wasn't reflected in the dataset
API. Is that something that should be addressed