Github user pwendell commented on the pull request:
https://github.com/apache/spark/pull/1056#issuecomment-46652485
Hey @sryza, that status update exists in the `CoarseGrainedExecutorBackend`
interface mostly for historical reasons. The very first scheduler Spark had was
mesos, and mesos offered a way to send control plan messages (like "my task is
finished") through its own messaging system. So this control plan messaging in
Spark was coupled with the cluster scheduler.
These things don't really need to be coupled and longer term I'd actually
like to see all of the control plane messages go directly from the executor to
the driver to keep it simpler and avoid fragmentation between deploy modes.
For your patch, I'd circumvent this interface rather than extending it.
Instead, I'd just piggy back your status updates on top of the existing
block manager heartbeat that goes directly to the driver. We've already started
cramming other stuff in there because it's the one control-plan message that
gets sent from executors to the driver that is _not_ the task status message.
At least, that's the first attempt I'd make at this patch.
A better solution would be to add a general heartbeat actor on the
executor, similar to the one in the block manager, but a shared one that would
heartbeat on behalf of all the components. That might not be too bad either.
---
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.
---
