Github user koeninger commented on the pull request:
https://github.com/apache/spark/pull/10197#issuecomment-163302006
The reason direct stream has some latency is because it is figuring out, in
advance, on the driver, which offsets are in each partition. That means that
all relevant state is on the driver (so it can be checkpointed esailty), and
rdds are immutable once defined (so can be freely retried in the case of an
executor failure).
Even that latency should be minimized if you have a batch size that's tuned
down close to what your hardware can support. I've done sub-second batches.
Maybe I'm confused, but...
- Why is this largely just a cut and paste of existing code?
- How does this handle multiple receivers?
- How does this handle errors? You copied the error handling code from the
direct stream, but that assumes it's in a task that will be retried, not a
receiver. Looks like you're just silently catching all exceptions.
- Why does this still mention checkpoints, when it doesn't seem to have any
interaction with checkpoints?
---
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]
