StephanEwen commented on pull request #13920:
URL: https://github.com/apache/flink/pull/13920#issuecomment-723639969
## On Volatile Reads and Performance Considerations
I think `volatile` read assumption is completely correct. There is a common
interpretation that this just means
StephanEwen commented on pull request #13920:
URL: https://github.com/apache/flink/pull/13920#issuecomment-723222632
A thought on performance: There are some metrics that use `volatile`
variables.
So far we always avoided that. Even "infrequently accessed" fields, like
watermarks,
StephanEwen commented on pull request #13920:
URL: https://github.com/apache/flink/pull/13920#issuecomment-723218943
The PR adds a new `numBytesIn` metric to the `OperatorIOMetricGroup`. That
means all operators now show that metric. But the metric is zero for the
majority of operators
StephanEwen commented on pull request #13920:
URL: https://github.com/apache/flink/pull/13920#issuecomment-723208144
@zentol Curious what your take is here.
This is an automated message from the Apache Git Service.
To
StephanEwen commented on pull request #13920:
URL: https://github.com/apache/flink/pull/13920#issuecomment-723207966
Thanks for implementing this.
At a first glance, I think moving the `MetricNames` to `flink-core` does not
work too well. The metric names don't have a meaning by
StephanEwen commented on pull request #13920:
URL: https://github.com/apache/flink/pull/13920#issuecomment-723196124
Taking a look now.
@becketqin The code currently fails with `NullPointerException` in some
tests.
StephanEwen commented on pull request #13920:
URL: https://github.com/apache/flink/pull/13920#issuecomment-723189562
Sorry to be late to the game here, but could you share a bit more
information on what the original setup was?
Specifically, what was your checkpoint storage system that
