pnowojski commented on a change in pull request #8409: [FLINK-12478] Decompose
monolithic run-loops in StreamTask implementa…
URL: https://github.com/apache/flink/pull/8409#discussion_r285551963
##########
File path:
flink-streaming-java/src/main/java/org/apache/flink/streaming/runtime/tasks/OneInputStreamTask.java
##########
@@ -98,13 +96,8 @@ public void init() throws Exception {
}
@Override
- protected void run() throws Exception {
- // cache processor reference on the stack, to make the code
more JIT friendly
- final StreamInputProcessor<IN> inputProcessor =
this.inputProcessor;
-
- while (running && inputProcessor.processInput()) {
- // all the work happens in the "processInput" method
- }
+ protected boolean performDefaultAction() throws Exception {
Review comment:
As I discussed offline with both of you, I think we should mostly focus on
the higher level benchmarks (as implemented by Stefan). Lower level benchmarks
might be valuable if there are various of special cases that we want to test
and we already know that the thing they are covering can be/is a bottleneck
(network, state backend accesses).
If we can not measure the performance regression in those higher level
benchmarks, I would say it will be impossible for it to be visible by a final
user on even higher level (cluster). Lower level benchmarks can either measure
non visible change or measure a change, that is not even there (JVM might not
be able to apply the same optimisation with larger context, or vice versa, when
running full code it might be able to optimise something that it wasn't able to
optimise during unit style benchmarks).
Regrading the potential improvement here. @StefanRRichter please open a pull
request with those two newly added benchmarks. I would like first to measure
couple of results on the master branch for those benchmarks to get a good
baseline, before merging this PR.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
