[
https://issues.apache.org/jira/browse/FLINK-15122?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16991584#comment-16991584
]
Piotr Nowojski edited comment on FLINK-15122 at 12/9/19 1:23 PM:
-----------------------------------------------------------------
Can you provide a benchmark results
([flink-benchmarks|https://github.com/dataArtisans/flink-benchmarks]), that
would motivate this change?
If we can motivate this change, then yes, there are two options:
# create a new flag
# reuse existing {{enableObjectReuse}} flag
In order to choose between those two, [~aljoscha] proposed an interesting
experiment: implement a PoC using {{ReusingDeserializationDelegate}} with the
existing {{ExecutionConfig.enableObjectReuse}} flag, enable it by default and
check how many and which tests are failing. I think that would be an important
data point. For example if all (almost all?) of the tests are passing, I would
vote for option 2.
was (Author: pnowojski):
Can you provide a benchmark
([flink-benchmarks|https://github.com/dataArtisans/flink-benchmarks]), that
would motivate this change?
If we can motivate this change, then yes, there are two options:
# create a new flag
# reuse existing {{enableObjectReuse}} flag
In order to choose between those two, [~aljoscha] proposed an interesting
experiment: implement a PoC using {{ReusingDeserializationDelegate}} with the
existing {{ExecutionConfig.enableObjectReuse}} flag, enable it by default and
check how many and which tests are failing. I think that would be an important
data point. For example if all (almost all?) of the tests are passing, I would
vote for option 2.
> Reusing record object in StreamTaskNetworkInput
> -----------------------------------------------
>
> Key: FLINK-15122
> URL: https://issues.apache.org/jira/browse/FLINK-15122
> Project: Flink
> Issue Type: Sub-task
> Components: Runtime / Network
> Reporter: Jingsong Lee
> Priority: Major
> Fix For: 1.11.0
>
>
> Now blink's batch is forced to open object reusing, but the data read from
> the network is not reused, which will lead to a large GC of the batch job.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
