[
https://issues.apache.org/jira/browse/FLINK-9913?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16582193#comment-16582193
]
ASF GitHub Bot commented on FLINK-9913:
---------------------------------------
zhijiangW commented on issue #6417: [FLINK-9913][runtime] Improve output
serialization only once in RecordWriter
URL: https://github.com/apache/flink/pull/6417#issuecomment-413468551
@pnowojski , I think I got the reason of regression in some non-broadcast
cases.
The key point is when to call `RecordSerializer#prune()` which is used to
decrease the intermediate serialization buffer. In the past, the prune method
was only called after the target copying buffer is full and the record is full.
But now after emitting each record we will call to prune method in order to
narrow down the intermediate buffer ASAP. So the performance may be regression
in sensitive job scenarios. I will modify to keep the previous behavior of
calling prune in serializer.
Thanks for your benchmark to find this potential issue. I only verified the
broadcast scenarios in benchmark before and the obvious advantage may hide the
potential regression. :)
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
> Improve output serialization only once in RecordWriter
> ------------------------------------------------------
>
> Key: FLINK-9913
> URL: https://issues.apache.org/jira/browse/FLINK-9913
> Project: Flink
> Issue Type: Improvement
> Components: Network
> Affects Versions: 1.6.0
> Reporter: zhijiang
> Assignee: zhijiang
> Priority: Minor
> Labels: pull-request-available
> Fix For: 1.7.0
>
>
> Currently the {{RecordWriter}} emits output into multi channels via
> {{ChannelSelector}} or broadcasts output to all channels directly. Each
> channel has a separate {{RecordSerializer}} for serializing outputs, that
> means the output will be serialized as many times as the number of selected
> channels.
> As we know, data serialization is a high cost operation, so we can get good
> benefits by improving the serialization only once.
> I would suggest the following changes for realizing it.
> # Only one {{RecordSerializer}} is created in {{RecordWriter}} for all the
> channels.
> # The output is serialized into the intermediate data buffer only once for
> different channels.
> # The intermediate serialization results are copied into different
> {{BufferBuilder}}s for different channels.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
