[
https://issues.apache.org/jira/browse/FLINK-34148?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17809874#comment-17809874
]
Matthias Pohl commented on FLINK-34148:
---------------------------------------
Thanks for looking into it, [~Sergey Nuyanzin]. It sounds it's the same issue
that was fixed for Flink's main repo in FLINK-28016 where we also introduced
the {{<flink.markBundledAsOptional/>}} to cover this topic with the upgrade of
the {{maven-shade-plugin}}. I'm wondering whether we should make the shade
plugin being done for flink-shaded, entirely, instead of having a dedicated
{{maven-shade-plugin}} upgrade for the jackson dependencies. Or was there a
reason why you only wanted to apply this plugin upgrade to the jackson
dependencies?
If I understand FLINK-28016 correctly, running the release with Maven 3.2.5
(i.e. something lower than 3.3) might also work around this issue.
Keep in mind that it might be useful to also add the {{ShadeOptionalChecker}}
(see
[.tools/ci/verify_bundled_optional.sh|https://github.com/apache/flink/blob/5bf5003f5c7baf19b0164f78558e495d8bb62b04/tools/ci/verify_bundled_optional.sh])
to {{flink-shaded}} analogously to what we're doing with the
{{LicenseChecker}} if we implement the {{<flink.markBundledAsOptional/>}}
approach. Anything to add from your side, [~chesnay]?
So, from my side, trying to fix the issue in flink-shaded and doing another
release 18.1 sounds reasonable. We can unblock the Apache Flink 1.19 release if
we want to by just reverting to flink-shaded 17.0. The JDK21 issue isn't
necessarily a blocking issue because JDK21 is still meant to be a beta feature.
> Potential regression (Jan. 13): stringWrite with Java8
> ------------------------------------------------------
>
> Key: FLINK-34148
> URL: https://issues.apache.org/jira/browse/FLINK-34148
> Project: Flink
> Issue Type: Improvement
> Components: API / Type Serialization System
> Reporter: Zakelly Lan
> Priority: Blocker
> Fix For: 1.19.0
>
>
> Significant drop of performance in stringWrite with Java8 from commit
> [881062f352|https://github.com/apache/flink/commit/881062f352f8bf8c21ab7cbea95e111fd82fdf20]
> to
> [5d9d8748b6|https://github.com/apache/flink/commit/5d9d8748b64ff1a75964a5cd2857ab5061312b51]
> . It only involves strings not so long (128 or 4).
> stringWrite.128.ascii(Java8) baseline=1089.107756 current_value=754.52452
> stringWrite.128.chinese(Java8) baseline=504.244575 current_value=295.358989
> stringWrite.128.russian(Java8) baseline=655.582639 current_value=421.030188
> stringWrite.4.chinese(Java8) baseline=9598.791964 current_value=6627.929927
> stringWrite.4.russian(Java8) baseline=11070.666415 current_value=8289.95767
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
