[ 
https://issues.apache.org/jira/browse/HIVE-29811?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18105208#comment-18105208
 ] 

László Bodor commented on HIVE-29811:
-------------------------------------

Attached the evaluations of my experiment with AppCDS on upstream hive 
docker-compose setup:  [^APPCDS-FINDINGS.pdf] 
https://github.com/abstractdog/hive/commit/fbb26d5040c8e18202b563dc6ba1a61d13fc9a8d

Even though the summary was generated by claude (having the full scope of the 
investigation session), the root of the reasoning is something I totally agree 
with, which is: "AppCDS on HS2 with Temurin 21 base CDS produces a measurable 
but operationally insignificant startup improvement."

There is still a chance that HS2 startup is slow, but it's not because of the 
classloading: further optimizing it isn't worth the effort.

> HS2 slow start - try AppCDS
> ---------------------------
>
>                 Key: HIVE-29811
>                 URL: https://issues.apache.org/jira/browse/HIVE-29811
>             Project: Hive
>          Issue Type: Improvement
>          Components: HiveServer2
>            Reporter: Attila Turoczy
>            Assignee: László Bodor
>            Priority: Major
>              Labels: cloud, modernization
>         Attachments: APPCDS-FINDINGS.pdf
>
>
> HiveServer2 is slow and cold-start time is dominated by JVM class loading of 
> the large *Hive/Hadoop/Tez/Calcite* class graph. Every process start 
> re-parses and verifies thousands of classes from jars. The base-JDK CDS 
> archive (enabled by \{{-Xshare:auto}}) covers only JDK classes. 
> Hive/Hadoop/Tez/Calcite application classes are NOT archived, so they are 
> re-loaded from jars every time.
> AI Recommendation:
> _AppCDS commonly reduces combined JVM + application startup by roughly 20–40% 
> for large class graphs by memory-mapping pre-parsed class metadata instead of 
> re-reading jars. The exact figure for HS2 should be confirmed with the 
> reproduction above on a reference image; the point of the ticket is to make 
> that win available and documented out of the box._



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to