[ 
https://issues.apache.org/jira/browse/HDDS-8580?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Zita Dombi resolved HDDS-8580.
------------------------------
    Resolution: Fixed

> Reduce memory usage in ContainerKeyMapperTask#reprocess
> -------------------------------------------------------
>
>                 Key: HDDS-8580
>                 URL: https://issues.apache.org/jira/browse/HDDS-8580
>             Project: Apache Ozone
>          Issue Type: Improvement
>          Components: Ozone Recon
>    Affects Versions: 1.3.0
>            Reporter: Ivan Andika
>            Assignee: Ivan Andika
>            Priority: Major
>              Labels: pull-request-available
>             Fix For: 1.4.0
>
>         Attachments: image-2023-05-11-11-46-01-929.png
>
>
> As OM performance test during Recon OM full snapshot, I removed the Recon DB 
> directory before restarting Recon to trigger full snapshot (essentially 
> bootstrapping a new Recon).
> However, it is found after OM DB is successfully downloaded, during the 
> {{reprocess}} of {{{}ContainerKeyMapperTask{}}}, the Recon heap usage 
> increased significantly for large keys table (our cluster has around 350 
> million keys).
> It is found that the issue was caused due to in-memory maps that store all 
> the OM keys during the reprocess. This is a regression introduced in 
> [HDDS-6783|https://jira.shopee.io/browse/HDDS-6783]. In essence, the patch is 
> to revert the implementation of 
> [HDDS-6783|https://jira.shopee.io/browse/HDDS-6783] only for 
> {{{}ContainerKeyMapperTask#reprocess{}}}.
>  * {{ContainerKeyMapperTask#process}} should not increase the heap memory 
> significantly since the number of delta updates are already limited by the 
> Recon configurations
>  * [HDDS-6783|https://jira.shopee.io/browse/HDDS-6783] aims for atomicity 
> during the Recon OM task updates. However since 
> {{ContainerKeyMapperTask#reprocess}} truncate all the Recon Container DB 
> before it starts and rebuilt the Recon Container DB, I think it's acceptable.
> After the patch is applied, the Recon heap size stays stable during the full 
> snapshot.
> Attached is Recon heap memory before and after the patch.
> !image-2023-05-11-11-46-01-929.png|width=716,height=546!



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

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to