[
https://issues.apache.org/jira/browse/KUDU-3516?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17776721#comment-17776721
]
ASF subversion and git services commented on KUDU-3516:
-------------------------------------------------------
Commit 3eb745f25dfa1975ab1aeba8d769697504f7fa54 in kudu's branch
refs/heads/master from 宋家成
[ https://gitbox.apache.org/repos/asf?p=kudu.git;h=3eb745f25 ]
KUDU-3516 No need to traverse all delta stores while computing score.
If we have many tables with many columns and each of them gets many
update requests, the maintenance scheduler thread might be stuck in
calculating the perf improvement scores of major delta compactions.
This is because we check all the updated columns while one single
updated column is actully enough.
This patch adds a new method to judge if a delta store needs to be
compacted. The scheduler will return as soon as it finds an updated
column. Please see KUDU-3516 for details.
Change-Id: I4b2c9525d8a12183130ee2cb7b5b00d1d2f60bfa
Reviewed-on: http://gerrit.cloudera.org:8080/20547
Tested-by: Kudu Jenkins
Reviewed-by: Alexey Serbin <ale...@apache.org>
> Tserver: Maintenance scheduler might be stuck in
> DeltaStats#AddColumnIdsWithUpdates
> ------------------------------------------------------------------------------------
>
> Key: KUDU-3516
> URL: https://issues.apache.org/jira/browse/KUDU-3516
> Project: Kudu
> Issue Type: Bug
> Components: tserver
> Reporter: Song Jiacheng
> Priority: Major
> Attachments: image-2023-10-09-15-58-47-267.png,
> image-2023-10-09-15-59-01-026.png
>
>
> If we have many tables with many columns and each of them get many update
> requests, the maintenance scheduler will be stuck in calculating the perf
> improvement score of major compaction.
> This tablet server has 6 maintenance manager but could only schedule 1 or 2
> tasks at one time, even if the tablet server is actually under high memory
> pressure.
> !image-2023-10-09-15-59-01-026.png|width=655,height=327!
> According to the stack showed below, I found out the scheduler was stuck in
> AddColumnIdsWithUpdates for a long time, but there is no need to get all the
> updated columns here.
> !image-2023-10-09-15-58-47-267.png|width=690,height=218!
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
