[
https://issues.apache.org/jira/browse/KUDU-2826?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16846554#comment-16846554
]
Xu Yao commented on KUDU-2826:
------------------------------
In our scenario, 90% of the write data for some tables is updated data, and is
mainly concentrated on several hotspot primary keys. So I think that this
optimization is necessary. :)
> Add redo tail for mutation list to accelerate UPDATE in memrowset
> -----------------------------------------------------------------
>
> Key: KUDU-2826
> URL: https://issues.apache.org/jira/browse/KUDU-2826
> Project: Kudu
> Issue Type: Improvement
> Reporter: ZhangYao
> Assignee: ZhangYao
> Priority: Major
> Attachments: 1.png, 2.png, 3.png
>
>
> We found kudu needs traversing the whole mutation list to find tail to
> append the new mutations when UPDATE or REINSERT in memrowset. And it cost
> considerable time when we update the same key tens of thousands times(We
> encountered this case in production environment).
> I reproduced this case in offline environment by updating 100 rows for
> about 60000 times, the find-tail costs would be more than 1ms when update
> frequently and the statistics was shown in picture 1. The total
> AppendToListAtomic cost statistics was shown in picture 2. It is meaningful
> for kudu to optimize the append and I try to add tail for mutation list to
> handle it. Picture 3 is the cost statistics for AppendToListAtomic after my
> modification.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
