[
https://issues.apache.org/jira/browse/FLINK-36870?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17904332#comment-17904332
]
Rui Fan commented on FLINK-36870:
---------------------------------
Thanks [~mxm] for creating this Jira and cc.
{quote}the first time it is scaled up should be the start of the time tracking
{quote}
>From my understanding, it cannot reduce the frequency of scale down. So I
>might misunderstand it, would you mind explaining the detailed mechanism?
Assuming current parallelism is 100, and scale down interval is 1 hour,
following is my understanding:
* 2024-12-09 20:00:00 -> from 100 to 120 (scale up)
* 2024-12-09 21:30:00 -> from 120 to 100 (First scale down)
* 2024-12-09 22:00:00 -> from 100 to 80 (Second scale down)
* 2024-12-09 22:30:01 -> from 80 to 60 (Third scale down)
If we use scaled up as the start of the time tracking, it will be 2024-12-09
20:00:00.
* When job wants to scale down at 21:30:00 or 22:00:00 or 22:30:01, the
(current time - start of the time tracking) > 1 hour, so job executes scale
down 3 times.
If we use the first scale down request as the start of the time tracking, it
will be 2024-12-09 21:30:00.
* So 2024-12-09 21:30:00 and 2024-12-09 22:00:00 don't execute scale down
since (current time - start of the time tracking) < 1 hour.
* We only execute scale down at 2024-12-09 22:30:01, since (current time -
start of the time tracking) > 1 hour.
Please correct me if I misunderstand anything, thank you very much :)
> Start scale down timer after initial scale up
> ---------------------------------------------
>
> Key: FLINK-36870
> URL: https://issues.apache.org/jira/browse/FLINK-36870
> Project: Flink
> Issue Type: Improvement
> Components: Autoscaler
> Affects Versions: kubernetes-operator-1.10.0
> Reporter: Maximilian Michels
> Priority: Major
> Fix For: kubernetes-operator-1.10.0
>
>
> FLINK-36018 added the option to combine multiple scale downs of vertices into
> a single scale down operation. The start time for tracking whether a vertex
> is allowed to be scaled down is the first time the algorithm decides to scale
> down the vertex. However, the first time it is scaled up should be the start
> of the time tracking, because that will guarantee the same behavior across
> all vertices.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
