[
https://issues.apache.org/jira/browse/FLINK-30464?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Maximilian Michels updated FLINK-30464:
---------------------------------------
Description:
The current code collects metrics as soon as the job goes into RUNNING mode,
regardless of whether we are still in the stabilization period. This means the
first metric collection window after rescaling will include metrics from the
stabilization period which is not desired. The stabilization period is supposed
to allow the job to stabilize and shouldn't be used to make new scaling
decisions.
Further, the collected metrics as input for the scaling decider initially do
not span a full metric window. Only after the first metric window is complete,
we are guaranteed to have a full windows available.
was:The current code collects metrics as soon as the job goes into RUNNING
mode, regardless of whether we are still in the stabilization period. This
means the first metric collection window after rescaling will include metrics
from the stabilization period which is not desired. The stabilization period is
supposed to allow the job to stabilize and shouldn't be used to make new
scaling decisions.
> Metrics are collected in stabilization phase and do not always span a full
> metric window
> ----------------------------------------------------------------------------------------
>
> Key: FLINK-30464
> URL: https://issues.apache.org/jira/browse/FLINK-30464
> Project: Flink
> Issue Type: Bug
> Components: Kubernetes Operator
> Affects Versions: kubernetes-operator-1.4.0
> Reporter: Maximilian Michels
> Assignee: Maximilian Michels
> Priority: Major
> Fix For: kubernetes-operator-1.4.0
>
>
> The current code collects metrics as soon as the job goes into RUNNING mode,
> regardless of whether we are still in the stabilization period. This means
> the first metric collection window after rescaling will include metrics from
> the stabilization period which is not desired. The stabilization period is
> supposed to allow the job to stabilize and shouldn't be used to make new
> scaling decisions.
> Further, the collected metrics as input for the scaling decider initially do
> not span a full metric window. Only after the first metric window is
> complete, we are guaranteed to have a full windows available.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
