[jira] [Closed] (FLINK-30571) Compute scale parallelism based on observed scalability

["Gyula Fora (Jira)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/FLINK-30571?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Gyula Fora closed FLINK-30571.
------------------------------
    Resolution: Fixed

> Compute scale parallelism based on observed scalability 
> --------------------------------------------------------
>
>                 Key: FLINK-30571
>                 URL: https://issues.apache.org/jira/browse/FLINK-30571
>             Project: Flink
>          Issue Type: New Feature
>          Components: Autoscaler, Kubernetes Operator
>            Reporter: Gyula Fora
>            Priority: Major
>              Labels: pull-request-available
>
> When computing target parallelism for job vertices we currently assume linear scaling with a fixed (1) coefficient.
> This assumes that in order to double the capacity we simply double the parallelism.
> While linearity already might be violated by many real time workloads this form of strong linearity rarely holds due to the overhead of increased network traffic, coordination etc.
> As we can access past (parallelism, processingRate) information based on the scaling history we should estimate the scalability coefficient either using a simple or weighted linear regression.



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