You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@flink.apache.org by "Yang Wang (Jira)" <ji...@apache.org> on 2022/06/22 03:34:00 UTC
[jira] [Commented] (FLINK-28189) The taskmanager dynamically created by flink native k8s creates pods directly, why not create a deployment?
[ https://issues.apache.org/jira/browse/FLINK-28189?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17557216#comment-17557216 ]
Yang Wang commented on FLINK-28189:
-----------------------------------
First, I believe you could also monitor a pod status while it is created by Flink ResourceManager, not a deployment. The monitoring should be irrelevant with how the pods are created.
Using naked pod has more advantages than k8s deployment for TaskManager.
* Flink could delete specific naked TaskManager pod easily. And it is more difficult to do this in k8s deployment. Imagine that some TaskManager pods in session cluster are idle and we want to release them.
* Flink now could not work normally when we allow the TaskManager pod could be restarted. For example, TaskManager with same hostname:port will fail to register to the ResourceManager with more than once.
* For fine-grained resource management, we might have various TaskManager pods with different spec(2G, 4G, etc.).
> The taskmanager dynamically created by flink native k8s creates pods directly, why not create a deployment?
> -----------------------------------------------------------------------------------------------------------
>
> Key: FLINK-28189
> URL: https://issues.apache.org/jira/browse/FLINK-28189
> Project: Flink
> Issue Type: Improvement
> Components: Deployment / Kubernetes
> Affects Versions: 1.14.4
> Reporter: hob007
> Priority: Major
>
> I am using Flink native k8s deployment. The pods is dynamically created. But there's some problem with K8S cluster monitor. I can't monitor taskmanager pods status, because these pods without deployment.
>
> So my question is:
> The taskmanager dynamically created by flink native k8s creates pods directly, why not create a deployment?
>
> For example:
> Define a deployment and using replicates to control taskmanager's pods.
>
> Chinese description reference:
> https://issues.apache.org/jira/browse/FLINK-28167
--
This message was sent by Atlassian Jira
(v8.20.7#820007)