[
https://issues.apache.org/jira/browse/FLINK-17176?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17087361#comment-17087361
]
Yang Wang commented on FLINK-17176:
-----------------------------------
In my opinion, they share the same retry logics(e.g. {{ExponentialBackoff}}),
but they happen in different situations. One is {{createTaskManagerPod}} failed
and then retry, another is pod launched failed, OOM kill, some other fatal
error in TaskManager(creation succeed).
If you insist, i could just do the first part and leave this ticket as a
separate one. When the retry policy is introduced, this ticket will not be very
difficult to implement.
> Slow down Pod re-creation in KubernetesResourceManager#PodCallbackHandler
> -------------------------------------------------------------------------
>
> Key: FLINK-17176
> URL: https://issues.apache.org/jira/browse/FLINK-17176
> Project: Flink
> Issue Type: Improvement
> Components: Deployment / Kubernetes
> Affects Versions: 1.10.0
> Reporter: Canbin Zheng
> Priority: Major
> Fix For: 1.11.0
>
>
> In the native K8s setups, there are some cases that we do not control the
> speed of pod re-creation which poses potential risks to flood the K8s API
> Server in the {{PodCallbackHandler}} implementation of
> {{KubernetesResourceManager.}}
> Here are steps to reproduce this kind of problems:
> # Mount the {{/opt/flink/log}} in the Container of TaskManager to a path on
> the K8s nodes via HostPath, make sure that the path exists but the
> TaskManager process has no write permission. We can achieve this via the
> [user-specified pod template
> support|https://issues.apache.org/jira/browse/FLINK-15656] or just hardcode
> it for testing only.
> # Launch a session cluster
> # Submit a new job to the session cluster, as expected, we can observe that
> the Pod constantly fails quickly during launching the main Container, then
> the {{KubernetesResourceManager#onModified}} is invoked to re-create a new
> Pod immediately, without any speed control.
> To sum up, once the {{KubernetesResourceManager}} receives the Pod *ADD*
> event and that Pod is terminated before successfully registering into the
> {{KubernetesResourceManager}}, the {{KubernetesResourceManager}} will send
> another creation request to K8s API Server immediately.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
