Unsubscribe

[yanglele via user]

Unsubscribe

----- 原始邮件 -----

发件人：Robin Cassan via user<user@flink.apache.org>

发送时间：2023-06-14 23:13:09

收件人：Gyula Fóra<gyula.f...@gmail.com>

抄送：user<user@flink.apache.org>

主 题：Re: Kubernetes operator: config for taskmanager.memory.process.size ignored

Thanks again, maybe the jvm overhead param will act as the margin I want, I'll try that :)

Robin

Le mer. 14 juin 2023 à 15:28, Gyula Fóra <gyula.f...@gmail.com> a écrit :

Again, this has absolutely nothing to do with the Kubernetes Operator, but simply how Flink Kubernetes Memory configs work:

https://nightlies.apache.org/flink/flink-docs-master/docs/deployment/memory/mem_tuning/#configure-memory-for-containers
 

 

You can probably play around with:  jobmanager.memory.jvm-overhead.fraction

 

You can set a larger memory size in the TM spec and increase the jvm overhead fraction.

 

Gyula

 

On Wed, Jun 14, 2023 at 2:46 PM Robin Cassan <robin.cas...@contentsquare.com> wrote:
 

Thanks Gyula for your answer! I'm wondering about your claim:
     > In Flink kubernetes the process is the pod so pod memory is always equal to process memory

Why should the flink TM process use the whole container (and so, the whole pod) memory?

   

Before migrating to the k8s operator, we still used Flink on kubernetes (without the operator) and left a little bit of margin between the process memory and the pod memory, which helped stability. It looks like it cannot be done with the k8s operator though and I wonder why the choice of removing this granularity in the settings

   

Robin

   

Le mer. 14 juin 2023 à 12:20, Gyula Fóra <gyula.f...@gmail.com> a écrit :
   

Basically what happens is that whatever you set to the spec.taskManager.resource.memory will be set in the config as process memory.

In Flink kubernetes the process is the pod so pod memory is always equal to process memory.

       

So basically the spec is a config shorthand, there is no reason to override it as you won't get a different behaviour at the end of the day.

       

Gyula

       

On Wed, Jun 14, 2023 at 11:55 AM Robin Cassan via user <user@flink.apache.org> wrote:
       

Hello all!

         

I am using the flink kubernetes operator and I would like to set the value for `taskmanager.memory.process.size`. I set the desired value in the flinkdeployment resource specs (here, I want 55gb), however it looks like the value that is effectively passed to the taskmanager is the same as the pod memory setting (which is set to 59gb).

         

For example, this flinkdeployment configuration:
         

```

Spec:
             Flink Configuration:
         

    taskmanager.memory.process.size:                                  55gb
         

  Task Manager:
               Resource:
                 Cpu:     6
                 Memory:  59Gb
         

```

will create a pod with 59Gb total memory (as expected) but will also give 59Gb to the memory.process.size instead of 55Gb, as seen in this TM log: `Loading configuration property: taskmanager.memory.process.size, 59Gb`
         

         Maybe this part of the flink k8s operator code is responsible:
         

https://github.com/apache/flink-kubernetes-operator/blob/d43e1ca9050e83b492b2e16b0220afdba4ffa646/flink-kubernetes-operator/src/main/java/org/apache/flink/kubernetes/operator/config/FlinkConfigBuilder.java#L393          

         

If so, I wonder what is the rationale for forcing the flink process memory to be the same as the pod memory?
           Is there a way to bypass that, for example by setting the desired process.memory configuration differently?

         

Thanks!