You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@flink.apache.org by Jay Ghiya <gh...@gmail.com> on 2022/05/06 18:03:27 UTC
Flink Kubernetes operator not having a scale subresource
Hi Team,
I have been experimenting the Flink Kubernetes operator. One of the biggest miss that we have is it does not support scale sub resource as of now to support reactive scaling. Without that commercially it becomes very difficult for products like us who have very varied loads for every hour.
Can I get some direction on the same to contribute on https://kubernetes.io/docs/tasks/extend-kubernetes/custom-resources/custom-resource-definitions/#scale-subresource <https://kubernetes.io/docs/tasks/extend-kubernetes/custom-resources/custom-resource-definitions/#scale-subresource> for our Kubernetes operator crd?
I have been a hard time reading -> https://github.com/apache/flink-kubernetes-operator/blob/main/helm/flink-kubernetes-operator/crds/flinkdeployments.flink.apache.org-v1.yml <https://github.com/apache/flink-kubernetes-operator/blob/main/helm/flink-kubernetes-operator/crds/flinkdeployments.flink.apache.org-v1.yml> to figure out the replicas, status,label selector json path of task manager? It may be due to lack of my knowledge so sense of direction will help me.
-Jay
GEHC
Re: Flink Kubernetes operator not having a scale subresource
Posted by Gyula Fóra <gy...@apache.org>.
Hi Team!
This is probably something for after the release but I created a simple
prototype for the scaling subresource based on taskmanager replica count.
You can take a look here:
https://github.com/apache/flink-kubernetes-operator/pull/227
After some consideration I decided against using parallelism and used tm
replicas instead (still with native integration), I describe this in the PR.
I will leave the PR open so people can experiment/comment and we should
definitely get back to this after the 1.0.0 release because it seems to be
a very lightweight yet useful feature.
Cheers,
Gyula
On Sat, May 7, 2022 at 11:25 AM Gyula Fóra <gy...@apache.org> wrote:
> Hi Jay!
>
> I will take a closer look into this and see if we can use the parallelism
> in the scale subresource.
>
> If you could experiment with this and see if it works with the current CRD
> that would be helpful . Not sure if we need to change the status or
> anything as parallelism is only part of the spec at the moment.
>
> If you have a working modified CRD I would appreciate if you could share
> it with us!
>
> Don’t worry about the release schedule, if we think that this is important
> and we need some changes for it , we can push the release out a few days if
> necessary.
>
> What is important at this point to understand what exactly we need to make
> the parallelism scaling work natively to avoid breaking changes to the
> spec/status after the release :)
>
> Cheers
> Gyula
>
> On Sat, 7 May 2022 at 11:14, Jay Ghiya <gh...@gmail.com> wrote:
>
>> Hi Team,
>>
>> Yes we can change the parallelism of flink job. So going through the
>> roadmap , what I understand that we have put the standalone mode as second
>> priority due to right reasons. So , if possible can I be of any help to
>> accelerate this as we have a tight release schedule so would want to close
>> this in next 10 days with your guys’ help.
>>
>> Looking forward to hear from you !
>>
>> -Jay
>>
>> Sent with a Spark <https://sparkmailapp.com/source?from=signature>
>> On 7 May 2022, 8:15 AM +0530, Yang Wang <da...@gmail.com>, wrote:
>>
>> Currently, the flink-kubernetes-operator is using Flink native K8s
>> integration[1], which means Flink ResourceManager will dynamically allocate
>> TaskManager on demand.
>> So the users do not need to specify the replicas of TaskManager.
>>
>> Just like Gyula said, one possible solution to make "kubectl scale" work
>> is to change the parallelism of Flink job.
>>
>> If the standalone mode[2] is introduced in the operator, then it is also
>> possible to directly change the replicas of TaskManager pods.
>>
>>
>> [1].
>> https://nightlies.apache.org/flink/flink-docs-release-1.15/docs/deployment/resource-providers/native_kubernetes/
>> [2].
>> https://cwiki.apache.org/confluence/display/FLINK/FLIP-225%3A+Implement+standalone+mode+support+in+the+kubernetes+operator
>>
>> Best,
>> Yang
>>
>> Gyula Fóra <gy...@apache.org> 于2022年5月7日周六 04:26写道：
>>
>>> Hi Jay!
>>>
>>> Interesting question/proposal to add the scale-subresource.
>>>
>>> I am not an expert on this area but we will look into this a little and
>>> give you some feedback and see if we can incorporate something into the
>>> upcoming release if it makes sense.
>>>
>>> On a high level there is not a single replicas value for a
>>> FlinkDeployment that would be easy to map, but maybe we could use the
>>> parallelism value for this purpose for Applications/Session jobs.
>>>
>>> Cheers,
>>> Gyula
>>>
>>> On Fri, May 6, 2022 at 8:04 PM Jay Ghiya <gh...@gmail.com> wrote:
>>>
>>>> Hi Team,
>>>>
>>>>
>>>> I have been experimenting the Flink Kubernetes operator. One of the
>>>> biggest miss that we have is it does not support scale sub resource as of
>>>> now to support reactive scaling. Without that commercially it becomes very
>>>> difficult for products like us who have very varied loads for every hour.
>>>>
>>>>
>>>>
>>>> Can I get some direction on the same to contribute on
>>>> https://kubernetes.io/docs/tasks/extend-kubernetes/custom-resources/custom-resource-definitions/#scale-subresource for
>>>> our Kubernetes operator crd?
>>>>
>>>> I have been a hard time reading -> *https://github.com/apache/flink-kubernetes-operator/blob/main/helm/flink-kubernetes-operator/crds/flinkdeployments.flink.apache.org-v1.yml
>>>> <https://github.com/apache/flink-kubernetes-operator/blob/main/helm/flink-kubernetes-operator/crds/flinkdeployments.flink.apache.org-v1.yml> to
>>>> figure out the replicas, status,label selector json path of task
>>>> manager? It may be due to lack of my knowledge so sense of direction will
>>>> help me.*
>>>>
>>>> *-Jay*
>>>> *GEHC*
>>>>
>>>
Re: Flink Kubernetes operator not having a scale subresource
Posted by Gyula Fóra <gy...@apache.org>.
Hi Team!
This is probably something for after the release but I created a simple
prototype for the scaling subresource based on taskmanager replica count.
You can take a look here:
https://github.com/apache/flink-kubernetes-operator/pull/227
After some consideration I decided against using parallelism and used tm
replicas instead (still with native integration), I describe this in the PR.
I will leave the PR open so people can experiment/comment and we should
definitely get back to this after the 1.0.0 release because it seems to be
a very lightweight yet useful feature.
Cheers,
Gyula
On Sat, May 7, 2022 at 11:25 AM Gyula Fóra <gy...@apache.org> wrote:
> Hi Jay!
>
> I will take a closer look into this and see if we can use the parallelism
> in the scale subresource.
>
> If you could experiment with this and see if it works with the current CRD
> that would be helpful . Not sure if we need to change the status or
> anything as parallelism is only part of the spec at the moment.
>
> If you have a working modified CRD I would appreciate if you could share
> it with us!
>
> Don’t worry about the release schedule, if we think that this is important
> and we need some changes for it , we can push the release out a few days if
> necessary.
>
> What is important at this point to understand what exactly we need to make
> the parallelism scaling work natively to avoid breaking changes to the
> spec/status after the release :)
>
> Cheers
> Gyula
>
> On Sat, 7 May 2022 at 11:14, Jay Ghiya <gh...@gmail.com> wrote:
>
>> Hi Team,
>>
>> Yes we can change the parallelism of flink job. So going through the
>> roadmap , what I understand that we have put the standalone mode as second
>> priority due to right reasons. So , if possible can I be of any help to
>> accelerate this as we have a tight release schedule so would want to close
>> this in next 10 days with your guys’ help.
>>
>> Looking forward to hear from you !
>>
>> -Jay
>>
>> Sent with a Spark <https://sparkmailapp.com/source?from=signature>
>> On 7 May 2022, 8:15 AM +0530, Yang Wang <da...@gmail.com>, wrote:
>>
>> Currently, the flink-kubernetes-operator is using Flink native K8s
>> integration[1], which means Flink ResourceManager will dynamically allocate
>> TaskManager on demand.
>> So the users do not need to specify the replicas of TaskManager.
>>
>> Just like Gyula said, one possible solution to make "kubectl scale" work
>> is to change the parallelism of Flink job.
>>
>> If the standalone mode[2] is introduced in the operator, then it is also
>> possible to directly change the replicas of TaskManager pods.
>>
>>
>> [1].
>> https://nightlies.apache.org/flink/flink-docs-release-1.15/docs/deployment/resource-providers/native_kubernetes/
>> [2].
>> https://cwiki.apache.org/confluence/display/FLINK/FLIP-225%3A+Implement+standalone+mode+support+in+the+kubernetes+operator
>>
>> Best,
>> Yang
>>
>> Gyula Fóra <gy...@apache.org> 于2022年5月7日周六 04:26写道：
>>
>>> Hi Jay!
>>>
>>> Interesting question/proposal to add the scale-subresource.
>>>
>>> I am not an expert on this area but we will look into this a little and
>>> give you some feedback and see if we can incorporate something into the
>>> upcoming release if it makes sense.
>>>
>>> On a high level there is not a single replicas value for a
>>> FlinkDeployment that would be easy to map, but maybe we could use the
>>> parallelism value for this purpose for Applications/Session jobs.
>>>
>>> Cheers,
>>> Gyula
>>>
>>> On Fri, May 6, 2022 at 8:04 PM Jay Ghiya <gh...@gmail.com> wrote:
>>>
>>>> Hi Team,
>>>>
>>>>
>>>> I have been experimenting the Flink Kubernetes operator. One of the
>>>> biggest miss that we have is it does not support scale sub resource as of
>>>> now to support reactive scaling. Without that commercially it becomes very
>>>> difficult for products like us who have very varied loads for every hour.
>>>>
>>>>
>>>>
>>>> Can I get some direction on the same to contribute on
>>>> https://kubernetes.io/docs/tasks/extend-kubernetes/custom-resources/custom-resource-definitions/#scale-subresource for
>>>> our Kubernetes operator crd?
>>>>
>>>> I have been a hard time reading -> *https://github.com/apache/flink-kubernetes-operator/blob/main/helm/flink-kubernetes-operator/crds/flinkdeployments.flink.apache.org-v1.yml
>>>> <https://github.com/apache/flink-kubernetes-operator/blob/main/helm/flink-kubernetes-operator/crds/flinkdeployments.flink.apache.org-v1.yml> to
>>>> figure out the replicas, status,label selector json path of task
>>>> manager? It may be due to lack of my knowledge so sense of direction will
>>>> help me.*
>>>>
>>>> *-Jay*
>>>> *GEHC*
>>>>
>>>
Re: Flink Kubernetes operator not having a scale subresource
Posted by Gyula Fóra <gy...@apache.org>.
Hi Jay!
I will take a closer look into this and see if we can use the parallelism
in the scale subresource.
If you could experiment with this and see if it works with the current CRD
that would be helpful . Not sure if we need to change the status or
anything as parallelism is only part of the spec at the moment.
If you have a working modified CRD I would appreciate if you could share it
with us!
Don’t worry about the release schedule, if we think that this is important
and we need some changes for it , we can push the release out a few days if
necessary.
What is important at this point to understand what exactly we need to make
the parallelism scaling work natively to avoid breaking changes to the
spec/status after the release :)
Cheers
Gyula
On Sat, 7 May 2022 at 11:14, Jay Ghiya <gh...@gmail.com> wrote:
> Hi Team,
>
> Yes we can change the parallelism of flink job. So going through the
> roadmap , what I understand that we have put the standalone mode as second
> priority due to right reasons. So , if possible can I be of any help to
> accelerate this as we have a tight release schedule so would want to close
> this in next 10 days with your guys’ help.
>
> Looking forward to hear from you !
>
> -Jay
>
> Sent with a Spark <https://sparkmailapp.com/source?from=signature>
> On 7 May 2022, 8:15 AM +0530, Yang Wang <da...@gmail.com>, wrote:
>
> Currently, the flink-kubernetes-operator is using Flink native K8s
> integration[1], which means Flink ResourceManager will dynamically allocate
> TaskManager on demand.
> So the users do not need to specify the replicas of TaskManager.
>
> Just like Gyula said, one possible solution to make "kubectl scale" work
> is to change the parallelism of Flink job.
>
> If the standalone mode[2] is introduced in the operator, then it is also
> possible to directly change the replicas of TaskManager pods.
>
>
> [1].
> https://nightlies.apache.org/flink/flink-docs-release-1.15/docs/deployment/resource-providers/native_kubernetes/
> [2].
> https://cwiki.apache.org/confluence/display/FLINK/FLIP-225%3A+Implement+standalone+mode+support+in+the+kubernetes+operator
>
> Best,
> Yang
>
> Gyula Fóra <gy...@apache.org> 于2022年5月7日周六 04:26写道：
>
>> Hi Jay!
>>
>> Interesting question/proposal to add the scale-subresource.
>>
>> I am not an expert on this area but we will look into this a little and
>> give you some feedback and see if we can incorporate something into the
>> upcoming release if it makes sense.
>>
>> On a high level there is not a single replicas value for a
>> FlinkDeployment that would be easy to map, but maybe we could use the
>> parallelism value for this purpose for Applications/Session jobs.
>>
>> Cheers,
>> Gyula
>>
>> On Fri, May 6, 2022 at 8:04 PM Jay Ghiya <gh...@gmail.com> wrote:
>>
>>> Hi Team,
>>>
>>>
>>> I have been experimenting the Flink Kubernetes operator. One of the
>>> biggest miss that we have is it does not support scale sub resource as of
>>> now to support reactive scaling. Without that commercially it becomes very
>>> difficult for products like us who have very varied loads for every hour.
>>>
>>>
>>>
>>> Can I get some direction on the same to contribute on
>>> https://kubernetes.io/docs/tasks/extend-kubernetes/custom-resources/custom-resource-definitions/#scale-subresource for
>>> our Kubernetes operator crd?
>>>
>>> I have been a hard time reading -> *https://github.com/apache/flink-kubernetes-operator/blob/main/helm/flink-kubernetes-operator/crds/flinkdeployments.flink.apache.org-v1.yml
>>> <https://github.com/apache/flink-kubernetes-operator/blob/main/helm/flink-kubernetes-operator/crds/flinkdeployments.flink.apache.org-v1.yml> to
>>> figure out the replicas, status,label selector json path of task
>>> manager? It may be due to lack of my knowledge so sense of direction will
>>> help me.*
>>>
>>> *-Jay*
>>> *GEHC*
>>>
>>
Re: Flink Kubernetes operator not having a scale subresource
Posted by Jay Ghiya <gh...@gmail.com>.
Hi Team,
Yes we can change the parallelism of flink job. So going through the roadmap , what I understand that we have put the standalone mode as second priority due to right reasons. So , if possible can I be of any help to accelerate this as we have a tight release schedule so would want to close this in next 10 days with your guys’ help.
Looking forward to hear from you !
-Jay
Sent with a Spark
On 7 May 2022, 8:15 AM +0530, Yang Wang <da...@gmail.com>, wrote:
> Currently, the flink-kubernetes-operator is using Flink native K8s integration[1], which means Flink ResourceManager will dynamically allocate TaskManager on demand.
> So the users do not need to specify the replicas of TaskManager.
>
> Just like Gyula said, one possible solution to make "kubectl scale" work is to change the parallelism of Flink job.
>
> If the standalone mode[2] is introduced in the operator, then it is also possible to directly change the replicas of TaskManager pods.
>
>
> [1]. https://nightlies.apache.org/flink/flink-docs-release-1.15/docs/deployment/resource-providers/native_kubernetes/
> [2]. https://cwiki.apache.org/confluence/display/FLINK/FLIP-225%3A+Implement+standalone+mode+support+in+the+kubernetes+operator
>
> Best,
> Yang
>
> > Gyula Fóra <gy...@apache.org> 于2022年5月7日周六 04:26写道：
> > > Hi Jay!
> > >
> > > Interesting question/proposal to add the scale-subresource.
> > >
> > > I am not an expert on this area but we will look into this a little and give you some feedback and see if we can incorporate something into the upcoming release if it makes sense.
> > >
> > > On a high level there is not a single replicas value for a FlinkDeployment that would be easy to map, but maybe we could use the parallelism value for this purpose for Applications/Session jobs.
> > >
> > > Cheers,
> > > Gyula
> > >
> > > > On Fri, May 6, 2022 at 8:04 PM Jay Ghiya <gh...@gmail.com> wrote:
> > > > > Hi Team,
> > > > >
> > > > >
> > > > > I have been experimenting the Flink Kubernetes operator. One of the biggest miss that we have is it does not support scale sub resource as of now to support reactive scaling. Without that commercially it becomes very difficult for products like us who have very varied loads for every hour.
> > > > >
> > > > >
> > > > >
> > > > > Can I get some direction on the same to contribute on https://kubernetes.io/docs/tasks/extend-kubernetes/custom-resources/custom-resource-definitions/#scale-subresource for our Kubernetes operator crd?
> > > > >
> > > > > I have been a hard time reading -> https://github.com/apache/flink-kubernetes-operator/blob/main/helm/flink-kubernetes-operator/crds/flinkdeployments.flink.apache.org-v1.yml to figure out the replicas, status,label selector json path of task manager? It may be due to lack of my knowledge so sense of direction will help me.
> > > > >
> > > > > -Jay
> > > > > GEHC
Re: Flink Kubernetes operator not having a scale subresource
Posted by Yang Wang <da...@gmail.com>.
Currently, the flink-kubernetes-operator is using Flink native K8s
integration[1], which means Flink ResourceManager will dynamically allocate
TaskManager on demand.
So the users do not need to specify the replicas of TaskManager.
Just like Gyula said, one possible solution to make "kubectl scale" work is
to change the parallelism of Flink job.
If the standalone mode[2] is introduced in the operator, then it is also
possible to directly change the replicas of TaskManager pods.
[1].
https://nightlies.apache.org/flink/flink-docs-release-1.15/docs/deployment/resource-providers/native_kubernetes/
[2].
https://cwiki.apache.org/confluence/display/FLINK/FLIP-225%3A+Implement+standalone+mode+support+in+the+kubernetes+operator
Best,
Yang
Gyula Fóra <gy...@apache.org> 于2022年5月7日周六 04:26写道：
> Hi Jay!
>
> Interesting question/proposal to add the scale-subresource.
>
> I am not an expert on this area but we will look into this a little and
> give you some feedback and see if we can incorporate something into the
> upcoming release if it makes sense.
>
> On a high level there is not a single replicas value for a
> FlinkDeployment that would be easy to map, but maybe we could use the
> parallelism value for this purpose for Applications/Session jobs.
>
> Cheers,
> Gyula
>
> On Fri, May 6, 2022 at 8:04 PM Jay Ghiya <gh...@gmail.com> wrote:
>
>> Hi Team,
>>
>>
>> I have been experimenting the Flink Kubernetes operator. One of the
>> biggest miss that we have is it does not support scale sub resource as of
>> now to support reactive scaling. Without that commercially it becomes very
>> difficult for products like us who have very varied loads for every hour.
>>
>>
>>
>> Can I get some direction on the same to contribute on
>> https://kubernetes.io/docs/tasks/extend-kubernetes/custom-resources/custom-resource-definitions/#scale-subresource for
>> our Kubernetes operator crd?
>>
>> I have been a hard time reading -> *https://github.com/apache/flink-kubernetes-operator/blob/main/helm/flink-kubernetes-operator/crds/flinkdeployments.flink.apache.org-v1.yml
>> <https://github.com/apache/flink-kubernetes-operator/blob/main/helm/flink-kubernetes-operator/crds/flinkdeployments.flink.apache.org-v1.yml> to
>> figure out the replicas, status,label selector json path of task
>> manager? It may be due to lack of my knowledge so sense of direction will
>> help me.*
>>
>> *-Jay*
>> *GEHC*
>>
>
Re: Flink Kubernetes operator not having a scale subresource
Posted by Gyula Fóra <gy...@apache.org>.
Hi Jay!
Interesting question/proposal to add the scale-subresource.
I am not an expert on this area but we will look into this a little and
give you some feedback and see if we can incorporate something into the
upcoming release if it makes sense.
On a high level there is not a single replicas value for a
FlinkDeployment that would be easy to map, but maybe we could use the
parallelism value for this purpose for Applications/Session jobs.
Cheers,
Gyula
On Fri, May 6, 2022 at 8:04 PM Jay Ghiya <gh...@gmail.com> wrote:
> Hi Team,
>
>
> I have been experimenting the Flink Kubernetes operator. One of the
> biggest miss that we have is it does not support scale sub resource as of
> now to support reactive scaling. Without that commercially it becomes very
> difficult for products like us who have very varied loads for every hour.
>
>
>
> Can I get some direction on the same to contribute on
> https://kubernetes.io/docs/tasks/extend-kubernetes/custom-resources/custom-resource-definitions/#scale-subresource for
> our Kubernetes operator crd?
>
> I have been a hard time reading -> *https://github.com/apache/flink-kubernetes-operator/blob/main/helm/flink-kubernetes-operator/crds/flinkdeployments.flink.apache.org-v1.yml
> <https://github.com/apache/flink-kubernetes-operator/blob/main/helm/flink-kubernetes-operator/crds/flinkdeployments.flink.apache.org-v1.yml> to
> figure out the replicas, status,label selector json path of task
> manager? It may be due to lack of my knowledge so sense of direction will
> help me.*
>
> *-Jay*
> *GEHC*
>