You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@flink.apache.org by Luca Castelli via dev <de...@flink.apache.org> on 2026/08/07 19:27:41 UTC

[DISCUSS] Flink Kubernetes Operator: One-shot statelessRestartNonce

 Hi all,

I would like to get feedback on adding a statelessRestartNonce to the Flink
Kubernetes Operator.

Today, if a job normally uses upgradeMode: savepoint or last-state,
triggering a fresh restart currently requires:

1. Changing upgradeMode to stateless
2. Bumping restartNonce
3. Changing upgradeMode back afterward

This temporarily mixes an operational action with the job's steady-state
recovery policy. Users sometimes forget the third step, leaving
subsequent upgrades (deploys or scaling events) stateless and creating a
risk of state loss.

Would the community be open to this approach?

 ```yaml
   spec:
     job:
       statelessRestartNonce: 42
 ```

Alternatively, we could name it statelessRedeployNonce.

Thanks,
Luca

退订

Posted by 18524655605 <18...@163.com>.

> 退订

退订

Posted by Yanko24 <ya...@foxmail.com>.
退订

Re: [DISCUSS] Flink Kubernetes Operator: One-shot statelessRestartNonce

Posted by Luca Castelli via dev <de...@flink.apache.org>.
Hi Gyula!

Great to hear from you, and thanks for your feedback.

I like your alternative approach. I'm curious what you are thinking
regarding one-shot configs/settings - is there a common use-case you have
in mind?

Luca

On Sat, Aug 8, 2026 at 2:47 AM Gyula Fóra <gy...@gmail.com> wrote:

> Hey Luca!
>
> I have thought about this a few times as well that it may be better to
> make an operation like this.
>
> I don't really like the idea of having multiple distinct nonces for
> overlapping operations (savepointRedeployNonce, statelessRedeployNonce etc.)
>
> An alternative approach could be if we want to make one-shot
> deployment operations more flexible is to introduce a new spec object field
> such as:
>
>  ```yaml
>    spec:
>      job:
>        redeploy:
>            nonce: x -> y // The main trigger
>            state: null / initial state reference
>            // Other future additions such as one-shot configs/settings etc.
>  ```
>
> What do you think?
> Gyula
>
>
> On Fri, Aug 7, 2026 at 9:28 PM Luca Castelli via dev <de...@flink.apache.org>
> wrote:
>
>>  Hi all,
>>
>> I would like to get feedback on adding a statelessRestartNonce to the
>> Flink
>> Kubernetes Operator.
>>
>> Today, if a job normally uses upgradeMode: savepoint or last-state,
>> triggering a fresh restart currently requires:
>>
>> 1. Changing upgradeMode to stateless
>> 2. Bumping restartNonce
>> 3. Changing upgradeMode back afterward
>>
>> This temporarily mixes an operational action with the job's steady-state
>> recovery policy. Users sometimes forget the third step, leaving
>> subsequent upgrades (deploys or scaling events) stateless and creating a
>> risk of state loss.
>>
>> Would the community be open to this approach?
>>
>>  ```yaml
>>    spec:
>>      job:
>>        statelessRestartNonce: 42
>>  ```
>>
>> Alternatively, we could name it statelessRedeployNonce.
>>
>> Thanks,
>> Luca
>>
>

Re: [DISCUSS] Flink Kubernetes Operator: One-shot statelessRestartNonce

Posted by Gyula Fóra <gy...@gmail.com>.
Hey Luca!

I have thought about this a few times as well that it may be better to make
an operation like this.

I don't really like the idea of having multiple distinct nonces for
overlapping operations (savepointRedeployNonce, statelessRedeployNonce etc.)

An alternative approach could be if we want to make one-shot
deployment operations more flexible is to introduce a new spec object field
such as:

 ```yaml
   spec:
     job:
       redeploy:
           nonce: x -> y // The main trigger
           state: null / initial state reference
           // Other future additions such as one-shot configs/settings etc.
 ```

What do you think?
Gyula


On Fri, Aug 7, 2026 at 9:28 PM Luca Castelli via dev <de...@flink.apache.org>
wrote:

>  Hi all,
>
> I would like to get feedback on adding a statelessRestartNonce to the Flink
> Kubernetes Operator.
>
> Today, if a job normally uses upgradeMode: savepoint or last-state,
> triggering a fresh restart currently requires:
>
> 1. Changing upgradeMode to stateless
> 2. Bumping restartNonce
> 3. Changing upgradeMode back afterward
>
> This temporarily mixes an operational action with the job's steady-state
> recovery policy. Users sometimes forget the third step, leaving
> subsequent upgrades (deploys or scaling events) stateless and creating a
> risk of state loss.
>
> Would the community be open to this approach?
>
>  ```yaml
>    spec:
>      job:
>        statelessRestartNonce: 42
>  ```
>
> Alternatively, we could name it statelessRedeployNonce.
>
> Thanks,
> Luca
>