> 3. Launcher placement - Here I'd go the opposite way of the current doc.
> Currently, every deployment shape keeps its bootstrap in the Flink
> distribution, and that convention is what decouples operator version from
> Flink version. My opinion is that the MiniCluster launcher should follow
> it. In practice, the launcher is close to `flink run -t local`, only adding
> HA-aware resubmission and pod exit-code semantics on top, so it is core
> code either way. I'd also say it makes sense to ship it in the distribution
> (ideally with a bin/ script, so the same setup runs locally too) because
> this means the operator depends on no MiniCluster Java API at all, and it
> can manage the deployment in the same unified way as the other shapes.

I really like this idea, thanks - that's much better than what I'd originally 
thought of. It simplifies a few things, such as migration between MiniCluster 
and Deployment by letting you just use the same image for both as-is.  I'll 
have a proper think through what it would need and update the doc. 

Do you think a Flink launcher would need a separate/dedicated FLIP to get 
broader scrutiny? 

Kind regards

D
--
dalelane.co.uk

On Tuesday, 25 August 2026 at 14:10, Dennis-Mircea Ciupitu 
<[email protected]> wrote:

> Hi Dale,
> 
> My response to the open points is the following:
> 
>    1. MiniCluster vs 2 containers in one pod - Robert's numbers and
>    argument make it clear to me. So, in order to ensure the lowest possible
>    resource consumption, we need to go with the MiniCluster approach and not
>    rely on something intermediary that does not reach this end goal and is
>    only better suited from the implementation/design perspective.
>    2. Separate CRD vs a third mode - Because the MiniCluster approach is
>    the way to go, I'd say that we should have an implementation that
>    integrates with it in the best way possible. For me, I see a separate CRD
>    as the best way to proceed here, as MiniCluster is different from many
>    points of view from a normal Flink cluster, and by going with a third mode
>    I think we just complicate ourselves and we introduce too much blast 
> radius.
>    3. Launcher placement - Here I'd go the opposite way of the current doc.
>    Currently, every deployment shape keeps its bootstrap in the Flink
>    distribution, and that convention is what decouples operator version from
>    Flink version. My opinion is that the MiniCluster launcher should follow
>    it. In practice, the launcher is close to `flink run -t local`, only adding
>    HA-aware resubmission and pod exit-code semantics on top, so it is core
>    code either way. I'd also say it makes sense to ship it in the distribution
>    (ideally with a bin/ script, so the same setup runs locally too) because
>    this means the operator depends on no MiniCluster Java API at all, and it
>    can manage the deployment in the same unified way as the other shapes.
>    4. Promoting the MiniCluster - Here I say if the entrypoint lands in
>    Flink core, then I'd follow Gyula's and Robert's guidance on this side, if
>    there is something that can be done further here in order to promote it
>    properly.
>    5. Resource and memory mapping - Deferring the exact ratios to
>    implementation is fine with me.
> 
> 
> Best regards,
> Dennis
> 
> 
> On Mon, Aug 24, 2026 at 2:00 PM Dale Lane <[email protected]>
> wrote:
> 
> > > Whether it should be a different CRD I think should
> > > depend on the overlap with the FlinkDeployment. If
> > > we are exposing essentially the same spec (podTemplate,
> > > jobSpec, flinkConfig, ingress..) then it feels like we
> > > should reuse the CRD and simply introduce a new
> > > deployment mode next to native and standalone.
> >
> > @Gyula, @Dennis - How would you like to proceed with this question?
> >
> > I can absolutely see arguments for/against both approaches. In the end I
> > leaned more towards a separate CRD, partly from an API purity perspective
> > (feeling that the separate topology deserved a separate API, and avoids
> > describing the pod spec and status as "jobManager" or "taskManager" when
> > it's both) and partly because it dodged the automated topology migration.
> >
> > But maybe we don't mind the "a MiniCluster includes a Job Manager so we
> > can just describe it as jobManager (and ignore the taskManager references)"
> > API smell. It could change the manual migration path I describe to
> > transition jobs between MiniCluster and full distributed clusters to just
> > flipping a mode value. And we can always reject/block such a mode switch in
> > the meantime unless/until the Operator is ready to support topology
> > transition/migration.
> >
> > Kind regards
> >
> > Dale
> > --
> > dalelane.co.uk
> >
> > Sent with Proton Mail secure email.
> >
> > On Thursday, 20 August 2026 at 10:47, Gyula Fóra <[email protected]>
> > wrote:
> >
> > > Thanks @Robert Metzger for the explanation,
> > >
> > > Overall I think this is a good idea and hopefully we can reuse most of
> > > the controller flow that we already have.
> > > It would be nice to share the spec/status as much as reasonably possible.
> > >
> > > Whether it should be a different CRD I think should depend on the
> > > overlap with the FlinkDeployment. If we are exposing essentially the
> > > same spec (podTemplate, jobSpec, flinkConfig, ingress..) then it feels
> > > like we should reuse the CRD and simply introduce a new deployment
> > > mode next to native and standalone.
> > >
> > > Cheers
> > > Gyula
> > >
> > > On Thu, Aug 20, 2026 at 11:26 AM Dale Lane
> > > <[email protected]> wrote:
> > > >
> > > > This bit that I added to the FLIP is perhaps worthy of a bit more
> > scrutiny, as it's based on my assumptions (misunderstandings?) of how the
> > FKO project works.
> > > >
> > > > In the section "Promoting MiniCluster to a stable API" I asserted:
> > > >
> > > > > For the Flink Kubernetes Operator project to provide a launcher
> > > > > as an example project would depend on the core Flink project
> > > > > promoting some MiniCluster components to @PublicEvolving ...
> > > > > This is not a technical requirement, as the current
> > > > > proof-of-concept demonstrates that the existing API is sufficient
> > > > >  - the promotion recommended here is of MiniCluster’s existing
> > > > > submission surface. The requirement is about the Operator project
> > > > > managing risk by building upon a stable contract. ...
> > > >
> > > > Is that really a requirement or have I just invented that? :)
> > > >
> > > > Looking more closely at the existing code, I can see we already have
> > plenty of uses of flink-runtime classes that don't have @Public /
> > @PublicEvolving annotations, so my usage of flink-runtime classes like the
> > MiniCluster and MiniClusterConfiguration wouldn't be without precedent.
> > > >
> > > > While I still think it'd be lovely to build this feature solely on
> > stable APIs, maybe I'm creating an unreasonably high bar to clear by
> > framing it this so strongly.
> > > >
> > > > What do you think? What are the norms here?
> > > >
> > > > Kind regards
> > > >
> > > > Dale
> > > > --
> > > > dalelane.co.uk
> > > >
> > > >
> > > > On Wednesday, 5 August 2026 at 16:02, Dale Lane <
> > [email protected]> wrote:
> > > >
> > > > > I'd like to start a discussion on
> > > > > FLIP-XXX : Running Flink jobs in MiniCluster using the Kubernetes
> > Operator
> > > > >
> > https://docs.google.com/document/d/1dtGjPYcsBkx1vxHPs1QnDtPxeH_Acz_pl8gx4b1BLB4/edit?usp=sharing
> > > > >
> > > > > The aim of the FLIP is to extend the Flink Kubernetes Operator to
> > offer a single-pod, light-weight deployment option for low-throughput jobs.
> > > > >
> > > > > From the motivation:
> > > > > A single-pod, self-contained Flink job that starts fast and needs no
> > multi-pod coordination could be a good fit for low-throughput jobs that
> > aren't suitable for session clusters because they need isolation.
> > > > >
> > > > > Looking forward to feedback, both on the general motivation (Have
> > you seen a need for very small lightweight Flink jobs where fast
> > crash-consistent resume is good enough without a full distributed Flink
> > cluster?) and the suggested implementation approach (Do you think a new
> > custom resource kind is the best way to represent this capability?)
> > > > >
> > > > > Kind regards
> > > > >
> > > > > Dale
> > > > > --
> > > > > dalelane.co.uk
> > > > >
> > >
> >
>

Reply via email to