Does flink support multi cluster deployment with HA on multi k8s clusters

Hi Team,

production requirement is to deploy flink in multi cluster mode,
i.e deploying flink cluster1 with HA on kubernetes cluster1 in data center1
& another flink cluster2  with HA  on kubernetes cluster2 in data center2 ..
if Flink cluster1 goes down on k8s cluster1 on DC1 ,it has to fail over to
Flink cluster2 on k8s cluster2 on DC2.

It has to failover with automatic HA mechanism.
Please let us know, whether is this possible ??
or any solution is provided to have flink in multi cluster mode with HA ..


Any solution please share the information

Note: Currently deployed Flink session cluster on standalone kubernetes with
HA(Kubernetes HA)



--
Sent from: http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/

[Till Rohrmann]

Hi Bhagi,

out of the box, Flink does not support this functionality. What you could
try to do is to configure a stretch K8s cluster running on different DCs
and then having an HA service (ZK or K8s) and blob storage being able to
survive a DC outage. That way Flink should also be able to survive a DC
outage.

Cheers,
Till

On Wed, Apr 28, 2021 at 3:26 PM bhagi@R <[hidden email]> wrote:

> Hi Team,
>
> production requirement is to deploy flink in multi cluster mode,
> i.e deploying flink cluster1 with HA on kubernetes cluster1 in data center1
> & another flink cluster2  with HA  on kubernetes cluster2 in data center2
> ..
> if Flink cluster1 goes down on k8s cluster1 on DC1 ,it has to fail over to
> Flink cluster2 on k8s cluster2 on DC2.
>
> It has to failover with automatic HA mechanism.
> Please let us know, whether is this possible ??
> or any solution is provided to have flink in multi cluster mode with HA ..
>
>
> Any solution please share the information
>
> Note: Currently deployed Flink session cluster on standalone kubernetes
> with
> HA(Kubernetes HA)
>
>
>
> --
> Sent from: http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/
>

Can you explain clearly...

How to configure HA service with k8s for different DC's for Flink cluster ..



--
Sent from: http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/

[Enrique]

Hi Till and Bhagi,

As part of our product, we also have a production requirement to deploy
Flink in multi zones and currently have it configured with Kubernetes HA.

Till, is the reason this is not supported out of the box also due to the
fact Kubernetes HA relies on RWX storage and this is considered an
anti-pattern in the cloud native space? I've seen in Cloud Native
deployments that the recommended way is to replicate the state through
software instead of replicating the state through a shared
filesystem/storage which is a common pattern in data-centers. The cloud
native approach results in each instance of Flink across nodes owning their
own RWO filesystem and communicating through the network to other instance
across nodes to replicate state. I was just wondering if this had been
considered in the move to a more cloud-native Flink deployment?

Cheers,

Enrique





--
Sent from: http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/

[Till Rohrmann]

The reason why this feature is not part of Flink is that it is not trivial
to solve. There are definitely different solutions to the problem. I think
that a solution will probably be built around Flink and not integrated into
it because it goes a bit beyond what Flink's current focus is. I can tell
you that companies are looking at this problem and want to solve it within
their commercial offerings.

Cheers,
Till

On Thu, Apr 29, 2021 at 12:40 PM Enrique <[hidden email]> wrote:

> Hi Till and Bhagi,
>
> As part of our product, we also have a production requirement to deploy
> Flink in multi zones and currently have it configured with Kubernetes HA.
>
> Till, is the reason this is not supported out of the box also due to the
> fact Kubernetes HA relies on RWX storage and this is considered an
> anti-pattern in the cloud native space? I've seen in Cloud Native
> deployments that the recommended way is to replicate the state through
> software instead of replicating the state through a shared
> filesystem/storage which is a common pattern in data-centers. The cloud
> native approach results in each instance of Flink across nodes owning their
> own RWO filesystem and communicating through the network to other instance
> across nodes to replicate state. I was just wondering if this had been
> considered in the move to a more cloud-native Flink deployment?
>
> Cheers,
>
> Enrique
>
>
>
>
>
> --
> Sent from: http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/
>