The main scenarios for us are upgrading NiFi and updating (and consequently
restarting) the Kubernetes workers for security updates. Both of these make
manual offloading basically impossible - the StatefulSet handles pod
restarts by itself, and the Kubernetes control plane drains and restarts
worker machines automatically.

Regards,
Alexis.

On Sat, 17 Aug 2024, 17:40 Joe Witt, <[email protected]> wrote:

> Alexis
>
> Yep we totally understand.
>
> What scenarios are you trying to plan/design for that require restarts?
>
> Thanks
>
> On Sat, Aug 17, 2024 at 8:24 AM Alexis Sarda-Espinosa <
> [email protected]> wrote:
>
>> Hi Joe,
>>
>> I agree the system can recover after the restart, but we want to deliver
>> data with as little latency as possible, so if FlowFiles stay queued during
>> the restart time window, by the time the node is up and running again the
>> data will be very late when time it finally leaves the NiFi cluster.
>>
>> Regards,
>> Alexis.
>>
>> On Sat, 17 Aug 2024, 16:59 Joe Witt, <[email protected]> wrote:
>>
>>> Alexis
>>>
>>>
>>> If you are simply restarting nodes you dont need to ensure everything is
>>> processed first. It will recover.
>>>
>>> If you want to scale down by removing a node there is an api to invoke
>>> offloading so you could scale down to a single node even if needed.
>>>
>>> Thanks
>>>
>>> On Sat, Aug 17, 2024 at 2:02 AM Alexis Sarda-Espinosa <
>>> [email protected]> wrote:
>>>
>>>> Hello,
>>>>
>>>> We use NiFi in cluster mode in Kubernetes via a StatefulSet. Sometimes,
>>>> we need to restart the cluster, so the StatefulSet restarts one pod at a
>>>> time. In this scenario, will a given node attempt to process all
>>>> "in-flight" FlowFiles to drain the queues as much as possible before
>>>> shutting down?
>>>>
>>>> I know it's probably impossible to have a completely graceful offload
>>>> in these cases, but since the process still receives a termination signal
>>>> and has some time before it must stop, it could try to prevent FlowFiles
>>>> from staying in the system for too long, but I don't know if there are any
>>>> mechanisms to handle this in NiFi.
>>>>
>>>> Regards,
>>>> Alexis.
>>>>
>>>

Reply via email to