Hi Matthias,

Thank you for the feedback. It’s an interesting idea, and I am more than happy 
to explore it.
The idea of having a dedicated RPC for pause/resume came from the original 
email thread [1] when I proposed this functionality.
I would be curious to hear David’s and Lucas’s thoughts on the subject as well.

Best,
Levani

[1] - https://lists.apache.org/thread/d45hpm86fxvc6f81svmzjo8hp8sjft5y

> On 22. Jul 2026, at 00:22, Matthias J. Sax <[email protected]> wrote:
> 
> Hello Levani,
> 
> I am very excited about this KIP -- I fully agree on the motivation, but I 
> think we need to do this very differently. I am glad so see that the KIP 
> already mentioned "streams" groups, but for this case we should take a 
> potential client-side state store reset into account. -- And yes, I think we 
> should also include "share" groups. Building out one unified solution for all 
> group types will provide the best user experience.
> 
> On a high-level, I don't think we should add pause/resume or similar RPCs as 
> currently proposed, but rather a new ResetOffsetsRPC, and this RPC should 
> encode all information the GC needs to orchestrate the reset.
> 
> I would also not necessarily revoke any partitions, but change the existing 
> heartbeats responses to include information about an ongoing offset reset, 
> and clients would get updated accordingly to do the right thing during an 
> ongoing reset. We need to define what "the right thing" would be. I believe 
> this would be a much better end-to-end design.
> 
> Yes, this description is high level, but I wanted to paint a direction 
> picture to see what everybody thinks about it.
> 
> We will also need to figure out how we keep track of "reset progress" and 
> many corner cases like membership changes (and failures) during an ongoing 
> reset, and how to decide when a reset was completed etc. For example, there 
> is a corner case for Kafka Streams w/ client side state store reset, which 
> could lead to bad behavior: a client with local state might drop out of the 
> group, a reset happens afterward, and the client rejoins later. For this 
> case, we need to still ensure that the client would reset it's local state 
> correctly, even if it was not part of the group during the reset 
> orchestration.
> 
> Curious to hear what you think about this direction.
> 
> 
> -Matthias
> 
> 
> 
> On 3/23/26 12:15 AM, Levani Kokhreidze wrote:
>> Hi,
>> Gentle bump on this KIP in case anyone else wants to share their feedback.
>> Best,
>> Levani
>>> On 16. Mar 2026, at 08:32, Sushant Mahajan <[email protected]> wrote:
>>> 
>>> Hi,
>>> Thanks for the reply.
>>> 
>>> On Tue, 10 Mar 2026, 02:58 Levani Kokhreidze, <[email protected]>
>>> wrote:
>>> 
>>>> Hi Sushant,
>>>> 
>>>> Thanks for the comments.
>>>> 
>>>> RE sm01:
>>>> I haven’t researched the share groups support as I wasn’t sure if it would
>>>> be even useful for the share groups to support this feature.
>>>> But usefulness of this feature was mostly discussed around the new
>>>> consumer group protocol and streams protocol [1].
>>>> Share groups support can be a followup KIP as well.
>>>> 
>>>> RE sm02:
>>>> For the new consumer group protocol we will be utilising the heartbeat API
>>>> to update the assignment based on the paused partitions state.
>>>> As for heartbeat API for the classic group, I do not think it gives any
>>>> hint around target assignment, that will be still up to the client side
>>>> assignor to handle.
>>>> Since classic groups will eventually be deprecated, I believe it makes
>>>> more sense to focus on the new consumer group protocol.
>>>> 
>>>> I hope this answers your questions.
>>>> 
>>>> Thanks,
>>>> Levani
>>>> 
>>>> [1] - https://lists.apache.org/thread/d45hpm86fxvc6f81svmzjo8hp8sjft5y
>>>> 
>>>>> On 6. Mar 2026, at 13:38, Sushant Mahajan <[email protected]> wrote:
>>>>> 
>>>>> Hi Levani,
>>>>> Thanks for the KIP.
>>>>> 
>>>>> sm01: Is there any reason for not including share groups as supported
>>>> group
>>>>> types?
>>>>> 
>>>>> sm02: Can heartbeat API be leveraged to inform the consumers about the
>>>>> paused partitions? That can extend to all group types and prevent future
>>>>> fetch calls.
>>>>> 
>>>>> Regards,
>>>>> Sushant Mahajan
>>>>> 
>>>>> 
>>>>> On Wed, 25 Feb 2026, 19:38 Sean Quah via dev, <[email protected]>
>>>> wrote:
>>>>> 
>>>>>> Hi Levani,
>>>>>> 
>>>>>> Thank you for the KIP!
>>>>>> 
>>>>>> sq01: Could you outline how the admin client will reset offsets for an
>>>>>> active group? After pausing a partition, we have to wait until the
>>>>>> partition has been released by its consumer. Will we poll DescribeGroup
>>>> or
>>>>>> retry OffsetCommits until it succeeds or do something else?
>>>>>> 
>>>>>> Thanks,
>>>>>> Sean
>>>>>> 
>>>>>> On Mon, Feb 23, 2026 at 8:53 PM Levani Kokhreidze <
>>>> [email protected]>
>>>>>> wrote:
>>>>>> 
>>>>>>> Hi all,
>>>>>>> 
>>>>>>> I’d like to start discussion on KIP-1249 [1] which aims to provide a
>>>>>>> better offset reset functionality for the new consumer group rebalance
>>>>>>> protocols.
>>>>>>> 
>>>>>>> Looking forward to hear community’s thoughts on this.
>>>>>>> 
>>>>>>> Best,
>>>>>>> Levani
>>>>>>> 
>>>>>>> [1] -
>>>>>>> 
>>>>>> 
>>>> https://cwiki.apache.org/confluence/display/KAFKA/KIP-1249%3A+Better+offset+reset+for+the+new+consumer+group+rebalance+protocol
>>>>>>> 
>>>>>>> 
>>>>>> 
>>>> 
>>>> 
> 

Reply via email to