Hi TengYao, Thanks for the KIP. Continuing on the point which Andrew mentioned as AS1.
As the memberId is now known to the client, and client might send the leave group heartbeat on shutdown prior to receiving the initial heartbeat response. If that's true then how do we guarantee that the 2 requests to join and leave will be processed in order, which could still leave stale members or throw unknown member id exceptions? Though the client side member id generation is helpful which will represent the same group perspective as from client and broker's end. But I think the major concern we want to solve here is Stale Partition Assignments which might still exist with the new approach. I am leaning towards the suggestion mentioned by Andrew where partition assignment triggers on subsequent heartbeat when client acknowledges the initial heartbeat, delayed partition assignment. Though on a separate note, I have a different question. What happens when there is an issue with the client which sends the initial heartbeat without memberId, then crashes and restarts? I think we must be re-triggering assignments and expiring members only after the heartbeat session timeout? If that's true then shall delayed partition assignment can help benefit us from this situation as well? Regards, Apoorv Mittal On Wed, Aug 14, 2024 at 12:51 PM David Jacot <dja...@confluent.io.invalid> wrote: > Hi Andrew, > > Personally, I don't like the lobby approach. It makes things more > complicated and it would require changing the records on the server too. > This is why I initially suggested the rejected alternative #2 which is > pretty close but also not perfect. > > I'd like to clarify one thing. The ConsumerGroupHeartbeat API already > supports generating the member id on the client so we don't need any > conditional logic on the client side. This is actually what we wanted to do > in the first place but the idea got pushed back by Magnus back then because > generating uuid from librdkafka required a new dependency. It turns out > that librdkafka has that dependency today. In retrospect, we should have > pushed back on this. Long story short, we can just do it. The proposal in > this KIP is to make the member id required in future versions. We could > also decide not to do it and to keep supporting both approaches. I would > also be fine with this. > > Best, > David > > On Wed, Aug 14, 2024 at 12:30 PM Andrew Schofield < > andrew_schofi...@live.com> > wrote: > > > Hi TengYao, > > Thanks for your response. I’ll have just one more try to persuade. > > I feel that I will need to follow the approach with KIP-932 when we’ve > > made a decision, so I do have more than a passing interest in this. > > > > A group member in the lobby is in the group, but it does not have any > > assignments. A member of a consumer group can have no assigned > > partitions (such as 5 CG members subscribed to a topic with 4 > partitions), > > so it’s a situation that consumer group members already expect. > > > > One of Kafka’s strengths is the way that we handle API versioning. > > But, there is a cost - the behaviour is different depending on the RPC > > version. KIP-848 is on the cusp of completion, but we’re already adding > > conditional logic for v0/v1 for ConsumerGroupHeartbeat. That’s a pity. > > Only a minor issue, but it’s unfortunate. > > > > Thanks, > > Andrew > > > > > On 14 Aug 2024, at 08:47, TengYao Chi <kiting...@gmail.com> wrote: > > > > > > Hello Andrew > > > Thank you for your thoughtful suggestions and getting the discussion > > going. > > > > > > To AS1: > > > In the current scenario where the server generates the UUID, if the > > client > > > shuts down before receiving the memberId generated by the GC > (regardless > > of > > > whether it’s a graceful shutdown or not), the GC will still have to > wait > > > for the heartbeat timeout because the client doesn’t know its memberId. > > > This KIP indeed cannot completely resolve the idempotency issue, but it > > can > > > better handle shutdown scenarios under normal circumstances because the > > > client always knows its memberId. Even if the client shuts down > > immediately > > > after the initial heartbeat, as long as it performs a graceful shutdown > > and > > > sends a leave heartbeat, the GC can manage the situation and remove the > > > member. Therefore, the goal of this KIP is to address the issue where > the > > > GC has to wait for the heartbeat timeout due to the client leaving > > without > > > knowing its memberId, which leads to reduced throughput and limited > > > scalability. > > > > > > The solution you suggest has also been proposed by David. The concern > > with > > > this approach is that it introduces additional complexity for > > > compatibility, as the new server would not immediately add the member > to > > > the group, while the old server would. This requires clients to > > > differentiate whether their memberId has been added to the group or > not, > > > which could result in unexpected logs. > > > > > > Best Regards, > > > TengYao > > > > > > Andrew Schofield <andrew_schofi...@live.com> 於 2024年8月14日 週三 > 上午12:29寫道： > > > > > >> Hi TengYao, > > >> Thanks for the KIP. I wonder if there’s a different way to close what > > >> is quite a small window. > > >> > > >> AS1: It is true that the initial heartbeat is not idempotent, but this > > >> remains > > >> true with this KIP. It’s just differently not idempotent. If the > client > > >> makes its > > >> own member ID, sends a request and dies, the GC will still have added > > >> the member to the group and it will hang around until the session > > expires. > > >> > > >> I wonder if the GC could still generate the member ID in response to > the > > >> first > > >> heartbeat, and put the member in a special PENDING state with no > > >> assignments until the client sends the next heartbeat, thus confirming > > it > > >> has received the member ID. This would not be a protocol change at > all, > > >> just > > >> a change to the GC to keep a new member in the lobby until it’s > > comfirmed > > >> it knows its member ID. > > >> > > >> > > >> Thanks, > > >> Andrew > > >> > > >>> On 13 Aug 2024, at 15:59, TengYao Chi <kiting...@gmail.com> wrote: > > >>> > > >>> Hi Chia-Ping, > > >>> > > >>> Thanks for review and suggestions. > > >>> I have updated the content of KIP accordingly. > > >>> Please take a look. > > >>> > > >>> Best regards, > > >>> TengYao > > >>> > > >>> Chia-Ping Tsai <chia7...@apache.org> 於 2024年8月13日 週二 下午9:45寫道： > > >>> > > >>>> hi TengYao > > >>>> > > >>>> thanks for this KIP. > > >>>> > > >>>> 1) could you please describe the before/after behavior in the > > "Proposed > > >>>> Changes" section? IIRC, current RPC allows HB having member id > > >> generated by > > >>>> client, right? If HB has no member ID, server will generate one and > > then > > >>>> return. The new behavior will enforce HB "must" have member ID. > > >>>> > > >>>> 2) could you please write the version number explicitly in the KIP > > >>>> > > >>>> 3) how new client code handle the old HB? Does it always generate > > member > > >>>> ID on client-side even though that is not restricted? > > >>>> > > >>>> Best, > > >>>> Chia-Ping > > >>>> > > >>>> On 2024/08/13 06:20:42 TengYao Chi wrote: > > >>>>> Hello everyone, > > >>>>> > > >>>>> I would like to start a discussion thread on KIP-1082, which > proposes > > >>>>> enabling id generation for clients over the ConsumerGroupHeartbeat > > RPC. > > >>>>> > > >>>>> Here is the KIP Link: KIP-1082 > > >>>>> < > > >>>> > > >> > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-1082%3A+Enable+ID+Generation+for+Clients+over+the+ConsumerGroupHeartbeat+RPC > > >>>>> > > >>>>> Please take a look and let me know what you think, and I would > > >> appreciate > > >>>>> any suggestions and feedback. > > >>>>> > > >>>>> Best regards, > > >>>>> TengYao > > >>>>> > > >>>> > > >> > > >> > > > > >
