Regarding 2), sounds good, I saw UNKNOWN_PRODUCER_ID is properly handled today in produce / add-partitions-to-txn / add-offsets-to-txn / end-txn responses, so that should be well covered.
Could you reflect this in the wiki page that the broker should be responsible for using different error codes given client request versions as well? Guozhang On Mon, Apr 6, 2020 at 9:20 AM Boyang Chen <reluctanthero...@gmail.com> wrote: > Thanks Guozhang for the review! > > On Sun, Apr 5, 2020 at 5:47 PM Guozhang Wang <wangg...@gmail.com> wrote: > > > Hello Boyang, > > > > Thank you for the proposed KIP. Just some minor comments below: > > > > 1. Could you also describe which producer APIs could potentially throw > the > > new TransactionTimedOutException, and also how should callers handle them > > differently (i.e. just to make your description more concrete as > javadocs). > > > > Good point, I will add example java doc changes. > > > > 2. It's straight-forward if client is on newer version while broker's on > > older version; however If the client is on older version while broker's > on > > newer version, today would the internal module of producers treat it as a > > general fatal error or not? If not, should the broker set a different > error > > code upon detecting older request versions? > > > > That's a good suggestion, my understanding is that the prerequisite of > this change is the new KIP-360 API which is going out with 2.5, > so we could just return UNKNOWN_PRODUCER_ID instead of PRODUCER_FENCED as > it could be interpreted as abortable error > in 2.5 producer and retry. Producers older than 2.5 will not be covered. > WDYT? > > > > > Guozhang > > > > On Thu, Apr 2, 2020 at 1:40 PM Boyang Chen <reluctanthero...@gmail.com> > > wrote: > > > > > Hey there, > > > > > > I would like to start discussion for KIP-588: > > > > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-588%3A+Allow+producers+to+recover+gracefully+from+transaction+timeouts > > > > > > which aims to improve Producer resilience to transaction timeout due to > > > transient system gaps. > > > > > > Best, > > > Boyang > > > > > > > > > -- > > -- Guozhang > > > -- -- Guozhang
