Hi Colin, Is it really true that "the period when the offset is unavailable should be brief"? I'm thinking about a producer with acks=1, so the old leader returns the ProduceResponse immediately and then is replaced before it can sent a FetchResponse to any followers. The new leader is then waiting for more messages from producers in order for its high watermark to increase (because it's log doesn't have the original messages in, so its HW can't catch up with this). This wait could be be arbitrarily long.
I rather suspect this isn't a problem really and that I misunderstand the precise details of the protocol, but it would be beneficial to me to discover my misconceptions. Thanks, Tom On 5 October 2017 at 19:23, Colin McCabe <cmcc...@apache.org> wrote: > Hi all, > > I created a KIP for discussion about fixing a corner case in > ListOffsetsResponse. Check it out at: > https://cwiki.apache.org/confluence/display/KAFKA/KIP- > 207%3A+Offsets+returned+by+ListOffsetsResponse+should+be+ > monotonically+increasing+even+during+a+partition+leader+change > > cheers, > Colin >
