Thanks Ismael and Jason, I filed a separate KIP to solve the problems identified through this discussion. I also incorporated Jason's comments in that document: https://cwiki.apache.org/confluence/display/KAFKA/KIP-192+%3A+Provide+cleaner+semantics+when+idempotence+is+enabled
Please have a look, Apurva On Tue, Aug 29, 2017 at 3:28 AM, Ismael Juma <ism...@juma.me.uk> wrote: > Thanks for the proposals. I think they make sense and I also agree with > Jason's suggestions. Also, it would be good to include the updated > ProduceRequest/Response schema in the KIP. > > Ismael > > On Tue, Aug 22, 2017 at 11:42 PM, Jason Gustafson <ja...@confluent.io> > wrote: > > > Thanks Apurva, > > > > On compatibility: I think the proposal makes sense. It's a pity that we > > can't support idempotence for 0.11.0.0 brokers in the "safe" mode even if > > it is supported by the broker. I can already imagine users complaining > > about this, but I guess it's the consequence of missing the impact of > that > > validation check and not thinking through the ultimate goal of enabling > > idempotence by default. A couple minor comments: > > > > 1. Instead of "safe," Ismael suggested "requested" as an alternative. > That > > seems to suggest more clearly that idempotence will only be used when the > > broker supports it. > > 2. Should we deprecate the "true" and "false" options? It's a little > weird > > long term to support them in addition to the descriptive names. > > > > On the OutOfOrderSequence proposal: high-level, the design makes sense. A > > couple questions: > > > > 1. With this proposal, OutOfOrderSequence means that we must have a last > > produced offset. Is the idea to expose that in the > > OutOfOrderSequenceException so that users know which data was lost? > > 2. Previously we discussed duplicate handling. Currently we raise > > OutOfOrderSequence if we happen to get a sequence number which is earlier > > than the sequence numbers we have cached. Alternatively, you suggested we > > can return a separate DuplicateError for this case, which clients can > > ignore if they do not care about the offset. I think it might make sense > to > > include that here so that the OutOfOrderSequence error is unambiguous. > > > > Finally, do you plan to roll these proposals into the current KIP or do > > them separately? Probably makes sense to combine them since they both > > require a bump to the ProduceRequest. > > > > Thanks, > > Jason > > > > > > > > On Fri, Aug 18, 2017 at 5:18 PM, Apurva Mehta <apu...@confluent.io> > wrote: > > > > > Thanks Jason and Ismael. > > > > > > The message format problem is an acute one: if we enable idempotence by > > > default, the UnsupportedVersionException when writing to topics with > the > > > older message format would mean that our prescribed upgrade steps would > > not > > > work. I have detailed the problems and the solutions on this page > (linked > > > to from the wiki): > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/ > > > Kafka+Exactly+Once+-+Dealing+with+older+message+formats+ > > > when+idempotence+is+enabled > > > > > > It is worth discussing the solution to the problem proposed there. If > it > > is > > > conceptually sound, it doesnt' seem too hard to implement. > > > > > > As far as the other problem of the spurious OutOfOrderSequence > problem, I > > > have documented a proposed solution here: > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/ > > > Kafka+Exactly+Once+-+Solving+the+problem+of+spurious+ > > > OutOfOrderSequence+errors > > > > > > This solution is a bit more involved in terms of effort. > > > > > > I think we cannot make the idempotent producer the default unless we > > solve > > > the message format compatibility problem. I would also prefer to solve > > the > > > second problem before making idempotence the default. > > > > > > I would be interested to hear everyone's thoughts on the two solutions > > > proposed above. > > > > > > Thanks, > > > Apurva > > > > > > On Fri, Aug 18, 2017 at 9:24 AM, Jason Gustafson <ja...@confluent.io> > > > wrote: > > > > > > > > > > > > > so this change will break client backward compatibility while > > > connecting > > > > > to 0.10.X brokers. > > > > > users need to change producer default settings while connecting > > older > > > > > brokers. > > > > > > > > > > > > At the moment, I think the answer is yes. The old broker will not > > support > > > > the InitProducerId request, so the producer will immediately fail. > > > Similar > > > > to the handling of old message formats mentioned above, we probably > > need > > > to > > > > change this so that we just revert to old producer semantics if the > > > broker > > > > can't support idempotence. > > > > > > > > -Jason > > > > > > > > > > > > On Fri, Aug 18, 2017 at 8:48 AM, Manikumar < > manikumar.re...@gmail.com> > > > > wrote: > > > > > > > > > > > > > > > > 3. The message format requirement is a good point. This should be > > > > > mentioned > > > > > > in the compatibility section. Users who are still using the old > > > message > > > > > > format will get an error after the upgrade, right? > > > > > > > > > > > > > > > > so this change will break client backward compatibility while > > > connecting > > > > > to 0.10.X brokers. > > > > > users need to change producer default settings while connecting > > older > > > > > brokers. > > > > > > > > > > > > > > >
