Thanks for the KIP! LGTM, +1 (non-binding).
On Wed, Sep 11, 2019 at 3:23 AM Matthias J. Sax <matth...@confluent.io> wrote: > I don't have a strong preference. So I am also fine to deprecate the > existing methods. Let's see what Jason thinks. > > Can you update the KIP to reflect the semantics of the return `Map` (ie, > does only contain entries for partitions with committed offsets, and > does not contain `null` values)? > > > +1 (binding) > > -Matthias > > > > > On 9/10/19 11:53 AM, Guozhang Wang wrote: > > Hi Jason / Matthias, > > > > I understand your concerns now. Just to clarify my main motivation on > > deprecating the old APIs is not only for aesthetics (confess I personally > > do have preference on succinct APIs), but to urge people to use the > batched > > API for better latency when possible --- as stated in the KIP, my > > observation is that in practice most callers are actually going to get > > committed offsets for more than one partitions, and without deprecating > the > > old APIs it would be hard for them to realize that the new API does have > a > > benefit in performance. > > > > This is different from some of the existing APIs though -- e.g., Matthias > > mentioned about seek / seekToBeginning / seekToEnd, where only seekToXX > > have plurals and seek only have singulars. We can, of course, make > seekToXX > > with plurals as well just like commitSync/Async, but since seeks are > > non-blocking APIs (they rely on the follow-up polling call to talk to the > > brokers) either calling it multiple times with one partition each v.s. > > calling it one time with a plural makes little difference (still, I'd > argue > > that today we do not have a same-named function overloaded with both > types > > :P) On the other hand, committed, commitSync, offsetsForTimes etc > blocking > > calls are all in the form of plurals except > > > > * committed > > * position > > * partitionsFor > > > > My rationale was that 1) for consecutive calls of #position, mostly it > > would only require a single round-trip to brokers since we are trying to > > refresh fetching positions for all partitions anyways, and 2) for > > #partitionsFor, theoretically we could also consider to ask for multiple > > topics in one call since each blocking call potentially incurs one round > > trip, but I did not include it in the scope of this KIP only because I > have > > not observed too many usage patterns that are commonly calling it > > consecutively for multiple topics. At the moment, what I truly want to > > "improve" on is the committed calls, as in many cases I've seen it being > > called consecutively for multiple topic-partitions. > > > > Therefore, I'm still more inclined to deprecate the old APIs so that we > can > > enforce people to discover the new batching APIs for efficiency in this > > KIP. But if you feel that this compatibility is very crucial to maintain > I > > could be convinced. > > > > > > Guozhang > > > > On Tue, Sep 10, 2019 at 10:18 AM Matthias J. Sax <matth...@confluent.io> > > wrote: > > > >> Thanks for the KIP Guozhang. > >> > >>> Another reason is that other functions of KafkaConsumers do not have > >> those > >>> overloaded functions to be consistent > >> > >> I tend to agree with Jason about keeping the existing methods. Your > >> argument does not seem to hold. I just checked the `Consumer` API, and > >> it's mix of overloads: > >> > >> Methods only talking `Collections` > >> > >> > >> > subscribe/assign/commitSync/commitAsyn/pause/resume/offsetsForTimes/beginningOffsets/endOffsets > >> > >> Method with overload taking `Collections` or as single value: > >> > >> seek/seekToBeginning/seekToEnd > >> > >> (those are strictly different methods, but they are semantically > related) > >> > >> Only talking single value: > >> > >> position/committed/partitionsFor > >> > >> > >> While you are strictly speaking correct, that there is no method with an > >> overload for `Collection` and single value, the API mix seems to suggest > >> that it might actually be worth to have corresponding overloads for all > >> methods instead of sticking to `Collections` only. > >> > >> > >> > >> About the return map: I agree that not containing any entry in the map > >> is better. It's not only consistent with other APIs but it also avoids > >> potential NPEs. > >> > >> > >> > >> -Matthias > >> > >> > >> On 9/10/19 10:04 AM, Jason Gustafson wrote: > >>>> I feel it not worth making committed to have both plurals and > >> singulars. > >>> > >>> Not sure I agree. If we had started with these new APIs from the > >> beginning, > >>> that may have been better, but we already have exposed the singular > APIs > >>> and users are depending on them. Not sure it's worth breaking > >> compatibility > >>> just for aesthetics. > >>> > >>> -Jason > >>> > >>> On Tue, Sep 10, 2019 at 9:41 AM Guozhang Wang <wangg...@gmail.com> > >> wrote: > >>> > >>>> Thanks Jason! > >>>> > >>>> On Tue, Sep 10, 2019 at 9:07 AM Jason Gustafson <ja...@confluent.io> > >>>> wrote: > >>>> > >>>>> Hi Guozhang, > >>>>> > >>>>> I think the motivation for the new API makes sense. I've wanted > >> something > >>>>> like this in the past. That said, do you think there is a substantial > >>>>> benefit from deprecating the old API? I can still see it being > >> convenient > >>>>> in some cases and it's no real cost to maintain. > >>>>> > >>>>> > >>>> That's a good question. > >>>> > >>>> Personally I would like to keep a succinct set of APIs out of the box > >> and > >>>> let users who want more syntax sugars to add themselves as extended > >> classes > >>>> for example (KafkaConsumer is not a final class). > >>>> Another reason is that other functions of KafkaConsumers do not have > >> those > >>>> overloaded functions to be consistent, e.g. we do not have a > >>>> subscribe(single-topic), pause/resume(single-topic-partition) or > >>>> seekToBeginning(single-topic-partition). I feel it not worth making > >>>> committed to have both plurals and singulars. > >>>> > >>>> WDYT? > >>>> > >>>> > >>>>> Also, just a minor detail. If the partition has no committed offset, > >> will > >>>>> it be present in the map with a null value? > >>>>> > >>>>> I looked into the admin client's listConsumerGroupOffsets call when > >>>> creating the KIP, and to be consistent with that API my intention is > to > >> NOT > >>>> include the entry if a topic-partition does not have committed > offsets. > >>>> That said, if we feel returning an entry with null value is better for > >>>> programmability I can also do that (and will update wiki page to > >> clarify as > >>>> well). LMK. > >>>> > >>>> > >>>>> -Jason > >>>>> > >>>>> On Tue, Sep 10, 2019 at 6:09 AM Mickael Maison < > >> mickael.mai...@gmail.com > >>>>> > >>>>> wrote: > >>>>> > >>>>>> +1 (non-binding), thanks Guozhang > >>>>>> > >>>>>> On Tue, Sep 10, 2019 at 1:14 AM Boyang Chen < > >>>> reluctanthero...@gmail.com> > >>>>>> wrote: > >>>>>>> > >>>>>>> Hey Guozhang, > >>>>>>> > >>>>>>> LGTM, +1 (non-binding) > >>>>>>> > >>>>>>> On Mon, Sep 9, 2019 at 5:07 PM Guozhang Wang <wangg...@gmail.com> > >>>>> wrote: > >>>>>>> > >>>>>>>> Hello folks, > >>>>>>>> > >>>>>>>> I've created a new KIP allowing consumer.committed to take a set > of > >>>>>>>> partitions instead of just one partition to allow batching effects > >>>> of > >>>>>> such > >>>>>>>> requests (the protocol already allows us to send multiple > >>>> partitions > >>>>>> in one > >>>>>>>> round-trip): > >>>>>>>> > >>>>>>>> > >>>>>>>> > >>>>>> > >>>>> > >>>> > >> > https://cwiki.apache.org/confluence/display/KAFKA/KIP-520%3A+Add+overloaded+Consumer%23committed+for+batching+partitions > >>>>>>>> > >>>>>>>> Since it is a pretty straight-forward KIP, I'm starting the VOTE > >>>> for > >>>>>> this > >>>>>>>> KIP directly. If there are any suggestions about this proposal, > >>>>> please > >>>>>> feel > >>>>>>>> free to share them in this thread. Thank you! > >>>>>>>> > >>>>>>>> > >>>>>>>> -- Guozhang > >>>>>>>> > >>>>>> > >>>>> > >>>> > >>>> > >>>> -- > >>>> -- Guozhang > >>>> > >>> > >> > >> > > > >
