By efficiency, I mean maximize throughput while minimize resources on both broker sides and consumer sides.
One example is if you have over 200 partitions on 10 brokers and you can start 5 consumer processes to consume data, if each one is single-thread and you do round-robin to distribute the load then each one will try to fetch from over 40 partitions one by one through 10 connections possibly(overall is 50), but if it's smart enough to group partitions by brokers, each process can have 2 separate threads(consuming from 2 different brokers concurrently). That seems a more optimal solution than another, right? On Tue, Nov 17, 2015 at 2:54 PM, Jason Gustafson <ja...@confluent.io> wrote: > Hi Siyuan, > > Your understanding about assign/subscribe is correct. We think of topic > subscription as enabling automatic assignment as opposed to doing manual > assignment through assign(). We don't currently them to be mixed. > > Can you elaborate on your findings with respect to using one thread per > broker? In what sense was it more efficient? Doing the same thing might be > tricky with the new consumer, but I think you could do it using > partitionsFor() to find the current partition leaders and assign() to set > the assignment in each thread. > > -Jason > > On Tue, Nov 17, 2015 at 10:25 AM, hsy...@gmail.com <hsy...@gmail.com> > wrote: > > > Thanks Guozhang, > > > > Maybe I should give a few words about what I'm going to achieve with new > > API > > > > Currently, I'm building a new kafka connector for Apache Apex( > > http://apex.incubator.apache.org/) using 0.9.0 API > > Apex support dynamic partition, so in the old version, We manage all the > > consumer partitions in either 1:1 strategy (each consumer process > consumes > > only from one kafka partition) or 1:n strategy (each consumer process > could > > consume from multiple kafka partitions, using round-robin to distribute) > > And we also have separate thread to monitor topic metadata change(leader > > broker change, new partition added, using internal API like ZkUtil etc) > > and do dynamic partition based on that(for example auto-reconnect to new > > leader broker, create new partition to consume from new kafka partition > at > > runtime). You can see High-level consumer doesn't work(It can only > balance > > between existing consumers unless you manually add new one) I'm thinking > > if the new consumer could be used to save some work we did before. > > > > I'm still confused with assign() and subscribe(). My understanding is if > > you use assign() only, the consumer becomes more like a simple consumer > > except if the leader broker changes it automatically reconnect to the new > > leader broker, is it correct? If you use subscribe() method only then > all > > the partitions will be distributed to running consumer process with same > " > > group.id" using "partition.assignment.strategy". Is it true? > > > > So I assume assign() and subscribe()(and group.id > > partition.assignment.strategy settings) can not be used together? > > > > Also in the old API we found one thread per broker is the most efficient > > way to consume data, for example, if one process consumes from p1, p2, p3 > > and p1,p2 are sitting on one broker b1, p3 is sitting on another one b2, > > the best thing is create 2 threads each thread use simple consumer API > and > > only consume from one broker. I'm thinking how do I use the new API to > do > > this. > > > > Thanks, > > Siyuan > > > > On Mon, Nov 16, 2015 at 4:43 PM, Guozhang Wang <wangg...@gmail.com> > wrote: > > > > > Hi Siyuan, > > > > > > 1) new consumer is single-threaded, it does not maintain any internal > > > threads as the old high-level consumer. > > > > > > 2) each consumer will only maintain one TCP connection with each > broker. > > > The only extra socket is the one with its coordinator. That is, if > there > > is > > > three brokers S1, S2, S3, and S1 is the coordinator for this consumer, > it > > > will maintain 4 sockets in total, 2 for S1 (one for fetching, one for > > > coordinating) and 1 for S2 and S3 (only for fetching). > > > > > > 3) Currently the connection is not closed by consumer, although the > > > underlying network client / selector will close idle ones after some > > > timeout. So in worst case it will only maintain N+1 sockets in total > for > > N > > > Kafka brokers at one time. > > > > > > Guozhang > > > > > > On Mon, Nov 16, 2015 at 4:22 PM, hsy...@gmail.com <hsy...@gmail.com> > > > wrote: > > > > > > > The new consumer API looks good. If I understand it correctly you can > > use > > > > it like simple consumer or high-level consumer. But I have couple > > > questions > > > > about it's internal implementation > > > > > > > > First of all does the consumer have any internal fetcher threads like > > > > high-level consumer? > > > > > > > > When you assign multiple TopicPartitions to a consumer, how many TCP > > > > connections it establish to the brokers. Is it same as number of > leader > > > > brokers that host those partitions or just number of TopicPartitions. > > If > > > > there is any leader broker change does it establish new > > connections/using > > > > existing connections to fetch the data? Can it continue consuming? > Also > > > is > > > > the connection kept until the consumer is closed? > > > > > > > > Thanks! > > > > > > > > Best, > > > > Siyuan > > > > > > > > > > > > > > > > -- > > > -- Guozhang > > > > > >
