Yes, but the ConsumerRebalanceListener is optional and by the default KafkaConsumer uses a NoOpConsumerRebalanceListener if no one is provided.
I think the seek() is already done internally when a consumer joins or quits the group. I'm not sure this line is actually needed. 2016-04-18 15:31 GMT+02:00 Kamal C <kamaltar...@gmail.com>: > When a new consumer joins to the group, it should start to read data > from where the other consumer left. > > --Kamal > > On Mon, Apr 18, 2016 at 6:58 PM, Florian Hussonnois <fhussonn...@gmail.com > > > wrote: > > > Thank you very much, the example is really helpful. > > > > My last question is : Why is it necessay to seek the consumer offsets > into > > the onPartitionsAssigned method ? > > > > > > > https://github.com/omkreddy/kafka-examples/blob/master/consumer/src/main/java/kafka/examples/consumer/advanced/AdvancedConsumer.java#L120 > > > > 2016-04-15 15:06 GMT+02:00 Kamal C <kamaltar...@gmail.com>: > > > > > Hi Florian, > > > > > > This may be helpful > > > > > > > > > https://github.com/omkreddy/kafka-examples/blob/master/consumer/src/main/java/kafka/examples/consumer/advanced/AdvancedConsumer.java > > > > > > --Kamal > > > > > > On Fri, Apr 15, 2016 at 2:57 AM, Jason Gustafson <ja...@confluent.io> > > > wrote: > > > > > > > Hi Florian, > > > > > > > > It's actually OK if processing takes longer than the heartbeat > > interval, > > > > but it does need to finish before the session timeout expires or the > > > > consumer will be kicked out of the group (which typically is revealed > > by > > > > commit failures). If the problem is just that the consumer is > handling > > > too > > > > many messages at once, then Kafka 0.10 has an option to tune the > number > > > of > > > > messages returned from poll() (max.poll.records), which may be > helpful. > > > We > > > > also have a pause/resume API which allows you to call poll() without > > > > consuming any data. That's the best option at the moment for 0.9 > > > consumers. > > > > > > > > For what it's worth, we've considered several times adding a > > heartbeat() > > > > API, but the challenge is figuring out how to handle rebalancing. > > > > Underneath the covers, we use heartbeats to find out when the group > is > > > > rebalancing, so a heartbeat() option would probably have to return a > > flag > > > > indicating whether a rebalance was needed. If the group has begun > > > > rebalancing, then you would need to call poll() before the expiration > > of > > > > the session timeout so that the consumer can join the rebalance. > > > > Alternatively, we could let heartbeat() complete the rebalance > itself, > > > but > > > > then you'd have to be prepared to abort processing from the rebalance > > > > callback. That's not really different from calling poll() after > pausing > > > > partitions though. The main problem in any case is that once a > > rebalance > > > > begins, you have the duration of the session timeout to stop > processing > > > and > > > > join the rebalance. We're seeing this problem pop up pretty much > > > everywhere > > > > that the consumer is used, so we're trying to think of some better > > > options > > > > to handle it. > > > > > > > > Thanks, > > > > Jason > > > > > > > > > > > > On Thu, Apr 14, 2016 at 12:32 PM, Florian Hussonnois < > > > > fhussonn...@gmail.com> > > > > wrote: > > > > > > > > > Hi everyone, > > > > > > > > > > I have a use case where a message can take longer than ' > > > > > heartbeat.interval.ms' to be processed by my application. As I > > > > understand > > > > > the heartbeats of consumer are done while the poll method is > invoked. > > > > > > > > > > I would like to instantiate a worker thread to process the messages > > > but I > > > > > need to wait for the messages completion before polling again. > > > > > > > > > > Is there a way to force the consumer to make an heartbeat without > > > polling > > > > > new messages ? > > > > > > > > > > Thanks, > > > > > > > > > > -- > > > > > Florian HUSSONNOIS > > > > > > > > > > > > > > > > > > > > -- > > Florian HUSSONNOIS > > > -- Florian HUSSONNOIS
