Hi Luke, Thanks for the details... so from explanation above, it seems that both of these scenarios, I won't be able to avoid duplicates processing, which is main purpose that I was looking to achieve
scenario 1: consumer shuts down, and doesn't commit offsets of already polled and processed batch of records (since auto.commit enabled which would commit on next poll which won't occur when closing consumer). This would give rise duplicate processing of that batch when partition is rebalanced to other consumer pod scenario 2: CooperativeStickyAssignor keeps working on partition before rebalancing which means again same thing i.e. consumer1 has polled and processed some records which are not yet committed before rebalancing and when partition moves over to next consumer, it can process those records again On Wed, Sep 21, 2022 at 7:32 AM Luke Chen <show...@gmail.com> wrote: > Hi > > 1. I was under impression, from documentation, that close method waits for > 30 seconds to complete processing of any in-flight events and then commits > offsets of last poll. Isn't that true? what does timeout of 30 seconds > mean? > > -> 30 seconds timeout is to have a buffer for graceful closing, ex: commit > offsets, leave groups,... > It won't complete processing any in-flight "fetch" events during closing. > > 2. How does CoperativeStickyAssignor solve my problem when partitions move > out to newly added consumer pod. i.e. consumer1 has polled 100 records from > partition1 and is midway processing those i.e. 50 completed, 50 remaining > and new consumer is added so partition1 has to move to new consumer2. Since > auto.commit is enabled, offsets of all 100 polled records will be committed > only during next poll. 2. How does CoperativeStickyAssignor solve my > problem when partitions move > out to newly added consumer pod. i.e. consumer1 has polled 100 records from > partition1 and is midway processing those i.e. 50 completed, 50 remaining > and new consumer is added so partition1 has to move to new consumer2. Since > auto.commit is enabled, offsets of all 100 polled records will be committed > only during next poll. > > -> You're right about the process. > > So how does CooperativeStickyAssignore help here to > wait for process of 100 records and commit their offsets before moving the > partition to new consumer? Looks like i am missing something > Looks like i am missing something > > -> CooperativeStickyAssignore does the same thing to it, except it will > keep all the partitions "during rebalancing". > So, the issue is: > In eagar protocol (ex: RangeAssignor) > consumer prepare rebalancing -> commit offsets -> revoke all owned > partitions -> rebalancing -> received new assignment -> start fetch data > In cooperative protocol (ex: CooperativeStickyAssignore) > consumer prepare rebalancing -> commit offsets (but no revoke) -> > rebalancing -> received new assignment -> revoke partitions not owned > anymore > > So you can see, in cooperative protocol, since it didn't revoke any > partition before rebalancing, it might fetch more data after offset > commits. > > Hope that's clear > Luke > > On Tue, Sep 20, 2022 at 9:36 PM Pushkar Deole <pdeole2...@gmail.com> > wrote: > > > Thanks Luke.. > > > > 1. I was under impression, from documentation, that close method waits > for > > 30 seconds to complete processing of any in-flight events and then > commits > > offsets of last poll. Isn't that true? what does timeout of 30 seconds > > mean? > > > > 2. How does CoperativeStickyAssignor solve my problem when partitions > move > > out to newly added consumer pod. i.e. consumer1 has polled 100 records > from > > partition1 and is midway processing those i.e. 50 completed, 50 remaining > > and new consumer is added so partition1 has to move to new consumer2. > Since > > auto.commit is enabled, offsets of all 100 polled records will be > committed > > only during next poll. So how does CooperativeStickyAssignore help here > to > > wait for process of 100 records and commit their offsets before moving > the > > partition to new consumer? Looks like i am missing something > > > > On Fri, Sep 16, 2022 at 7:59 AM Luke Chen <show...@gmail.com> wrote: > > > > > Hi Pushkar, > > > > > > Here's the answer to your questions: > > > > > > > 1. During scale-down operation, I am adding a shutdown hook to the > Java > > > Runtime, and calling close on the consumer. As per kafka docs, close > > > provides 30 sec to commit current offsets if auto.commit is enabled: > so, > > i > > > assume that it will process the current batch of polled records within > 30 > > > sec timeout before committing offsets and then close the consumer. Is > my > > > understanding correct? > > > > > > No, close() method is only doing some cleanup and offset commit if > > needed. > > > It won't care if the polled records are processed or not. > > > So, to be clear, the 30 seconds is for consumer to do: > > > (1) commit offset if auto.commit is enabled (2) leave consumer group > (3) > > > other cleanup > > > > > > > 2. During scale out operation, new pod (consumer) will be added to > the > > > consumer group, so partitions of existing consumers will be rebalanced > to > > > new consumer. In this case, I want to ensure that the current batch of > > > records polled and being processed by the consumer is processed and > > offsets > > > are committed before partition rebalance happens to new consumer. > > > How can I ensure this with auto-commit enabled? > > > > > > It depends on which version of Kafka you're running, and which > > > `partition.assignment.strategy` you are setting. > > > In Kafka v3.2.1, we found a bug that it'll have chance to process > > duplicate > > > records during rebalance: KAFKA-14196 > > > <https://issues.apache.org/jira/browse/KAFKA-14196> > > > So, assuming you're using default `partition.assignment.strategy` > > setting, > > > and not in v3.2.1, we can ensure it will not have duplicated > consumption. > > > If you set the `partition.assignment.strategy` to > > > cooperativeStickyAssignor, there's a bug that we're still working on: > > > KAFKA-14224 <https://issues.apache.org/jira/browse/KAFKA-14224> > > > > > > Thank you. > > > Luke > > > > > > On Wed, Sep 14, 2022 at 3:09 PM Pushkar Deole <pdeole2...@gmail.com> > > > wrote: > > > > > > > Hi All, > > > > > > > > I am hosting kafka consumers inside microservice hosted as kubernetes > > > pods, > > > > 3 consumers in a consumer group. > > > > There is a requirement to add auto-scaling where there will be a > single > > > pod > > > > which will be auto-scaled out or scaled-in based on the load on > > > > microservice. > > > > So, 1 pod can be scaled out to 2 or 3 pods, and similarly 3 pods can > be > > > > scaled down to 2 or 1 pod. > > > > > > > > Currently, I am using enabled.auto.commit set to 'true' in the > > consumers > > > > and during scale out or scale-in, i want to commit offset of polled > and > > > > processed records so duplicates won't occur. > > > > I have narrowed the problem to 2 scenarios: > > > > > > > > 1. During scale-down operation, I am adding a shutdown hook to the > Java > > > > Runtime, and calling close on the consumer. As per kafka docs, close > > > > provides 30 sec to commit current offsets if auto.commit is enabled: > > so, > > > i > > > > assume that it will process the current batch of polled records > within > > 30 > > > > sec timeout before committing offsets and then close the consumer. Is > > my > > > > understanding correct? > > > > > > > > public void close() > > > > > > > > Close the consumer, waiting for up to the default timeout of 30 > seconds > > > for > > > > any needed cleanup. If auto-commit is enabled, this will commit the > > > current > > > > offsets if possible within the default timeout. See close(Duration) > > > > < > > > > > > > > > > https://kafka.apache.org/25/javadoc/org/apache/kafka/clients/consumer/KafkaConsumer.html#close-java.time.Duration- > > > > > > > > > for > > > > details. Note that wakeup() > > > > < > > > > > > > > > > https://kafka.apache.org/25/javadoc/org/apache/kafka/clients/consumer/KafkaConsumer.html#wakeup-- > > > > > > > > > cannot > > > > be used to interrupt close. > > > > > > > > 2. During scale out operation, new pod (consumer) will be added to > the > > > > consumer group, so partitions of existing consumers will be > rebalanced > > to > > > > new consumer. In this case, I want to ensure that the current batch > of > > > > records polled and being processed by the consumer is processed and > > > offsets > > > > are committed before partition rebalance happens to new consumer. > > > > How can I ensure this with auto-commit enabled? > > > > > > > > > >
