Kirk True created KAFKA-16556:
---------------------------------

             Summary: Race condition between ConsumerRebalanceListener and 
SubscriptionState
                 Key: KAFKA-16556
                 URL: https://issues.apache.org/jira/browse/KAFKA-16556
             Project: Kafka
          Issue Type: Bug
          Components: clients, consumer
    Affects Versions: 3.7.0
            Reporter: Kirk True
             Fix For: 3.8.0


There appears to be a race condition between invoking the 
{{ConsumerRebalanceListener}} callbacks on reconciliation and 
{{initWithCommittedOffsetsIfNeeded}} in the consumer.
 
The membership manager adds the newly assigned partitions to the 
{{{}SubscriptionState{}}}, but marks them as {{{}pendingOnAssignedCallback{}}}. 
Then, after the {{ConsumerRebalanceListener.onPartitionsAssigned()}} completes, 
the membership manager will invoke {{enablePartitionsAwaitingCallback}} to set 
all of those partitions' 'pending' flag to false.
 
During the main {{Consumer.poll()}} loop, {{AsyncKafkaConsumer}} may need to 
call {{initWithCommittedOffsetsIfNeeded()}} if the positions aren't already 
cached. Inside {{{}initWithCommittedOffsetsIfNeeded{}}}, the consumer calls the 
subscription's {{initializingPartitions}} method to get a set of the partitions 
for which to fetch their committed offsets. However, 
{{SubscriptionState.initializingPartitions()}} only returns partitions that 
have the {{pendingOnAssignedCallback}} flag set to to false.
 
The result is: * If the {{MembershipManagerImpl.assignPartitions()}} future  is 
completed on the background thread first, the 'pending' flag is set to false. 
On the application thread, when {{SubscriptionState.initializingPartitions()}} 
is called, it returns the partition, and we fetch its committed offsets
 * If instead the application thread calls 
{{SubscriptionState.initializingPartitions()}} first, the partitions's 
'pending' flag is still set to false, and so the partition is omitted from the 
returned set. The {{updateFetchPositions()}} method then continues on and 
re-initializes the partition's fetch offset to 0.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to