spring integration kafka listener thread reads multiple partitions when concurrency = partition count

Question

I set up a Spring Integration flow to process a topic having 3 partitions and set the listener container's concurrency to 3. As expected, I see three threads processing batches from all 3 partitions. However, I see that in some cases, one of the listener threads may process a single batch containing messages from multiple partitions. My data is partitioned in kafka by an id so that it may be processed concurrently with other ids, but not with the same ids on another thread (which is what I was surprised to observe is happening). I thought from reading the docs that each thread would be assigned a partition. I'm using a KafkaMessageDrivenChannelAdapter like this:

private static final Class<List<MyEvent>> payloadClass = (Class<List<MyEvent>>)(Class) List.class;

public KafkaMessageDrivenChannelAdapterSpec.KafkaMessageDrivenChannelAdapterListenerContainerSpec<String, MyEvent> myChannelAdapterSpec() {
        return Kafka.messageDrivenChannelAdapter(tstatEventConsumerFactory(),
                KafkaMessageDrivenChannelAdapter.ListenerMode.batch, "my-topic") //3 partitions
                .configureListenerContainer(c -> {
                    c.ackMode(ContainerProperties.AckMode.BATCH);
                    c.id(_ID);
                    c.concurrency(3);
                    RecoveringBatchErrorHandler errorHandler = new RecoveringBatchErrorHandler(
                            (record, exception) -> log.error("failed to handle record at offset {}: {}",
                                    record.offset(), record.value(), exception),
                            new FixedBackOff(FixedBackOff.DEFAULT_INTERVAL, 2)
                    );
                    c.errorHandler(errorHandler);
                });
    }
@Bean
public IntegrationFlow myIntegrationFlow() {
        return IntegrationFlows.from(myChannelAdapterSpec())
                .handle(payloadClass, (payload, headers) -> {
                    service.performSink(payload);
                    return null;
                })
                .get();
    }

How do I set this up so that each listener container thread only processes messages from one partition?

Answer 1

But is there additionally a way that I can keep from ever getting a batch with messages from multiple partitions, even if a rebalance does occur?

That's not how consumer group works. If you would like to have a "sticky" consumers, then consider to use a manual assignment. See the channel adapter factory based on the TopicPartitionOffset... topicPartitions :

/**
 * Create an initial
 * {@link KafkaMessageDrivenChannelAdapterSpec.KafkaMessageDrivenChannelAdapterListenerContainerSpec}.
 * @param consumerFactory the {@link ConsumerFactory}.
 * @param listenerMode the {@link KafkaMessageDrivenChannelAdapter.ListenerMode}.
 * @param topicPartitions the {@link TopicPartitionOffset} vararg.
 * @param <K> the Kafka message key type.
 * @param <V> the Kafka message value type.
 * @return the KafkaMessageDrivenChannelAdapterSpec.KafkaMessageDrivenChannelAdapterListenerContainerSpec.
 */
public static <K, V>
KafkaMessageDrivenChannelAdapterSpec.KafkaMessageDrivenChannelAdapterListenerContainerSpec<K, V> messageDrivenChannelAdapter(
        ConsumerFactory<K, V> consumerFactory,
        KafkaMessageDrivenChannelAdapter.ListenerMode listenerMode,
        TopicPartitionOffset... topicPartitions) {

Then it is not going to be treated as consumer group and you have to create several channel adapters pointing each to its specific partition. All of this channel adapters may emit messages to the same MessageChannel .