Skip to content

Receiving Kafka Records

The Kafka Connector retrieves Kafka Records from Kafka Brokers and maps each of them to Reactive Messaging Messages.

Example

Let’s imagine you have a Kafka broker running, and accessible using the kafka:9092 address (by default it would use localhost:9092). Configure your application to receive Kafka records from a Kafka topic on the prices channel as follows:

1
2
3
4
5
kafka.bootstrap.servers=kafka:9092 # <1>

mp.messaging.incoming.prices.connector=smallrye-kafka # <2>
mp.messaging.incoming.prices.value.deserializer=org.apache.kafka.common.serialization.DoubleDeserializer # <3>
mp.messaging.incoming.prices.broadcast=true # <4>

  1. Configure the broker location. You can configure it globally or per channel

  2. Configure the connector to manage the prices channel

  3. Sets the (Kafka) deserializer to read the record’s value

  4. Make sure that we can receive from more than one consumer (see KafkaPriceConsumer and KafkaPriceMessageConsumer below)

Note

You don’t need to set the Kafka topic. By default, it uses the channel name (prices). You can configure the topic attribute to override it.

Then, your application receives Message<Double>. You can consume the payload directly:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
package kafka.inbound;

import jakarta.enterprise.context.ApplicationScoped;

import org.eclipse.microprofile.reactive.messaging.Incoming;

@ApplicationScoped
public class KafkaPriceConsumer {

    @Incoming("prices")
    public void consume(double price) {
        // process your price.
    }

}

Or, you can retrieve the Message<Double>:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
package kafka.inbound;

import java.util.concurrent.CompletionStage;

import jakarta.enterprise.context.ApplicationScoped;

import org.eclipse.microprofile.reactive.messaging.Incoming;
import org.eclipse.microprofile.reactive.messaging.Message;

@ApplicationScoped
public class KafkaPriceMessageConsumer {

    @Incoming("prices")
    public CompletionStage<Void> consume(Message<Double> price) {
        // process your price.

        // Acknowledge the incoming message (commit the offset)
        return price.ack();
    }

}

Deserialization

The deserialization is handled by the underlying Kafka Client. You need to configure the:

  • mp.messaging.incoming.[channel-name].value.deserializer to configure the value deserializer (mandatory)

  • mp.messaging.incoming.[channel-name].key.deserializer to configure the key deserializer (optional, default to String)

If you want to use a custom deserializer, add it to your CLASSPATH and configure the associate attribute.

In addition, the Kafka Connector also provides a set of message converters. So you can receive payloads representing records from Kafka using:

  • Record - a pair key/value
  • ConsumerRecord - a structure representing the record with all its metadata
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
@Incoming("topic-a")
public void consume(Record<String, String> record) {
    String key = record.key(); // Can be `null` if the incoming record has no key
    String value = record.value(); // Can be `null` if the incoming record has no value
}

@Incoming("topic-b")
public void consume(ConsumerRecord<String, String> record) {
    String key = record.key(); // Can be `null` if the incoming record has no key
    String value = record.value(); // Can be `null` if the incoming record has no value
    String topic = record.topic();
    int partition = record.partition();
    // ...
}

Inbound Metadata

Messages coming from Kafka contains an instance of IncomingKafkaRecordMetadata in the metadata. It provides the key, topic, partitions, headers and so on:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
IncomingKafkaRecordMetadata<String, Double> metadata = incoming.getMetadata(IncomingKafkaRecordMetadata.class)
        .orElse(null);
if (metadata != null) {
    // The topic
    String topic = metadata.getTopic();

    // The key
    String key = metadata.getKey();

    // The timestamp
    Instant timestamp = metadata.getTimestamp();

    // The underlying record
    ConsumerRecord<String, Double> record = metadata.getRecord();

    // ...
}

Acknowledgement

When a message produced from a Kafka record is acknowledged, the connector invokes a commit strategy. These strategies decide when the consumer offset for a specific topic/partition is committed. Committing an offset indicates that all previous records have been processed. It is also the position where the application would restart the processing after a crash recovery or a restart.

Committing every offset has performance penalties as Kafka offset management can be slow. However, not committing the offset often enough may lead to message duplication if the application crashes between two commits.

The Kafka connector supports three strategies:

  • throttled keeps track of received messages and commit to the next offset after the latest acked message in sequence. This strategy guarantees at-least-once delivery even if the channel performs asynchronous processing. The connector tracks the received records and periodically (period specified by auto.commit.interval.ms (default: 5000)) commits the highest consecutive offset. The connector will be marked as unhealthy if a message associated with a record is not acknowledged in throttled.unprocessed-record-max-age.ms (default: 60000). Indeed, this strategy cannot commit the offset as soon as a single record processing fails (see failure-strategy to configure what happens on failing processing). If throttled.unprocessed-record-max-age.ms is set to less than or equal to 0, it does not perform any health check verification. Such a setting might lead to running out of memory if there are poison pill messages. This strategy is the default if enable.auto.commit is not explicitly set to true.

  • checkpoint allows persisting consumer offsets on a "state store", instead of committing them back to the Kafka broker. Using the CheckpointMetadata API, consumer code can persist a processing state with the offset to mark the progress of a consumer. When the processing continues from a previously persisted offset, it seeks the Kafka consumer to that offset and also restores the persisted state, continuing the stateful processing from where it left off. The checkpoint strategy holds locally the processing state associated with the latest offset, and persists it periodically to the state store (period specified by auto.commit.interval.ms (default: 5000)). The connector will be marked as unhealthy if no processing state is persisted to the state store in checkpoint.unsynced-state-max-age.ms (default: 10000). Using the CheckpointMetadata API the user code can force to persist the state on message ack. If checkpoint.unsynced-state-max-age.ms is set to less than or equal to 0, it does not perform any health check verification. For more information, see Stateful processing with Checkpointing

  • latest commits the record offset received by the Kafka consumer as soon as the associated message is acknowledged (if the offset is higher than the previously committed offset). This strategy provides at-least-once delivery if the channel processes the message without performing any asynchronous processing. This strategy should not be used on high-load as offset commit is expensive. However, it reduces the risk of duplicates.

  • ignore performs no commit. This strategy is the default strategy when the consumer is explicitly configured with enable.auto.commit to true. It delegates the offset commit to the Kafka client. When enable.auto.commit is true this strategy DOES NOT guarantee at-least-once delivery. However, if the processing failed between two commits, messages received after the commit and before the failure will be re-processed.

Important

The Kafka connector disables the Kafka auto commit if not explicitly enabled. This behavior differs from the traditional Kafka consumer.

If high-throughout is important for you, and not limited by the downstream, we recommend to either:

  • Use the throttled policy
  • or set enable.auto.commit to true and annotate the consuming method with @Acknowledgment(Acknowledgment.Strategy.NONE)

Failure Management

If a message produced from a Kafka record is nacked, a failure strategy is applied. The Kafka connector supports 3 strategies:

  • fail - fail the application, no more records will be processed. (default) The offset of the record that has not been processed correctly is not committed.

  • ignore - the failure is logged, but the processing continue. The offset of the record that has not been processed correctly is committed.

  • dead-letter-queue - the offset of the record that has not been processed correctly is committed, but the record is written to a (Kafka) dead letter queue topic.

  • delayed-retry-topic - the offset of the record that has not been processed correctly is still committed, but the record is written to a series of Kafka topics for retrying the processing with some delay. This allows retrying failed records by reconsuming them later without blocking the processing of the latest records.

The strategy is selected using the failure-strategy attribute.

Dead Letter Queue

In the case of dead-letter-queue, you can configure the following attributes:

  • dead-letter-queue.topic: the topic to use to write the records not processed correctly, default is dead-letter-topic-$channel, with $channel being the name of the channel.

  • dead-letter-queue.producer-client-id: the client id used by the kafka producer when sending records to dead letter queue topic. If not specified it will default to kafka-dead-letter-topic-producer-$client-id, with $client-id being the value obtained from consumer client id.

  • dead-letter-queue.key.serializer: the serializer used to write the record key on the dead letter queue. By default, it deduces the serializer from the key deserializer.

  • dead-letter-queue.value.serializer: the serializer used to write the record value on the dead letter queue. By default, it deduces the serializer from the value deserializer.

The record written on the dead letter topic contains the original record’s headers, as well as a set of additional headers about the original record:

  • dead-letter-reason - the reason of the failure (the Throwable passed to nack())

  • dead-letter-cause - the cause of the failure (the getCause() of the Throwable passed to nack()), if any

  • dead-letter-topic - the original topic of the record

  • dead-letter-partition - the original partition of the record (integer mapped to String)

  • dead-letter-offset - the original offset of the record (long mapped to String)

When using dead-letter-queue, it is also possible to change some metadata of the record that is sent to the dead letter topic. To do that, use the Message.nack(Throwable, Metadata) method:

1
2
3
4
5
6
7
8
9
@Incoming("in")
public CompletionStage<Void> consume(Message<String> message) {
    return message.nack(new Exception("Failed!"), Metadata.of(
            OutgoingKafkaRecordMetadata.builder()
                    .withKey("failed-record")
                    .withHeaders(new RecordHeaders()
                            .add("my-header", "my-header-value".getBytes(StandardCharsets.UTF_8)))
                    .build()));
}

The Metadata may contain an instance of OutgoingKafkaRecordMetadata. If the instance is present, the following properties will be used:

  • key; if not present, the original record’s key will be used

  • topic; if not present, the configured dead letter topic will be used

  • partition; if not present, partition will be assigned automatically

  • headers; combined with the original record’s headers, as well as the dead-letter-* headers described above

Delayed Retry Topic

Experimental

Delayed retry topic feature is experimental.

The delayed retry topic strategy allows failed records to be automatically retried by forwarding them to a series of retry topics. Each retry topic is associated with a specific delay time, which is expressed in milliseconds. When a record processing fails, it is forwarded to the first retry topic. The failure strategy then consumes these records and dispatches them to be retried again once the delay time of the topic has elapsed.

If the processing of a record fails again, the message is forwarded to the next topic in the list, with possibly a longer delay time. If the processing of a record keeps failing, it will eventually be abandoned. Alternatively, if the dead-letter-queue.topic property is configured, the record will be sent to the dead letter queue.

The Kafka producer client used when forwarding records to retry topics can be configured using the dead-letter-queue properties namely, dead-letter-queue.producer-client-id, dead-letter-queue.key.serializer and dead-letter-queue.value.serializer.

Delayed retry topics and delays can be configured with following attributes:

  • delayed-retry-topic.topics : The comma-separated list of retry topics, each one suffixed with _[DELAY_IN_MILLISECONDS] for indicating the delay time. For example, my_retry_topic_2000,my_retry_topic_4000,my_retry_topic_10000 will use three topics my_retry_topic_2000, my_retry_topic_4000 and my_retry_topic_10000, with 2000ms 4000ms and 10000ms respectively.

    If not configured the source channel name is used, with 10, 20 and 50 seconds of delay, ex. for a channel named source, retry topics will be source_retry_10000, source_retry_20000, source_retry_50000.

  • delayed-retry-topic.max-retries : The maximum number of retries before abandoning the retries. If configured higher than the number of retry topics the last topic is used until maximum number of retries is reached. This can be configured to use a single retry topic with a fixed delay and multiple retries.

    For example, delayed-retry-topic.topics=source_retry_10000 and delayed-retry-topic.max-retries=4 will forward failed records to the topic source_retry_10000 with maximum of 4 retries.

  • delayed-retry-topic.timeout : The global timeout in milliseconds for a retried record. The timeout is calculated from the first failure for a record. If the next retry will reach the timeout, instead of forwarding to the retry topic the retry is abandoned and, if configured, the record is forwarded to the dead letter queue.

    The default is 120 seconds.

Important

While you can use Smallrye Fault Tolerance to retry processing, it will block the processing of further messages until the retried record is processed successfully, or abandoned.

Delayed retry topic failure strategy allows effectively implementing non-blocking retries. But it will not preserve the order of messages inside a topic-partition.

The record written on the delayed retry topics will preserve the key and partition of the original record. It also contains the original record’s headers, as well as a set of additional headers about the original record:

  • delayed-retry-count the current number of retries
  • delayed-retry-original-timestamp the original timestamp of the record
  • delayed-retry-first-processing-timestamp the first processing timestamp of the record
  • delayed-retry-reason the reason of the failure (the Throwable passed to nack())
  • delayed-retry-cause the cause of the failure (the getCause() of the Throwable passed to nack()), if any
  • delayed-retry-topic the original topic of the record
  • delayed-retry-partition the original partition of the record
  • delayed-retry-offset the original offset of the record
  • delayed-retry-exception-class-name the class name of the throwable passed to nack()
  • delayed-retry-cause-class-name the class name of the the getCause() of the Throwable passed to nack(), if any

As for the dead letter queue it is possible to change forwarded values by providing a OutgoingKafkaRecordMetadata when the message is nacked using Message.nack(Throwable, Metadata).

Multiple partitions

The delayed retry topic strategy does not create retry topics automatically. If the source topic has multiple partitions, delayed retry and dead letter queue topics would need to be setup with the same number of partitions.

It is possible to scale consumer application instances according to the number of partitions. But it is not guaranteed that the retry topics consumer will be assigned the same partition(s) as the main topic consumer. Therefore, retry processing of a record can happen in an other instance.

Custom commit and failure strategies

In addition to provided strategies, it is possible to implement custom commit and failure strategies and configure Kafka channels with them.

For example, for a custom commit strategy, implement the KafkaCommitHandler interface, and provide a managed bean implementing the KafkaCommitHandler.Factory interface, identified using @Identifier qualifier.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
package kafka.inbound;

import java.util.Collection;
import java.util.function.BiConsumer;

import jakarta.enterprise.context.ApplicationScoped;

import org.apache.kafka.common.TopicPartition;

import io.smallrye.common.annotation.Identifier;
import io.smallrye.mutiny.Uni;
import io.smallrye.reactive.messaging.kafka.IncomingKafkaRecord;
import io.smallrye.reactive.messaging.kafka.KafkaConnectorIncomingConfiguration;
import io.smallrye.reactive.messaging.kafka.KafkaConsumer;
import io.smallrye.reactive.messaging.kafka.commit.KafkaCommitHandler;
import io.vertx.mutiny.core.Vertx;

public class KafkaCustomCommit implements KafkaCommitHandler {

    @Override
    public <K, V> Uni<Void> handle(IncomingKafkaRecord<K, V> record) {
        // called on message ack
        return Uni.createFrom().voidItem();
    }

    @Override
    public <K, V> Uni<IncomingKafkaRecord<K, V>> received(IncomingKafkaRecord<K, V> record) {
        // called before message processing
        return Uni.createFrom().item(record);
    }

    @Override
    public void terminate(boolean graceful) {
        // called on channel shutdown
    }

    @Override
    public void partitionsAssigned(Collection<TopicPartition> partitions) {
        // called on partitions assignment
    }

    @Override
    public void partitionsRevoked(Collection<TopicPartition> partitions) {
        // called on partitions revoked
    }

    @ApplicationScoped
    @Identifier("custom")
    public static class Factory implements KafkaCommitHandler.Factory {

        @Override
        public KafkaCommitHandler create(KafkaConnectorIncomingConfiguration config,
                Vertx vertx,
                KafkaConsumer<?, ?> consumer,
                BiConsumer<Throwable, Boolean> reportFailure) {
            return new KafkaCustomCommit(/* ... */);
        }
    }

}

Finally, to use the custom commit strategy, set the commit-strategy attribute to the identifier of the commit handler factory: mp.messaging.incoming.$channel.commit-strategy=custom. Similarly, custom failure strategies can be configured using failure-strategy attribute.

Note

If the custom strategy implementation inherits ContextHolder class it can access the Vert.x event-loop context created for the Kafka consumer

Retrying processing

You can combine Reactive Messaging with SmallRye Fault Tolerance, and retry processing when it fails:

1
2
3
4
5
6
@Incoming("kafka")
@Outgoing("processed")
@Retry(delay = 10, maxRetries = 5)
public String process(String v) {
   // ... retry if this method throws an exception
}

You can configure the delay, the number of retries, the jitter...

If your method returns a Uni, you need to add the @NonBlocking annotation:

1
2
3
4
5
6
7
@Incoming("kafka")
@Outgoing("processed")
@Retry(delay = 10, maxRetries = 5)
@NonBlocking
public Uni<String> process(String v) {
   // ... retry if this method throws an exception or the returned Uni produce a failure
}

The incoming messages are acknowledged only once the processing completes successfully. So, it commits the offset after the successful processing. If after the retries the processing still failed, the message is nacked and the failure strategy is applied.

You can also use @Retry on methods only consuming incoming messages:

1
2
3
4
5
@Incoming("kafka")
@Retry(delay = 10, maxRetries = 5)
public void consume(String v) {
   // ... retry if this method throws an exception
}

Handling deserialization failures

Because deserialization happens before creating a Message, the failure strategy presented above cannot be applied. However, when a deserialization failure occurs, you can intercept it and provide a fallback value. To achieve this, create a CDI bean implementing the DeserializationFailureHandler interface:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
@ApplicationScoped
@Identifier("failure-retry") // Set the name of the failure handler
public class MyDeserializationFailureHandler
    implements DeserializationFailureHandler<JsonObject> { // Specify the expected type

    @Override
    public JsonObject decorateDeserialization(Uni<JsonObject> deserialization,
            String topic, boolean isKey, String deserializer, byte[] data,
            Headers headers) {
        return deserialization
                    .onFailure().retry().atMost(3)
                    .await().atMost(Duration.ofMillis(200));
    }
}

The bean must be exposed with the @Identifier qualifier specifying the name of the bean. Then, in the connector configuration, specify the following attribute:

  • mp.messaging.incoming.$channel.key-deserialization-failure-handler: name of the bean handling deserialization failures happening for the record’s key

  • mp.messaging.incoming.$channel.value-deserialization-failure-handler: name of the bean handling deserialization failures happening for the record’s value,

The handler is called with the deserialization action as a Uni<T>, the record’s topic, a boolean indicating whether the failure happened on a key, the class name of the deserializer that throws the exception, the corrupted data, the exception, and the records headers augmented with headers describing the failure (which ease the write to a dead letter). On the deserialization Uni failure strategies like retry, providing a fallback value or applying timeout can be implemented. Note that the method must await on the result and return the deserialized object. Alternatively, the handler can only implement handleDeserializationFailure method and provide a fallback value, which may be null.

If you don’t configure a deserialization failure handlers and a deserialization failure happens, the application is marked unhealthy. You can also ignore the failure, which will log the exception and produce a null value. To enable this behavior, set the mp.messaging.incoming.$channel.fail-on-deserialization-failure attribute to false.

If the fail-on-deserialization-failure attribute is set to false and the failure-strategy attribute is dead-letter-queue the failed record will be sent to the corresponding dead letter queue topic. The forwarded record will have the original key and value, and the following headers set:

  • deserialization-failure-reason: The deserialization failure message
  • deserialization-failure-cause: The deserialization failure cause if any
  • deserialization-failure-key: Whether the deserialization failure happened on a key
  • deserialization-failure-topic: The topic of the incoming message when a deserialization failure happen
  • deserialization-failure-deserializer: The class name of the underlying deserializer
  • deserialization-failure-key-data: If applicable the key data that was not able to be deserialized
  • deserialization-failure-value-data: If applicable the value data that was not able to be deserialized

Receiving Cloud Events

The Kafka connector supports Cloud Events. When the connector detects a structured or binary Cloud Events, it adds a IncomingKafkaCloudEventMetadata in the metadata of the Message. IncomingKafkaCloudEventMetadata contains the various (mandatory and optional) Cloud Event attributes.

If the connector cannot extract the Cloud Event metadata, it sends the Message without the metadata.

Binary Cloud Events

For binary Cloud Events, all mandatory Cloud Event attributes must be set in the record header, prefixed by ce_ (as mandated by the protocol binding). The connector considers headers starting with the ce_ prefix but not listed in the specification as extensions. You can access them using the getExtension method from IncomingKafkaCloudEventMetadata. You can retrieve them as String.

The datacontenttype attribute is mapped to the content-type header of the record. The partitionkey attribute is mapped to the record’s key, if any.

Note that all headers are read as UTF-8.

With binary Cloud Events, the record’s key and value can use any deserializer.

Structured Cloud Events

For structured Cloud Events, the event is encoded in the record’s value. Only JSON is supported, so your event must be encoded as JSON in the record’s value.

Structured Cloud Event must set the content-type header of the record to application/cloudevents or prefix the value with application/cloudevents such as: application/cloudevents+json; charset=UTF-8.

To receive structured Cloud Events, your value deserializer must be:

  • org.apache.kafka.common.serialization.StringDeserializer

  • org.apache.kafka.common.serialization.ByteArrayDeserializer

  • io.vertx.kafka.client.serialization.JsonObjectDeserializer

As mentioned previously, the value must be a valid JSON object containing at least all the mandatory Cloud Events attributes.

If the record is a structured Cloud Event, the created Message’s payload is the Cloud Event data.

The partitionkey attribute is mapped to the record’s key if any.

Consumer Rebalance Listener

To handle offset commit and assigned partitions yourself, you can provide a consumer rebalance listener. To achieve this, implement the io.smallrye.reactive.messaging.kafka.KafkaConsumerRebalanceListener interface, make the implementing class a bean, and add the @Identifier qualifier. A usual use case is to store offset in a separate data store to implement exactly-once semantic, or starting the processing at a specific offset.

The listener is invoked every time the consumer topic/partition assignment changes. For example, when the application starts, it invokes the partitionsAssigned callback with the initial set of topics/partitions associated with the consumer. If, later, this set changes, it calls the partitionsRevoked and partitionsAssigned callbacks again, so you can implement custom logic.

Note that the rebalance listener methods are called from the Kafka polling thread and must block the caller thread until completion. That’s because the rebalance protocol has synchronization barriers, and using asynchronous code in a rebalance listener may be executed after the synchronization barrier.

When topics/partitions are assigned or revoked from a consumer, it pauses the message delivery and restarts once the rebalance completes.

If the rebalance listener handles offset commit on behalf of the user (using the ignore commit strategy), the rebalance listener must commit the offset synchronously in the partitionsRevoked callback. We also recommend applying the same logic when the application stops.

Unlike the ConsumerRebalanceListener from Apache Kafka, the io.smallrye.reactive.messaging.kafka.KafkaConsumerRebalanceListener methods pass the Kafka Consumer and the set of topics/partitions.

Example

In this example we set-up a consumer that always starts on messages from at most 10 minutes ago (or offset 0). First we need to provide a bean that implements the io.smallrye.reactive.messaging.kafka.KafkaConsumerRebalanceListener interface and is annotated with @Identifier. We then must configure our inbound connector to use this named bean.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
package kafka.inbound;

import java.util.Collection;
import java.util.HashMap;
import java.util.Map;
import java.util.logging.Logger;

import jakarta.enterprise.context.ApplicationScoped;

import org.apache.kafka.clients.consumer.Consumer;
import org.apache.kafka.clients.consumer.OffsetAndTimestamp;

import io.smallrye.common.annotation.Identifier;
import io.smallrye.reactive.messaging.kafka.KafkaConsumerRebalanceListener;

@ApplicationScoped
@Identifier("rebalanced-example.rebalancer")
public class KafkaRebalancedConsumerRebalanceListener implements KafkaConsumerRebalanceListener {

    private static final Logger LOGGER = Logger.getLogger(KafkaRebalancedConsumerRebalanceListener.class.getName());

    /**
     * When receiving a list of partitions will search for the earliest offset within 10 minutes
     * and seek the consumer to it.
     *
     * @param consumer underlying consumer
     * @param partitions set of assigned topic partitions
     */
    @Override
    public void onPartitionsAssigned(Consumer<?, ?> consumer,
            Collection<org.apache.kafka.common.TopicPartition> partitions) {
        long now = System.currentTimeMillis();
        long shouldStartAt = now - 600_000L; //10 minute ago

        Map<org.apache.kafka.common.TopicPartition, Long> request = new HashMap<>();
        for (org.apache.kafka.common.TopicPartition partition : partitions) {
            LOGGER.info("Assigned " + partition);
            request.put(partition, shouldStartAt);
        }
        Map<org.apache.kafka.common.TopicPartition, OffsetAndTimestamp> offsets = consumer
                .offsetsForTimes(request);
        for (Map.Entry<org.apache.kafka.common.TopicPartition, OffsetAndTimestamp> position : offsets.entrySet()) {
            long target = position.getValue() == null ? 0L : position.getValue().offset();
            LOGGER.info("Seeking position " + target + " for " + position.getKey());
            consumer.seek(position.getKey(), target);
        }
    }

}

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
package kafka.inbound;

import java.util.concurrent.CompletableFuture;
import java.util.concurrent.CompletionStage;

import jakarta.enterprise.context.ApplicationScoped;

import org.eclipse.microprofile.reactive.messaging.Acknowledgment;
import org.eclipse.microprofile.reactive.messaging.Incoming;
import org.eclipse.microprofile.reactive.messaging.Message;

@ApplicationScoped
public class KafkaRebalancedConsumer {

    @Incoming("rebalanced-example")
    @Acknowledgment(Acknowledgment.Strategy.NONE)
    public CompletionStage<Void> consume(Message<String> message) {
        // We don't need to ACK messages because in this example we set offset during consumer re-balance
        return CompletableFuture.completedFuture(null);
    }

}

To configure the inbound connector to use the provided listener we either set the consumer rebalance listener’s name:

  • mp.messaging.incoming.rebalanced-example.consumer-rebalance-listener.name=rebalanced-example.rebalancer

Or have the listener’s name be the same as the group id:

  • mp.messaging.incoming.rebalanced-example.group.id=rebalanced-example.rebalancer

Setting the consumer rebalance listener’s name takes precedence over using the group id.

Receiving Kafka Records in Batches

By default, incoming methods receive each Kafka record individually. Under the hood, Kafka consumer clients poll the broker constantly and receive records in batches, presented inside the ConsumerRecords container.

In batch mode, your application can receive all the records returned by the consumer poll in one go.

To achieve this you need to set mp.messaging.incoming.$channel.batch=true and specify a compatible container type to receive all the data:

1
2
3
4
5
6
@Incoming("prices")
public void consume(List<Double> prices) {
    for (double price : prices) {
        // process price
    }
}

The incoming method can also receive Message<List<Payload>, KafkaBatchRecords<Payload> ConsumerRecords<Key, Payload> types, They give access to record details such as offset or timestamp :

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
@Incoming("prices")
public CompletionStage<Void> consumeMessage(KafkaRecordBatch<String, Double> records) {
    for (KafkaRecord<String, Double> record : records) {
        record.getMetadata(IncomingKafkaRecordMetadata.class).ifPresent(metadata -> {
            int partition = metadata.getPartition();
            long offset = metadata.getOffset();
            Instant timestamp = metadata.getTimestamp();
            //... process messages
        });
    }
    // ack will commit the latest offsets (per partition) of the batch.
    return records.ack();
}

@Incoming("prices")
public void consumeRecords(ConsumerRecords<String, Double> records) {
    for (TopicPartition partition : records.partitions()) {
        for (ConsumerRecord<String, Double> record : records.records(partition)) {
            //... process messages
        }
    }
}

Note that the successful processing of the incoming record batch will commit the latest offsets for each partition received inside the batch. The configured commit strategy will be applied for these records only.

Conversely, if the processing throws an exception, all messages are nacked, applying the failure strategy for all the records inside the batch.

Manual topic-partition assignment

The default behavior of Kafka incoming channels is to subscribe to one or more topics in order to receive records from the Kafka broker. Channel attributes topic and topics allow specifying topics to subscribe to, or pattern attribute allows to subscribe to all topics matching a regular expression. Subscribing to topics allows partitioning consumption of topics by dynamically assigning (rebalancing) partitions between members of a consumer group.

The assign-seek configuration attribute allows manually assigning topic-partitions to a Kafka incoming channel, and optionally seek to a specified offset in the partition to start consuming records. If assign-seek is used, the consumer will not be dynamically subscribed to topics, but instead will statically assign the described partitions. In manual topic-partition rebalancing doesn't happen and therefore rebalance listeners are never called.

The attribute takes a list of triplets separated by commas: <topic>:<partition>:<offset>.

For example, the following configuration

1
mp.messaging.incoming.data.assign-seek=topic1:0:10, topic2:1:20

assigns the consumer to: - Partition 0 of topic 'topic1', setting the initial position at offset 10. - Partition 1 of topic 'topic2', setting the initial position at offset 20.

The topic, partition, and offset in each triplet can have the following variations: - If the topic is omitted, the configured topic will be used. - If the offset is omitted, partitions are assigned to the consumer but won't be seeked to offset. - If offset is 0, it seeks to the beginning of the topic-partition. - If offset is -1, it seeks to the end of the topic-partition.

Stateful processing with Checkpointing

Experimental

Checkpointing is experimental, and APIs and features are subject to change in the future.

The checkpoint commit strategy allows for a Kafka incoming channel to manage topic-partition offsets, not by committing on the Kafka broker, but by persisting consumers' advancement on a state store.

In addition to that, if the consumer builds an internal state as a result of consumed records, the topic-partition offset persisted to the state store can be associated with a processing state, saving the local state to the persistent store. When a consumer restarts or consumer group instances scale, i.e. when new partitions get assigned to the consumer, the checkpointing works by resuming the processing from the latest offset and its saved state.

The @Incoming channel consumer code can manipulate the processing state through the CheckpointMetadata API:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
@Incoming("prices")
public CompletionStage<Void> consume(Message<Double> record) {
    // Get the `CheckpointMetadata` from the incoming message
    CheckpointMetadata<Double> checkpoint = CheckpointMetadata.fromMessage(record);

    // `CheckpointMetadata` allows transforming the processing state
    // Applies the given function, starting from the value `0.0` when no previous state exists
    checkpoint.transform(0.0, current -> current + record.getPayload(), /* persistOnAck */ true);

    // `persistOnAck` flag set to true, ack will persist the processing state
    // associated with the latest offset (per partition).
    return record.ack();
}

The transform method allows applying a transformation function to the current state, producing a changed state and registering it locally for checkpointing. By default, the local state is synced (persisted) to the state store periodically, period specified by auto.commit.interval.ms, (default: 5000). If persistOnAck flag is given, the latest state is persisted to the state store eagerly on message acknowledgment. The setNext method works similarly directly setting the latest state.

The checkpoint commit strategy tracks when a processing state is last persisted for each topic-partition. If an outstanding state change can not be persisted for checkpoint.unsynced-state-max-age.ms (default: 10000), the channel is marked unhealthy.

Where and how processing states are persisted is decided by the state store implementation. This can be configured on the incoming channel using checkpoint.state-store configuration property, using the state store factory identifier name. The serialization of state objects depends on the state store implementation. In order to instruct state stores for serialization can require configuring the type name of state objects using checkpoint.state-type property.

In order to keep Smallrye Reactive Messaging free of persistence-related dependencies, this library includes only a default state store named file. It is based on Vert.x Filesystem API and stores the processing state in Json formatted files, in a local directory configured by the checkpoint.file.state-dir property. State files follow the naming scheme [consumer-group-id]:[topic]:[partition].

Implementing State Stores

State store implementations are required to implement CheckpointStateStore interface, and provide a managed bean implementing CheckpointStateStore.Factory, identified with @Identifier bean qualifier indicating the name of the state-store. The factory bean identifier indicates the name to configure on checkpoint.state-store config property. The factory is discovered as a CDI managed bean and state store is created once per channel:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
package kafka.inbound;

import java.util.Collection;
import java.util.Map;

import jakarta.enterprise.context.ApplicationScoped;

import org.apache.kafka.clients.consumer.ConsumerConfig;
import org.apache.kafka.common.TopicPartition;

import io.smallrye.common.annotation.Identifier;
import io.smallrye.mutiny.Uni;
import io.smallrye.reactive.messaging.kafka.KafkaConnectorIncomingConfiguration;
import io.smallrye.reactive.messaging.kafka.KafkaConsumer;
import io.smallrye.reactive.messaging.kafka.commit.CheckpointStateStore;
import io.smallrye.reactive.messaging.kafka.commit.ProcessingState;
import io.vertx.mutiny.core.Vertx;

public class MyCheckpointStateStore implements CheckpointStateStore {

    private final String consumerGroupId;
    private final Class<?> stateType;

    public MyCheckpointStateStore(String consumerGroupId, Class<?> stateType) {
        this.consumerGroupId = consumerGroupId;
        this.stateType = stateType;
    }

    /**
     * Can be used with
     * {@code mp.reactive.messaging.incoming.my-channel.commit-strategy=checkpoint}
     * {@code mp.reactive.messaging.incoming.my-channel.checkpoint.state-store=my-store}
     */
    @ApplicationScoped
    @Identifier("my-store")
    public static class Factory implements CheckpointStateStore.Factory {

        @Override
        public CheckpointStateStore create(KafkaConnectorIncomingConfiguration config,
                Vertx vertx,
                KafkaConsumer<?, ?> consumer,
                Class<?> stateType /* if configured, otherwise null */) {
            String consumerGroupId = (String) consumer.configuration().get(ConsumerConfig.GROUP_ID_CONFIG);
            return new MyCheckpointStateStore(consumerGroupId, stateType);
        }
    }

    @Override
    public Uni<Map<TopicPartition, ProcessingState<?>>> fetchProcessingState(Collection<TopicPartition> partitions) {
        // Called on Vert.x event loop
        // Return a Uni completing with the map of topic-partition to processing state
        // The Uni will be subscribed also on Vert.x event loop
        return Uni.createFrom().nullItem();
    }

    @Override
    public Uni<Void> persistProcessingState(Map<TopicPartition, ProcessingState<?>> state) {
        // Called on Vert.x event loop
        // Return a Uni completing with void when the given states are persisted
        // The Uni will be subscribed also on Vert.x event loop
        return Uni.createFrom().voidItem();
    }

    @Override
    public void close() {
        /* Called when channel is closing, no-op by default */
    }
}

The checkpoint commit strategy calls the state store in following events:

  • fetchProcessingState : on partitions assigned, to seek the consumer to the latest offset.
  • persistProcessingState on partitions revoked, to persist the state of last processed record.
  • persistProcessingState on message acknowledgement, if a new state is set during the processing and persistOnAck flag is set.
  • persistProcessingState on auto.commit.interval.ms intervals, if a new state is set during processing.
  • persistProcessingState on channel shutdown.
  • close on channel shutdown.

Configuration Reference

Attribute (alias) Description Type Mandatory Default
assign-seek Assign partitions and optionally seek to offsets, instead of subscribing to topics. A comma-separating list of triplets in form of <topic>:|<partition>|:<offset> to assign statically to the consumer and seek to the given offsets. Offset 0 seeks to beginning and offset -1 seeks to the end of the topic-partition. If the topic is omitted the configured topic will be used. If the offset is omitted partitions are assigned to the consumer but won't be seeked to offset. string false
auto.offset.reset What to do when there is no initial offset in Kafka.Accepted values are earliest, latest and none string false latest
batch Whether the Kafka records are consumed in batch. The channel injection point must consume a compatible type, such as List<Payload> or KafkaRecordBatch<Payload>. boolean false false
bootstrap.servers (kafka.bootstrap.servers) A comma-separated list of host:port to use for establishing the initial connection to the Kafka cluster. string false localhost:9092
broadcast Whether the Kafka records should be dispatched to multiple consumer boolean false false
checkpoint.state-store While using the checkpoint commit-strategy, the name set in @Identifier of a bean that implements io.smallrye.reactive.messaging.kafka.StateStore.Factory to specify the state store implementation. string false
checkpoint.state-type While using the checkpoint commit-strategy, the fully qualified type name of the state object to persist in the state store. When provided, it can be used by the state store implementation to help persisting the processing state object. string false
checkpoint.unsynced-state-max-age.ms While using the checkpoint commit-strategy, specify the max age in milliseconds that the processing state must be persisted before the connector is marked as unhealthy. Setting this attribute to 0 disables this monitoring. int false 10000
client-id-prefix Prefix for Kafka client client.id attribute. If defined configured or generated client.id will be prefixed with the given value. string false
cloud-events Enables (default) or disables the Cloud Event support. If enabled on an incoming channel, the connector analyzes the incoming records and try to create Cloud Event metadata. If enabled on an outgoing, the connector sends the outgoing messages as Cloud Event if the message includes Cloud Event Metadata. boolean false true
commit-strategy Specify the commit strategy to apply when a message produced from a record is acknowledged. Values can be latest, ignore or throttled. If enable.auto.commit is true then the default is ignore otherwise it is throttled string false
consumer-rebalance-listener.name The name set in @Identifier of a bean that implements io.smallrye.reactive.messaging.kafka.KafkaConsumerRebalanceListener. If set, this rebalance listener is applied to the consumer. string false
dead-letter-queue.key.serializer When the failure-strategy is set to dead-letter-queue indicates the key serializer to use. If not set the serializer associated to the key deserializer is used string false
dead-letter-queue.producer-client-id When the failure-strategy is set to dead-letter-queue indicates what client id the generated producer should use. Defaults is kafka-dead-letter-topic-producer-$client-id string false
dead-letter-queue.topic When the failure-strategy is set to dead-letter-queue indicates on which topic the record is sent. Defaults is dead-letter-topic-$channel string false
dead-letter-queue.value.serializer When the failure-strategy is set to dead-letter-queue indicates the value serializer to use. If not set the serializer associated to the value deserializer is used string false
delayed-retry-topic.max-retries When the failure-strategy is set to delayed-retry-topic indicates the maximum number of retries. If higher than the number of delayed retry topics, last topic is used. int false
delayed-retry-topic.timeout When the failure-strategy is set to delayed-retry-topic indicates the global timeout per record. int false 120000
delayed-retry-topic.topics When the failure-strategy is set to delayed-retry-topic indicates topics to use. If not set the source channel name is used, with 10, 20 and 50 seconds delayed topics. string false
enable.auto.commit If enabled, consumer's offset will be periodically committed in the background by the underlying Kafka client, ignoring the actual processing outcome of the records. It is recommended to NOT enable this setting and let Reactive Messaging handles the commit. boolean false false
fail-on-deserialization-failure When no deserialization failure handler is set and a deserialization failure happens, report the failure and mark the application as unhealthy. If set to false and a deserialization failure happens, a null value is forwarded. boolean false true
failure-strategy Specify the failure strategy to apply when a message produced from a record is acknowledged negatively (nack). Values can be fail (default), ignore, or dead-letter-queue string false fail
fetch.min.bytes The minimum amount of data the server should return for a fetch request. The default setting of 1 byte means that fetch requests are answered as soon as a single byte of data is available or the fetch request times out waiting for data to arrive. int false 1
graceful-shutdown Whether or not a graceful shutdown should be attempted when the application terminates. boolean false true
group.id A unique string that identifies the consumer group the application belongs to. If not set, a unique, generated id is used string false
health-enabled Whether health reporting is enabled (default) or disabled boolean false true
health-readiness-enabled Whether readiness health reporting is enabled (default) or disabled boolean false true
health-readiness-timeout deprecated - During the readiness health check, the connector connects to the broker and retrieves the list of topics. This attribute specifies the maximum duration (in ms) for the retrieval. If exceeded, the channel is considered not-ready. Deprecated: Use 'health-topic-verification-timeout' instead. long false
health-readiness-topic-verification deprecated - Whether the readiness check should verify that topics exist on the broker. Default to false. Enabling it requires an admin connection. Deprecated: Use 'health-topic-verification-enabled' instead. boolean false
health-topic-verification-enabled Whether the startup and readiness check should verify that topics exist on the broker. Default to false. Enabling it requires an admin client connection. boolean false false
health-topic-verification-readiness-disabled Whether the topic verification is disabled for the readiness health check. boolean false false
health-topic-verification-startup-disabled Whether the topic verification is disabled for the startup health check. boolean false false
health-topic-verification-timeout During the startup and readiness health check, the connector connects to the broker and retrieves the list of topics. This attribute specifies the maximum duration (in ms) for the retrieval. If exceeded, the channel is considered not-ready. long false 2000
kafka-configuration Identifier of a CDI bean that provides the default Kafka consumer/producer configuration for this channel. The channel configuration can still override any attribute. The bean must have a type of Map and must use the @io.smallrye.common.annotation.Identifier qualifier to set the identifier. string false
key-deserialization-failure-handler The name set in @Identifier of a bean that implements io.smallrye.reactive.messaging.kafka.DeserializationFailureHandler. If set, deserialization failure happening when deserializing keys are delegated to this handler which may retry or provide a fallback value. string false
key.deserializer The deserializer classname used to deserialize the record's key string false org.apache.kafka.common.serialization.StringDeserializer
lazy-client Whether Kafka client is created lazily or eagerly. boolean false false
max-queue-size-factor Multiplier factor to determine maximum number of records queued for processing, using max.poll.records * max-queue-size-factor. Defaults to 2. In batch mode max.poll.records is considered 1. int false 2
partitions The number of partitions to be consumed concurrently. The connector creates the specified amount of Kafka consumers. It should match the number of partition of the targeted topic int false 1
pattern Indicate that the topic property is a regular expression. Must be used with the topic property. Cannot be used with the topics property boolean false false
pause-if-no-requests Whether the polling must be paused when the application does not request items and resume when it does. This allows implementing back-pressure based on the application capacity. Note that polling is not stopped, but will not retrieve any records when paused. boolean false true
poll-timeout The polling timeout in milliseconds. When polling records, the poll will wait at most that duration before returning records. Default is 1000ms int false 1000
requests When partitions is greater than 1, this attribute allows configuring how many records are requested by each consumers every time. int false 128
retry Whether or not the connection to the broker is re-attempted in case of failure boolean false true
retry-attempts The maximum number of reconnection before failing. -1 means infinite retry int false -1
retry-max-wait The max delay (in seconds) between 2 reconnects int false 30
throttled.unprocessed-record-max-age.ms While using the throttled commit-strategy, specify the max age in milliseconds that an unprocessed message can be before the connector is marked as unhealthy. Setting this attribute to 0 disables this monitoring. int false 60000
topic The consumed / populated Kafka topic. If neither this property nor the topics properties are set, the channel name is used string false
topics A comma-separating list of topics to be consumed. Cannot be used with the topic or pattern properties string false
tracing-enabled Whether tracing is enabled (default) or disabled boolean false true
value-deserialization-failure-handler The name set in @Identifier of a bean that implements io.smallrye.reactive.messaging.kafka.DeserializationFailureHandler. If set, deserialization failure happening when deserializing values are delegated to this handler which may retry or provide a fallback value. string false
value.deserializer The deserializer classname used to deserialize the record's value string true

You can also pass any property supported by the underlying Kafka consumer.

For example, to configure the max.poll.records property, use:

1
mp.messaging.incoming.[channel].max.poll.records=1000

Some consumer client properties are configured to sensible default values:

If not set, reconnect.backoff.max.ms is set to 10000 to avoid high load on disconnection.

If not set, key.deserializer is set to org.apache.kafka.common.serialization.StringDeserializer.

The consumer client.id is configured according to the number of clients to create using mp.messaging.incoming.[channel].partitions property.

  • If a client.id is provided, it is used as-is or suffixed with client index if partitions property is set.

  • If a client.id is not provided, it is generated as kafka-consumer-[channel][-index].