Develop Reference

Handling commits for errors with #KafkaListener in batch consumers

We have a Kafka Consumer setup like below
#Bean
public ConsumerFactory<String, Object> consumerFactory() {
final Map<String, Object> props = kafkaProperties.buildConsumerProperties();
return new DefaultKafkaConsumerFactory<>(props);
}
#Bean
public ConcurrentKafkaListenerContainerFactory<String, Object> batchFactory(
final ConsumerFactory<String, Object> consumerFactory,
#Value("${someProp.batch}") final boolean enableBatchListener,
#Value("${someProp.concurrency}") final int consumerConcurrency,
#Value("${someProp.error.backoff.ms}") final int errorBackoffInterval
) {
final SeekToCurrentBatchErrorHandler errorHandler = new SeekToCurrentBatchErrorHandler();
errorHandler.setBackOff(new FixedBackOff(errorBackoffInterval, UNLIMITED_ATTEMPTS));
final var containerFactory = new ConcurrentKafkaListenerContainerFactory<String, Object>();
containerFactory.setConsumerFactory(consumerFactory);
containerFactory.getContainerProperties().setAckMode(MANUAL_IMMEDIATE);
containerFactory.getContainerProperties().setMissingTopicsFatal(false);
containerFactory.setBatchListener(enableBatchListener);
containerFactory.setConcurrency(consumerConcurrency);
containerFactory.setBatchErrorHandler(errorHandler);
return containerFactory;
}
someProp:
concurrency: 16
batch: true
error.backoff.ms: 2000
spring:
kafka:
bootstrap-servers: ${KAFKA_BOOTSTRAP_SERVERS}
consumer:
groupId: some-grp
autoOffsetReset: earliest
keyDeserializer: org.apache.kafka.common.serialization.StringDeserializer
valueDeserializer: io.confluent.kafka.serializers.KafkaAvroDeserializer
properties:
schema.registry.url: ${SCHEMA_REGISTRY_URL}
specific.avro.reader: true
security.protocol: SSL
In batch listener method annotated with #KafkaListener, we call acknowledgment.acknowledge() at the end of processing of the list. Assuming that when the service comes up, I already have a million messages in the topic ready to be consumed by the service, I have following questions with respect to this scenario as I could not find documentation which talks in detail regarding the batch listening:
The listener will read 500 messages in the list. 500 because max.poll.records is not set and hence defaults to 500, so the list will have 500 messages. Is this understanding correct?
Given the above, where does the consumer concurrency come into picture? Does the stated configuration mean I will have 16 consumers each of which can read 500 messages in parallel from the same topic?
I understand, in this case I must have at least 16 partitions to make use of all the consumers otherwise I would be left with consumers who do nothing?
Due to SeekToCurrentBatchErrorHandler, the batch will be replayed in case there is any exception in processing inside the listener method. So, if in a particular batch there is an exception processing the 50th message, first 49 will be played again (basically duplicates, which I am fine with), next 50 to 500 messages will be played and tried for processing as usual. Is this understanding correct?
If there are multiple batches being read continuously and a particular consumer thread gets stuck with the SeekToCurrentBatchErrorHandler, how is the offset commit handled, as other consumer threads would still be processing the messages successfully thus moving the offset pointer way forward then the stuck consumers offsets
The doc for MANUAL_IMMEDIATE states
/**
* User takes responsibility for acks using an
* {#link AcknowledgingMessageListener}. The consumer
* immediately processes the commit.
*/
MANUAL_IMMEDIATE,
Does this mean calling acknowledgment.acknowledge() is not sufficient and AcknowledgingMessageListener has to be used in some way? If yes, what is the preferred approach.

You will get "up to" 500; there is no guarantee you will get exactly 500.
Yes; 16 consumers (assuming you have at least 16 partitions).
Correct.
Correct; but version 2.5 now has the RecoveringBatchErrorHandler whereby you can throw a special exception to tell it where in the batch the error occurred; it will commit the offsets of the successful records and seek the remaining ones.
The consumers get unique partitions so a consumer that is "stuck" has no impact on other consumers.
I am not sure what you are asking there; if you are calling ack.acknowledge() you are already using an AcknowledgingMessageListener (#KafkaListener always has that capability; we only populate the ack with a manual ack mode.
However, you really don't need to use manual acks for this use case; the container will commit the offsets automatically when the listener exits normally; no need to unnecessarily complicate your code.

Assert Kafka send worked

I'm writing an application with Spring Boot so to write to Kafka I do:
#Autowired
private KafkaTemplate<String, String> kafkaTemplate;
and then inside my method:
kafkaTemplate.send(topic, data)
But I feel like I'm just relying on this to work, how can I know if this has worked? If it's asynchronous, is it a good practice to return a 200 code and hoped it did work? I'm confused. If Kafka isn't available, won't this fail? Shouldn't I be prompted to catch an exception?

Along with what #mjuarez has mentioned you can try playing with two Kafka producer properties. One is ProducerConfig.ACKS_CONFIG, which lets you set the level of acknowledgement that you think is safe for your use case. This knob has three possible values. From Kafka doc
acks=0: Producer doesn't care about acknowledgement from server, and considers it as sent.
acks=1: This will mean the leader will write the record to its local log but will respond without awaiting full acknowledgement from all followers.
acks=all: This means the leader will wait for the full set of in-sync replicas to acknowledge the record.
The other property is ProducerConfig.RETRIES_CONFIG. Setting a value greater than zero will cause the client to resend any record whose send fails with a potentially transient error.

Yes, if Kafka is not available, that .send() call will fail, but if you send it async, no one will be notified. You can specify a callback that you want to be executed when the future finally finishes. Full interface spec here: https://kafka.apache.org/20/javadoc/org/apache/kafka/clients/producer/Callback.html
From the official Kafka javadoc here: https://kafka.apache.org/20/javadoc/index.html?org/apache/kafka/clients/producer/KafkaProducer.html
Fully non-blocking usage can make use of the Callback parameter to
provide a callback that will be invoked when the request is complete.
ProducerRecord<byte[],byte[]> record = new ProducerRecord<byte[],byte[]>("the-topic", key, value);
producer.send(myRecord,
new Callback() {
public void onCompletion(RecordMetadata metadata, Exception e) {
if(e != null) {
e.printStackTrace();
} else {
System.out.println("The offset of the record we just sent is: " + metadata.offset());
}
}
});

you can use below command while sending messages to kafka:
bin/kafka-console-consumer.sh --bootstrap-server localhost:9092 --topic topic-name
while above command is running you should run your code and if sending messages being successful then the message must be printed on the console.
Furthermore, likewise any other connection to any resources if the connection could not be established, then doing any kinds of operations would result some exception raises.

Why Kafka KTable is missing entries?

I have a single instance java application that uses KTable from Kafka Streams. Until recently I could retrieve all data using KTable when suddenly some of the messages seemed to vanish. There should be ~33k messages with unique keys there.
When I want to retrieve messages by key I don't get some of the messages. I use ReadOnlyKeyValueStore to retrieve messages:
final ReadOnlyKeyValueStore<GenericRecord, GenericRecord> store = ((KafkaStreams)streams).store(storeName, QueryableStoreTypes.keyValueStore());
store.get(key);
These are the configuration settings I set to the KafkaStreams.
final Properties config = new Properties();
config.put(StreamsConfig.APPLICATION_SERVER_CONFIG, serverId);
config.put(StreamsConfig.APPLICATION_ID_CONFIG, applicationId);
config.put(StreamsConfig.BOOTSTRAP_SERVERS_CONFIG, bootstrapServers);
config.put(ConsumerConfig.AUTO_OFFSET_RESET_CONFIG, "earliest");
config.put(AbstractKafkaAvroSerDeConfig.SCHEMA_REGISTRY_URL_CONFIG, schemaRegistryUrl);
config.put(StreamsConfig.KEY_SERDE_CLASS_CONFIG, GenericAvroSerde.class);
config.put(StreamsConfig.VALUE_SERDE_CLASS_CONFIG, GenericAvroSerde.class);
config.put(StreamsConfig.CACHE_MAX_BYTES_BUFFERING_CONFIG, 0);
Kafka: 0.10.2.0-cp1
Confluent: 3.2.0
Investigations brought me to some very worrying insights. Using REST Proxy I manually read partitions and found out that some offsets return error.
Request:
/topics/{topic}/partitions/{partition}/messages?offset={offset}
{
"error_code": 50002,
"message": "Kafka error: Fetch response contains an error code: 1"
}
No client, neither java nor command line however return any error. They just skip over the faulty missing messages resulting in missing data in KTables. Everything was fine and without notice it seems that somehow some of the messages got corrupt.
I have two brokers and all the topics have the replication factor of 2 and are fully replicated. Both brokers separately return the same. Restarting brokers makes no difference.
What could possibly be the cause?
How to detect this case in a client?

By default Kafka Broker config key cleanup.policy is set to delete. Set it to compact to keep the latest message for each key. See compaction.
Deletion of old messages does not change the minimum offset so trying to retrieve message below it causes an error. The error is very vague. The Kafka Streams client will start reading messages from minimum offset so there is no error. The only visible effect is missing data in KTables.
While the application is running thanks to the caches all data might still be available even after messages are deleted from Kafka itself. They will vanish after cleanup.

How to handle kafka publishing failure in robust way

I'm using Kafka and we have a use case to build a fault tolerant system where not even a single message should be missed. So here's the problem:
If publishing to Kafka fails due to any reason (ZooKeeper down, Kafka broker down etc) how can we robustly handle those messages and replay them once things are back up again. Again as I say we cannot afford even a single message failure.
Another use case is we also need to know at any given point in time how many messages were failed to publish to Kafka due to any reason i.e. something like counter functionality and now those messages needs to be re-published again.
One of the solution is to push those messages to some database (like Cassandra where writes are very fast but we also need counter functionality and I guess Cassandra counter functionality is not that great and we don't want to use that.) which can handle that kind of load and also provide us with the counter facility which is very accurate.
This question is more from architecture perspective and then which technology to use to make that happen.
PS: We handle some where like 3000TPS. So when system start failing those failed messages can grow very fast in very short time. We're using java based frameworks.
Thanks for your help!

The reason Kafka was built in a distributed, fault-tolerant way is to handle problems exactly like yours, multiple failures of core components should avoid service interruptions. To avoid a down Zookeeper, deploy at least 3 instances of Zookeepers (if this is in AWS, deploy them across availability zones). To avoid broker failures, deploy multiple brokers, and ensure you're specifying multiple brokers in your producer bootstrap.servers property. To ensure that the Kafka cluster has written your message in a durable manor, ensure that the acks=all property is set in the producer. This will acknowledge a client write when all in-sync replicas acknowledge reception of the message (at the expense of throughput). You can also set queuing limits to ensure that if writes to the broker start backing up you can catch an exception and handle it and possibly retry.
Using Cassandra (another well thought out distributed, fault tolerant system) to "stage" your writes doesn't seem like it adds any reliability to your architecture, but does increase the complexity, plus Cassandra wasn't written to be a message queue for a message queue, I would avoid this.
Properly configured, Kafka should be available to handle all your message writes and provide suitable guarantees.

I am super late to the party. But I see something missing in above answers :)
The strategy of choosing some distributed system like Cassandra is a decent idea. Once the Kafka is up and normal, you can retry all the messages that were written into this.
I would like to answer on the part of "knowing how many messages failed to publish at a given time"
From the tags, I see that you are using apache-kafka and kafka-consumer-api.You can write a custom call back for your producer and this call back can tell you if the message has failed or successfully published. On failure, log the meta data for the message.
Now, you can use log analyzing tools to analyze your failures. One such decent tool is Splunk.
Below is a small code snippet than can explain better about the call back I was talking about:
public class ProduceToKafka {
private ProducerRecord<String, String> message = null;
// TracerBulletProducer class has producer properties
private KafkaProducer<String, String> myProducer = TracerBulletProducer
.createProducer();
public void publishMessage(String string) {
ProducerRecord<String, String> message = new ProducerRecord<>(
"topicName", string);
myProducer.send(message, new MyCallback(message.key(), message.value()));
}
class MyCallback implements Callback {
private final String key;
private final String value;
public MyCallback(String key, String value) {
this.key = key;
this.value = value;
}
#Override
public void onCompletion(RecordMetadata metadata, Exception exception) {
if (exception == null) {
log.info("--------> All good !!");
} else {
log.info("--------> not so good !!");
log.info(metadata.toString());
log.info("" + metadata.serializedValueSize());
log.info(exception.getMessage());
}
}
}
}
If you analyze the number of "--------> not so good !!" logs per time unit, you can get the required insights.
God speed !

Chris already told about how to keep the system fault tolerant.
Kafka by default supports at-least once message delivery semantics, it means when it try to send a message something happens, it will try to resend it.
When you create a Kafka Producer properties, you can configure this by setting retries option more than 0.
Properties props = new Properties();
props.put("bootstrap.servers", "localhost:4242");
props.put("acks", "all");
props.put("retries", 0);
props.put("batch.size", 16384);
props.put("linger.ms", 1);
props.put("buffer.memory", 33554432);
props.put("key.serializer", "org.apache.kafka.common.serialization.StringSerializer");
props.put("value.serializer", "org.apache.kafka.common.serialization.StringSerializer");
Producer<String, String> producer = new KafkaProducer<>(props);
For more info check this.

Kafka pattern subscription. Rebalancing is not being triggered on new topic

According to the documentation on kafka javadocs if I:
Subscribe to a pattern
Create a topic that matches the pattern
A rebalance should occur, which makes the consumer read from that new topic. But that's not happening.
If I stop and start the consumer, it does pick up the new topic. So I know the new topic matches the pattern. There's a possible duplicate of this question in https://stackoverflow.com/questions/37120537/whitelist-filter-in-kafka-doesnt-pick-up-new-topics but that question got nowhere.
I'm seeing the kafka logs and there are no errors, it just doesn't trigger a rebalance. The rebalance is triggered when consumers join or die, but not when new topics are created (not even when partitions are added to existing topics, but that's another subject).
I'm using kafka 0.10.0.0, and the official Java client for the "New Consumer API", meaning broker GroupCoordinator instead of fat client + zookeeper.
This is the code for the sample consumer:
public class SampleConsumer {
public static void main(String[] args) throws IOException {
KafkaConsumer<String, String> consumer;
try (InputStream props = Resources.getResource("consumer.props").openStream()) {
Properties properties = new Properties();
properties.load(props);
properties.setProperty("group.id", "my-group");
System.out.println(properties.get("group.id"));
consumer = new KafkaConsumer<>(properties);
}
Pattern pattern = Pattern.compile("mytopic.+");
consumer.subscribe(pattern, new SampleRebalanceListener());
while (true) {
ConsumerRecords<String, String> records = consumer.poll(1000);
for (ConsumerRecord<String, String> record : records) {
System.out.printf("%s %s\n", record.topic(), record.value());
}
}
}
}
In the producer, I'm sending messages to topics named mytopic1, mytopic2, etc.
Patterns are pretty much useless if the rebalance is not triggered.
Do you know why the rebalance is not happening?

The documentation mentions "The pattern matching will be done periodically against topics existing at the time of check.". It turns out the "periodically" corresponds to the metadata.max.age.ms property. By setting that property (inside "consumer.props" in my code sample) to i.e. 5000 I can see it detects new topics and partitions every 5 seconds.
This is as designed, according to this jira ticket https://issues.apache.org/jira/browse/KAFKA-3854:
The final note on the JIRA stating that a later created topic that matches a consumer's subscription pattern would not be assigned to the consumer upon creation seems to be as designed. A repeat subscribe() to the same pattern would be needed to handle that case.
The refresh metadata polling does the "repeat subscribe()" mentioned in the ticket.
This is confusing coming from Kafka 0.8 where there was true triggering based on zookeper watches, instead of polling. IMO 0.9 is more of a downgrade for this scenario, instead of "just in time" rebalancing, this becomes either high frequency polling with overhead, or low frequency polling with long times before it reacts to new topics/partitions.

to trigger a rebalance immediately, you can explicitly make a poll call after subscribe to the topic:
kafkaConsumer.poll(pollDuration);
refer to:
https://cwiki.apache.org/confluence/display/KAFKA/KIP-568%3A+Explicit+rebalance+triggering+on+the+Consumer

In your consumer code, use the following:
properties.setProperty(ConsumerConfig.AUTO_OFFSET_RESET_CONFIG, EARLIEST)
and try again