Hi there, I have probably found a bug in Java Kafka producer client.
Scenario & current behavior:
- Start Kafka broker, single instance.
- Start application that produces messages to Kafka.
- Let the application to load partitions for a topic to warm up the producer, e.g. send a message to Kafka. I'm not sure if this is necessary step, but our code does it.
- Gracefully stop the Kafka broker.
- Application logs now contains "org.apache.kafka.clients.NetworkClient: [Producer clientId=...] Connection to node 0 could not be established. Broker may not be available." so the client is aware about the Kafka unavailability.
- Trigger the producer to send a message using KafkaProducer.send(ProducerRecord, Callback) method.
- The callback that notifies business code receives non-null RecordMetadata and null Exception after request.timeout.ms. The metadata contains offset -1 which is value of ProduceResponse.INVALID_OFFSET.
- If the Kafka is not running and the message is not appended to the log, the callback should contain null RecordMetadata and non-null Exception. At least I subjectively understand the Javadoc this way, "exception on production error" in simple words.
- Developer that is not aware of this behavior and that doesn't test for offset -1, may consider the message as successfully send and properly acked by the broker.
- Together with checking for non-null exception in the callback, add another condition for ProduceResponse.INVALID_OFFSET.
- Latest Kafka 2.0.0 for both broker and Java client.
- Originally found with broker 0.11.0.1 and client 2.0.0.
- Code is analogy of the one in Javadoc of KafkaProducer.send().
- Used producer configuration (others use defaults).