[KAFKA-2135] New Kafka Producer Client: Send requests wait indefinitely if no broker is available.

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Duplicate
• Affects Version/s: 0.8.2.0
• Component/s: producer
• Assignee: Jun Rao
• Reporter: David Hay
• Created: 20/Apr/15 18:04
• Updated: 04/May/15 22:26
• Resolved: 04/May/15 22:26

Description

I'm seeing issues when sending a message with the new producer client API. The future returned from Producer.send() will block indefinitely if the cluster is unreachable for some reason. Here are the steps:

1. Start up a single node kafka cluster locally.
2. Start up application and create a KafkaProducer with the following config:

   KafkaProducerWrapper values: 
   	compression.type = snappy
   	metric.reporters = []
   	metadata.max.age.ms = 300000
   	metadata.fetch.timeout.ms = 60000
   	acks = all
   	batch.size = 16384
   	reconnect.backoff.ms = 10
   	bootstrap.servers = [localhost:9092]
   	receive.buffer.bytes = 32768
   	retry.backoff.ms = 100
   	buffer.memory = 33554432
   	timeout.ms = 30000
   	key.serializer = class com.mycompany.kafka.serializer.ToStringEncoder
   	retries = 3
   	max.request.size = 1048576
   	block.on.buffer.full = true
   	value.serializer = class com.mycompany.kafka.serializer.JsonEncoder
   	metrics.sample.window.ms = 30000
   	send.buffer.bytes = 131072
   	max.in.flight.requests.per.connection = 5
   	metrics.num.samples = 2
   	linger.ms = 0
   	client.id = site-json
   

3. Send some messages...they are successfully sent
4. Shut down the kafka broker
5. Send another message.

At this point, calling get() on the returned Future will block indefinitely until the broker is restarted.

It appears that there is some logic in org.apache.kafka.clients.producer.internal.Sender that is supposed to mark the Future as "done" in response to a disconnect event (towards the end of the run(long) method). However, the while loop earlier in this method seems to remove the broker from consideration entirely, so the final loop over ClientResponse objects is never executed.

It seems like "timeout.ms" configuration should be honored in this case, or perhaps introduce another timeout, indicating that we should give up waiting for the cluster to return.

Issue Links

• duplicates: KAFKA-1788 producer record can stay in RecordAccumulator forever if leader is no available