Kafka QuickStart, advertised.host.name gives kafka.common.LeaderNotAvailableException

Question

I am able to get a simple one-node Kafka (kafka_2.11-0.8.2.1) working locally on one linux machine, but when I try to run a producer remotely I'm getting some confusing errors.

I'm following the quickstart guide at http://kafka.apache.org/documentation.html#quickstart. I stopped the kafka processes and deleted all the zookeeper & karma files in /tmp. I am on a local 10.0.0.0/24 network NAT-ed with an external IP address, so I modified server.properties to tell zookeeper how to broadcast my external address, as per https://medium.com/@thedude_rog/running-kafka-in-a-hybrid-cloud-environment-17a8f3cfc284:

advertised.host.name=MY.EXTERNAL.IP

Then I'm running this:

$ bin/zookeeper-server-start.sh config/zookeeper.properties
--> ...
$ export KAFKA_HEAP_OPTS="-Xmx256M -Xms128M" # small test server!
$ bin/kafka-server-start.sh config/server.properties
--> ...

I opened up the firewall for my producer on the remote machine, and created a new topic and verified it:

$ bin/kafka-topics.sh --create --zookeeper MY.EXTERNAL.IP:2181 --replication-factor 1 --partitions 1 --topic test123
--> Created topic "test123".
$ bin/kafka-topics.sh --list --zookeeper MY.EXTERNAL.IP:2181
--> test123

However, the producer I'm running remotely gives me errors:

$ bin/kafka-console-producer.sh --broker-list MY.EXTERNAL.IP:9092 --topic test123
--> [2015-06-16 14:41:19,757] WARN Property topic is not valid (kafka.utils.VerifiableProperties)
My Test Message
--> [2015-06-16 14:42:43,347] WARN Error while fetching metadata [{TopicMetadata for topic test123 -> 

No partition metadata for topic test123 due to kafka.common.LeaderNotAvailableException}] for topic [test123]: class kafka.common.LeaderNotAvailableException (kafka.producer.BrokerPartitionInfo) --> (repeated several times)

(I disabled the whole firewall to make sure that wasn't the problem.)

The stdout errors in the karma-startup are repeated: [2015-06-16 20:42:42,768] INFO Closing socket connection to /MY.EXTERNAL.IP. (kafka.network.Processor)

And the controller.log gives me this, several times:

java.nio.channels.ClosedChannelException
    at kafka.network.BlockingChannel.send(BlockingChannel.scala:100)
    at kafka.controller.RequestSendThread.liftedTree1$1(ControllerChannelManager.scala:132)
    at kafka.controller.RequestSendThread.doWork(ControllerChannelManager.scala:131)
    at kafka.utils.ShutdownableThread.run(ShutdownableThread.scala:60)
[2015-06-16 20:44:08,128] INFO [Controller-0-to-broker-0-send-thread], Controller 0 connected to id:0,host:MY.EXTERNAL.IP,port:9092 for sending state change requests (kafka.controller.RequestSendThread)
[2015-06-16 20:44:08,428] WARN [Controller-0-to-broker-0-send-thread], Controller 0 epoch 1 fails to send request Name:LeaderAndIsrRequest;Version:0;Controller:0;ControllerEpoch:1;CorrelationId:7;ClientId:id_0-host_null-port_9092;Leaders:id:0,host:MY.EXTERNAL.IP,port:9092;PartitionState:(test123,0) -> (LeaderAndIsrInfo:(Leader:0,ISR:0,LeaderEpoch:0,ControllerEpoch:1),ReplicationFactor:1),AllReplicas:0) to broker id:0,host:MY.EXTERNAL.IP,port:9092. Reconnecting to broker. (kafka.controller.RequestSendThread)

Running this seems to indicate that there is a leader at 0:

$ ./bin/kafka-topics.sh --zookeeper MY.EXTERNAL.IP:2181 --describe --topic test123
--> Topic:test123   PartitionCount:1    ReplicationFactor:1 Configs:
Topic: test123  Partition: 0    Leader: 0   Replicas: 0 Isr: 0

I reran this test and my server.log indicates that there is a leader at 0:

...
[2015-06-16 21:58:04,498] INFO 0 successfully elected as leader (kafka.server.ZookeeperLeaderElector)
[2015-06-16 21:58:04,642] INFO Registered broker 0 at path /brokers/ids/0 with address MY.EXTERNAL.IP:9092. (kafka.utils.ZkUtils$)
[2015-06-16 21:58:04,670] INFO [Kafka Server 0], started (kafka.server.KafkaServer)
[2015-06-16 21:58:04,736] INFO New leader is 0 (kafka.server.ZookeeperLeaderElector$LeaderChangeListener)

I see this error in the logs when I send a message from the producer:

[2015-06-16 22:18:24,584] ERROR [KafkaApi-0] error when handling request Name: TopicMetadataRequest; Version: 0; CorrelationId: 7; ClientId: console-producer; Topics: test123 (kafka.server.KafkaApis)
kafka.admin.AdminOperationException: replication factor: 1 larger than available brokers: 0
    at kafka.admin.AdminUtils$.assignReplicasToBrokers(AdminUtils.scala:70)

I assume this means that the broker can't be found for some reason? I'm confused what this means...

Scarificator

I've had this before. We set a replication factor of 3 on our topics and have 3 brokers, however one of our brokers was not working because we had accidentally given it the same broker ID as one of the other ones. We changed the broker ID and restarted and this fixed the issue. – Sloppy

@JonHunter It's strange because I only have one broker and one topic. I set this up the same way on EC2 last night to see if it was related to my hosting setup but I had the same error there too. – Scarificator

Could you try changing your broker ID. I've not used a broker ID of 0 before (even though the Kafka documentation says that is OK). Also you can use the Zookeeper client "zkCli.sh" which will show you which brokers are registered with Zookeeper (it allows you to browse through the Zookeeper znode structure). – Sloppy

Ok, I've finally figured this out---and it's my own stupid firewall/routing error. The problem is that the I didn't explicitly allow zookeeper to access the broker via the external IP address, even though it was on the same machine. So it worked fine when "advertised.host.name" was unconfigured because zookeeper was using localhost. And it meant that I could connect fine from the remote client to the server and zookeeper, but it wasn't communicating internally. Sigh, at least I learned a lot about kafka trying to figure that out. – Scarificator

Thanks for the help, BTW. I did confirm that using a broker of "0" works, though. – Scarificator

Welcome to the world of Kafka, I've learnt so far that it's all about trial and error as there's very little support documentation on the web! – Sloppy

@JonHunter Yes, I'm a little concerned. – Scarificator

Answer 1

For the recent versions of Kafka (0.10.0 as of this writing), you don't want to use advertised.host.name at all. In fact, even the [documentation] states that advertised.host.name is already deprecated. Moreover, Kafka will use this not only as the "advertised" host name for the producers/consumers, but for other brokers as well (in a multi-broker environment)...which is kind of a pain if you're using using a different (perhaps internal) DNS for the brokers...and you really don't want to get into the business of adding entries to the individual /etc/hosts of the brokers (ew!)

So, basically, you would want the brokers to use the internal name, but use the external FQDNs for the producers and consumers only. To do this, you will update advertised.listeners instead.

Nader

Answer 2

Set advertised.host.name to a host name, not an IP address. The default is to return a FQDN using getCanonicalHostName(), but this is only best effort and falls back to an IP. See the java docs for getCanonicalHostName().

The trick is to get that host name to always resolve to the correct IP. For small environments I usually setup all of the hosts with all of their internal IPs in /etc/hosts. This way all machines know how to talk to each other over the internal network, by name. In fact, configure your Kafka clients by name now too, not by IP. If managing all the /etc/hosts files is a burden then setup an internal DNS server to centralize it, but internal DNS should return internal IPs. Either of these options should be less work than having IP addresses scattered throughout various configuration files on various machines.

Once everything is communicating by name all that's left is to configure external DNS with the external IPs and everything just works. This includes configuring Kafka clients with the server names, not IPs.

Cetus

Answer 3

So to summarize, the solution to this was to add a route via NAT so that the machine can access its own external IP address.

Zookeeper uses the address it finds in advertised.host.name both to tell clients where to find the broker as well as to communicate with the broker itself. The error that gets reported doesn't make this very clear, and it's confusing because a client has no problem opening a TCP connection.

Scarificator

Would this work if you're using multiple brokers (= multiple kafka containers running behind the same advertised hostname)? – Melinite

Answer 4

Taking cue from above: for my single node (while still learning) I modified server.properties file having text "advertised.host.name" to value=127.0.01. So finally it looks something like this

advertised.host.name=127.0.0.1

While starting producer it still shows warning, but now it is atleast working while I can see messages on consumer terminal perfectly comming

Brooklime

Answer 5

On your machine where Kafka is installed, check if it is up and running. The error states, 0 brokers are available that means Kafka is not up and running.

On linux machine you can use the netstat command to check if the service is running.

netstat -an|grep port_kafka_is_Listening ( default is 9092)

Drench

Yes, it is indeed running and listening on port 9092. I can connect to kafka with no problem, it's just that it gives me errors on both the producer and the server when I produce a message. It is easy to reproduce this error on EC2. – Scarificator

Answer 6

conf/server.properties: host.name DEPRECATED: only used when listeners is not set. Use listeners instead. hostname of broker. If this is set, it will only bind to this address. If this is not set, it will bind to all interfaces

Pitchblack