Apache Kafka on Docker

This repository holds a build definition and supporting files for building a Docker image to run Kafka in containers. It is published as an Automated Build on the Docker registry, as p1ck/kafka.

This build intends to provide an operator-friendly Kafka deployment suitable for usage in a production Docker environment:

• It runs one service, no bundled ZooKeeper (for more convenient development, use Docker Compose!).
• Configuration is parameterized, enabling a Kafka cluster to be run from multiple container instances.
• Kafka data and logs can be handled outside the container(s) using volumes.
• JMX is exposed, for Kafka and JVM metrics visibility.

If you find any shortcomings with the build regarding operability, pull requests or feedback via GitHub issues are welcomed.

Usage Quick Start

Here is a minimal-configuration example running the Kafka broker service, then using the container as a client to run the basic producer and consumer example from the Kafka Quick Start:

$ docker run -d --name zookeeper jplock/zookeeper:3.4.6
$ docker run -d --name kafka --link zookeeper:zookeeper p1ck/kafka

$ ZK_IP=$(docker inspect --format '{{ .NetworkSettings.IPAddress }}' zookeeper)
$ KAFKA_IP=$(docker inspect --format '{{ .NetworkSettings.IPAddress }}' kafka)

$ docker run --rm p1ck/kafka \
>   kafka-topics.sh --create --topic test --replication-factor 1 --partitions 1 --zookeeper $ZK_IP:2181
Created topic "test".

# In separate terminals:
$ docker run --rm --interactive p1ck/kafka \
>   kafka-console-producer.sh --topic test --broker-list $KAFKA_IP:9092
<type some messages followed by newline>

$ docker run --rm p1ck/kafka \
>   kafka-console-consumer.sh --topic test --from-beginning --zookeeper $ZK_IP:2181

Volumes

The container exposes two volumes that you may wish to bind-mount, or process elsewhere with --volumes-from:

• /data: Path where Kafka's data is stored (log.dirs in Kafka configuration)
• /logs: Path where Kafka's logs (INFO level) will be written, via log4j

Ports and Linking

The container publishes two ports:

• 9092: Kafka's standard broker communication
• 7203: JMX publishing, for e.g. jconsole or VisualVM connection

Kafka requires Apache ZooKeeper. You can satisfy the dependency by simply linking another container that exposes ZooKeeper on its standard port of 2181, as shown in the above example, ensuring that you link using an alias of zookeeper.

Alternatively, you may configure a specific address for Kafka to find ZK. See the Configuration section below.

A more complex local development setup

This example shows more configuration options and assumes that you wish to run a development environment with Kafka ports mapped directly to localhost, for instance if you're writing a producer or consumer and want to avoid rebuilding a container for it to run in as you iterate. This requires that localhost is your Docker host, i.e. your workstation runs Linux. If you're using something like boot2docker, substitute the value of boot2docker ip below.

$ mkdir -p kafka-ex/{data,logs} && cd kafka-ex
$ docker run -d --name zookeeper --publish 2181:2181 jplock/zookeeper:3.4.6
$ docker run -d \
    --hostname localhost \
    --name kafka \
    --volume ./data:/data --volume ./logs:/logs \
    --publish 9092:9092 --publish 7203:7203 \
    --env KAFKA_ADVERTISED_HOST_NAME=127.0.0.1 --env ZOOKEEPER_IP=127.0.0.1 \
    p1ck/kafka

Configuration

Some parameters of Kafka configuration can be set through environment variables when running the container (docker run -e VAR=value). These are shown here with their default values, if any:

• KAFKA_BROKER_ID=0

  Maps to Kafka's broker.id setting. Must be a unique integer for each broker in a cluster.

• KAFKA_PORT=9092

  Maps to Kafka's port setting. The port that the broker service listens on. You will need to explicitly publish a new port from container instances if you change this.

• KAFKA_ADVERTISED_HOST_NAME=<container's IP within docker0's subnet>

  Maps to Kafka's advertised.host.name setting. Kafka brokers gossip the list of brokers in the cluster to relieve producers from depending on a ZooKeeper library. This setting should reflect the address at which producers can reach the broker on the network, i.e. if you build a cluster consisting of multiple physical Docker hosts, you will need to set this to the hostname of the Docker host's interface where you forward the container KAFKA_PORT.

• KAFKA_ADVERTISED_PORT=9092

  As above, for the port part of the advertised address. Maps to Kafka's advertised.port setting. If you run multiple broker containers on a single Docker host and need them to be accessible externally, this should be set to the port that you forward to on the Docker host.

• JAVA_RMI_SERVER_HOSTNAME=$KAFKA_ADVERTISED_HOST_NAME

  Maps to the java.rmi.server.hostname JVM property, which is used to bind the interface that will accept remote JMX connections. Like KAFKA_ADVERTISED_HOST_NAME, it may be necessary to set this to a reachable address of the Docker host if you wish to connect a JMX client from outside of Docker.

• ZOOKEEPER_IP=<taken from linked "zookeeper" container, if available>

  Required if no container is linked with the alias "zookeeper" and publishing port 2181, or not using ZOOKEEPER_CONNECTION_STRING instead. Used in constructing Kafka's zookeeper.connect setting.

• ZOOKEEPER_PORT=2181

  Used in constructing Kafka's zookeeper.connect setting.

• ZOOKEEPER_CONNECTION_STRING=<comma separated string of host:port pairs>

  Set a string with host:port pairs for connecting to a ZooKeeper Cluster. This setting overrides ZOOKEEPER_IP and ZOOKEEPER_PORT.

• ZOOKEEPER_CHROOT, ex: /v0_8_1

  ZooKeeper root path used in constructing Kafka's zookeeper.connect setting. This is blank by default, which means Kafka will use the ZK /. You should set this if the ZK instance/cluster is shared by other services, or to accommodate Kafka upgrades that change schema. Starting in Kafka 0.8.2, it will create the path in ZK automatically; with earlier versions, you must ensure it is created before starting brokers.

JMX

Remote JMX access can be a bit of a pain to set up. The start script for this container tries to make it as painless as possible, but it's important to understand that if you want to connect a client like VisualVM from outside other Docker containers (e.g. directly from your host OS in development), then you'll need to configure RMI to be addressed as the Docker host IP or hostname. If you have set KAFKA_ADVERTISED_HOST_NAME, that value will be used and is probably what you want. If not (you're only using other containers to talk to Kafka brokers) or you need to override it for some reason, then you can instead set JAVA_RMI_SERVER_HOSTNAME.

For example in practice, if your Docker host is VirtualBox run by Docker Machine, a run command like this should allow you to connect VisualVM from your host OS to $(docker-machine ip docker-vm):7203:

$ docker run -d --name kafka -p 7203:7203 \
    --link zookeeper:zookeeper \
    --env JAVA_RMI_SERVER_HOSTNAME=$(docker-machine ip docker-vm) \
    p1ck/kafka

Note that it is fussy about port as well---it may not work if the same port number is not used within the container and on the host (any advice for workarounds is welcome).

Finally, please note that by default remote JMX has authentication and SSL turned off (these settings are taken from Kafka's own default start scripts). If you expose the JMX hostname/port from the Docker host in a production environment, you should make make certain that access is locked down appropriately with firewall rules or similar. A more advisable setup in a Docker setting would be to run a metrics collector in another container, and link it to the Kafka container(s).

If you need finer-grained configuration, you can totally control the relevant Java system properties by setting KAFKA_JMX_OPTS yourself---see start.sh.

Fork Legacy

This image is forked from ches/kafka but has been rebased on a different Linux distribution and Java JRE. (openjdk:8-jre-alpine) There is probably no interest in merging these changes upstream, since it is just an alternative installation which does not improve on any specific issues in the original.

The ches/kafka image/repo was originally forked from relateiq/kafka. My original motivations for forking were:

• Change the Kafka binary source to an official Apache artifact. RelateIQ's was on a private S3 bucket, and this opaqueness is not suitable for a publicly-shared image for reasons of trust.
• Changes described in this pull request.

After a period of unresponsiveness from upstream on pull requests and my repo tallying far more downloads on Docker Hub, I have made further updates and changes with the expectation of maintaining independently from here on. This project's changelog file describes these in detail.