Developing a Deeper Understanding of Apache Kafka Architecture

Print Friendly, PDF & Email

The Apache Kafka distributed streaming platform features an architecture that – ironically, given the name – provides application messaging that is markedly clearer and less Kafkaesque when compared with alternatives. In this article, we’ll take a detailed look at how Kafka’s architecture accomplishes this.

The Kafka Components – Universal Modeling Language (UML)

Kafka’s main architectural components include Producers, Topics, Consumers, Consumer Groups, Clusters, Brokers, Partitions, Replicas, Leaders, and Followers. This simplified UML diagram describes the ways these components relate to one another:

It’s important to note the relationships between broker, replica, and partition components that are highlighted, such as:

  • Kafka clusters can have one or more brokers.
  • Brokers can host multiple replicas.
  • Topics can have one or more partitions.
  • A broker can host zero or one replica per partition.
  • A partition has one leader replica and zero or more follower replicas.
  • Each replica for a partition needs to be on a separate broker.
  • Every partition replica needs to fit on a broker, and a partition can’t be divided over multiple brokers.
  • Every broker can have one or more leaders, covering different partitions and topics.

Let’s have a closer-look example of the the relationship between producers, topics, and consumers:

In this diagram, producers send to single topics:

Notice here that consumers – such as Consumer 3 in the above diagram – can be simultaneously subscribed to multiple topics and will therefore receive messages from those topics in a single poll. Received messages can be checked and filtered by topic if necessary.

In this example, a producer sends to multiple topics:

While producers can only message to one topic at a time, they’re able to send messages asynchronously. Using this technique allows a producer to functionally send multiple messages to multiple topics at once.

Because Kafka is designed for broker scalability and performance, producers (rather than brokers) are responsible for choosing which partition each message is sent to.  The default partition is determined by a hashing function on the message key, or round-robin in the absence of a key. However, this may not always provide the desired behaviour (e.g. message ordering, fair distribution of messages to consumers, etc). Producers can therefore send messages to specific partitions – through the use of a custom partitioner, or by using manual or hashing options available with the default partitioner.

Consumers Rule!

A fundamental explanation of Kafka’s inner workings goes as follows: Every topic is associated with one or more partitions, which are spread over one or more brokers. Every partition gets replicated to those one or more brokers depending on the replication factor that is set. The replication factor is then responsible for determining the reliability, while the number of partitions is responsible for the parallelism for consumers. A partition is associated with only a single consumer instance per consumer group. Since the total consumer instances per group is less than – or the same as – the number of partitions, adding support for extra consumer instances requires that more partitions be added as well, but ensures read scalability.

Because of this, events in a partition have an order to them. Within a consumer group, each event will only be processed by a single consumer. At the same time, when multiple consumer groups subscribe to a topic and have a consumer in each, every consumer will receive every message that is broadcast.

The examples below demonstrate various ways in which a single topic with multiple partitions, consumers, and consumer groups can be utilized.

Here, the same number of partitions and consumers within one group are used:

Normally, a partition can only support a single consumer in a group. This limit can be overcome, however, by manually connecting consumers to a specific partition in a topic, effectively overruling the dynamic protocol for those consumers. Such consumers should be in different groups. This has no effect on consumers that still connect dynamically:

Another scenario is to have a more consumers in a group than partitions, with certain consumers remaining idle. Kafka is able to failover to such idle consumers in cases where an active consumer dies, or when a new partition is added:

In this next example there are fewer consumers in a group that partitions, causing consumer A2 to be responsible for processing more messages than consumer A1:

Finally, this last example includes multiple consumer groups, with the result that every event from each partition gets broadcast to each group:

The Kafka protocol will dynamically perform the work of maintaining a consumer’s membership in its group. Any new instances that join the group will automatically take over partitions from other group members. And, any instances that die will have their partitions reassigned among the instances that persist.

In understanding Kafka consumers from an architectural and resourcing standpoint, it’s critical to note that consumers and producers don’t run on Kafka brokers, but they do they require CPU and IO resources of their own. This is advantageous in offering the flexibility to run consumers however, wherever, and in whatever quantity is needed without concern for deployment to brokers or sharing their resources. However, a challenge remains in that the best method for deploying and resourcing consumers and producers must be thoroughly thought out. An optimal strategy might be to enlist scalable and elastic microservices for the task.

In this article, we gained an understanding of the main Kafka components and how Kafka consumers work. In the next article, we’ll see how these contribute to the ability of Kafka to provide extreme scalability and reliability for both streaming write and read workloads.

About the Author

Paul Brebner is Tech Evangelist at Instaclustr, which provides a managed service platform of open source technologies such as Apache Cassandra, Apache Spark, Elasticsearch and Apache Kafka.


Sign up for the free insideBIGDATA newsletter.

Speak Your Mind



  1. Marcel Bricman says

    “a topic is replicated over 1 or more partitions” – this is incorrect. a partition is not a duplication-device, it is a splitting-device. the text should rather read “a topic is split into 1 or more partitions”.

  2. Alain Christian Moune says

    Thank you very much for the Great post, it should be unique in the way it presents Kafka.

    And concerning the previous comment (Marcel), the topic can indeed be replicated over 1 or more partitions, for high availability.

    Let’s say you have 2 partitions PartA and PartB. The topic will first be split in let’s say Set1 and Set2, and each partition will receive its set, PartA -> Set1, and PartB -> Set2. But there is no redundancy here, If the broker owning PartA goes down you loose data. So base on your replication config, Kafka can maintain a copy of Set1 in PartB, and of Set2 in PartA. So you will end up with this: PartA -> (Set1, Set2_copy), and PartB -> (Set2, Set1_copy).

    In this post, Set1 is called the Learder replica, and Set2_copy is called the Follower replica.
    So a topic can indeed be replicated over 1 or more partitions.

  3. Universal Modeling Language (UML)?
    U should be Unified.

  4. I need help with understanding for designing of topics like when should we add events to same topic and when should we create a new topic? When do we add to the same partition in single topic? Any help would be appreciated