Kafka (CCDAK)

#infrastructure #scalability #queues

What is Apache Kafka?

Kafka is a distributed streaming platform. This means that kafka enables you to:

Kafka is written in java and it was created by linkedin in 2011

Kafka use cases

Messaging - building real-time streaming data pipelines that reliably get data between systems or applications

Kafka is mainly used by 2 types of applications:

Benefits

Main concepts and terminology

Events - Also known as record or message, they store information of an event that happened in your applications. An event has a key, value, timestamp and optional metadata headers.

key: alice
value: "Made a payment of 200 to bob"
timestamp: "Jun. 25, 2020 at 2:06 p.m"

Producers - Are applications that publish events to kafka
Offset - A Sequential and unique Id of event in a partition

Consumers - are applications that subscribe to kafka.
Consumer group - By default consumers will re-consume the same records. Consumer groups allow to change this behaviour, allowing for consumers to only consume different events/partitions.

Garantees - There are multiple types of delivery garantees that could be provided:

Topics - Events are stored in topics. Topics in Kafka are always multi-producer and multi-subscriber - They can have zero, one or many producers or consumers. Events can be read as often as needed, since they are not deleted on consumption. Deletion is defined by a configuration which defines how long a topic should retain messages.

Partition - Topics are partitioned, meaning that a topic is spread over a number of buckets located in different Kafka Brokers. Wen a new event is published to a topic it is appended to a topic partition. Events with the same key (eg. customer id) are written to the same partition and kafka garantees that consumers will read events in order.

kafka_partition.png

Log - Immutable and partitioned Data structure used to store a sequence of events

Replication - To make your data fault tolerant and highly available every topic can be replicated in different brokers. A common configuration is a replication factor of 3.

Architecture

Brokers are servers that make a Kafka Cluster. Producers and consumers communicate with them to publish and consume messages.

Kafka depends on zookeeper to manage clusters. It coordinates cluster communication, adds, removes and monitor brokers/nodes.

Kafka elects a broker to be the Controller, which coordinates assigning partitions and replicas to other nodes. If the controller goes down another one will be elected.

Kafka uses TCP to handle message communication

Design

https://kafka.apache.org/documentation/#design

Partitions and replication

A partition is assigned to an specific broker and to achieve fault tolerance this partition is replicated to other clusters. The replication factor is the number of replicas that will kept in different brokers.

To ensure consistency kafka chooses a leader for each partition, the leader handles all reads and writes for the partition. If the leader is down the cluster tries to choose a new one by leader election.

By default Kafka will only set as electable replicas that are in sync (ISR), if there are no ISR Kafka will just wait until one becomes available. It is possible to configure this behaviour so it will work as unclean leader election.

Kafka Streams

Work as Node.js Transformation streams (get an input and pipe to an output).
There are 2 types of transformations in Kafka Streams:

Stateless Transformations

Stateful Transformations