Introduction to Kafka and Event-Driven

Introduction to Kafka
and Event-Driven

Inter-process communication- Previous Slides
Recap
one-to-one one-to-many
Synchronous Request / Response -
Asynchronous Asynchronous Request /
Response
One way notification
Publish / Subscribe
Publish / Async responses

Asynchronous Messaging
Two types of async messaging:
● Brokerless
● Broker-based (Kafka, RabbitMQ, AWS
kinesis…)
The message represents a change of the state, and
is triggered when a user or service performs an
action.

Brokers
● Point to Point (1-to-1): message is sent from
one producer to a consumer via a queue (e.g.
AWS SQS). Multiple consumers can listen on the
queue but only one of them will get the
message.
● Publish/Subscribe (1-to-many): message is sent
from one producer to multiple
consumers/subscribers via a topic (e.g. Kafka,
AWS Kinesis). The subscribers may or may not
acknowledge the published message

What is Kafka?
● An event streaming platform implemented as “distributed commit log”.
● Uses an append-only commit log which provides a durable and replayable record.
● Designed to provide protection against failures and scaling performance.
● Emphasis on high throughput stream processing.
● Is a database inside out, a tool for storing data, processing it in real time, and creating views,
optimized for continual computation rather than batch processing.
● Real-time processing first, long-term storage second.

Kafka Main Concepts
● Events have a key (optional), value, timestamp and metadata headers (optional).
● Events are written to topics for a configurable period of time called retention period.
● Topics are always multi-producer and multi-consumer and can be read as often as needed.
● Topics are partitioned and spread over a number of partitions (just a bucket located on different
Kafka brokers).
● Events with the same key are always written to same the partition.
● For efficiency, messages are written into batches (collection of messages produced to same topic
and partition)
● Kafka guarantees event time-ordering inside a partition.

Kafka Main Concepts 2
● Partitions enable high throughput and scalability.
● Kafka enforces durability by replicating partitions.
● Leader replica where all producer / consumer requests go through.
● Follower replicas which replicate messages from the leader and aim to stay in-sync.
● Events written to the partition leader are not immediately readable by consumers.
● Only when all in-sync replicas have acknowledged the write the message is considered committed.
● This ensures that messages cannot be lost by a broker failure.
● Partition replication provides data fault tolerance and high-availability.

Kafka How It works - High Level
● Producers send messages to a Kafka broker.
● Messages are stored in a topic.
● Consumers read messages by subscribing to topics.
● Topics are splitted into partitions and are append-only.
● All messages inside a partition are ordered in the sequence
they came in.
● Kafka follows the principle of a dumb broker and smart
consumer, as doesn’t keep track of what records are read by
the consumer.
● Consumers use offset to keep track of the latest consumed
message.

Kafka Broker Concepts (1)
● Broker is a single Kafka server allocated in a cluster (can contain multiple).
● The goal of broker is to:
○ receive messages from producers.
○ assign offsets to messages.
○ store the messages to disk.
○ respond to consumers requests for messages in specific partition.

Kafka Broker Concepts (2)
● One broker is elected as controller and is responsible for
○ admin operations
○ assign partitions
○ electing partition leader
○ monitor for broker failures
● When controller notices that a broker left cluster:
○ goes over all partitions who need new leader
○ finds new leader
○ sends request to all broker with the list of new leaders and followers replicas.

Kafka Broker Configuration
Let’s explore some core broker configurations:
1. broker.id -> Every broker must have an unique integer identifier in a single cluster.
2. port -> Listener on TCP port default value 9092.
3. logs.dirs -> Directories used to persist data on disk.
4. auto.create.topics.enable -> Specifies if the broker can automatically create topics.

Kafka Topic Configuration
Let’s explore some core topic configurations.
1. num.partitions -> Defines how many partitions per topic. Number of partitions can only be
increased, never decreased.
2. log.retention.ms -> Defines how long Kafka will keep messages. Default value 1 week.
3. log.retention.bytes -> Defines the max number of bytes of messages per partition, before they are
considered as expired.
4. message.max.bytes -> Defines the max size of a message. Default value is 1MB.

Kafka Producer - Concepts (1)
● Emits messages (ProducerRecord) to Kafka.
● ProducerRecord is delivered to Partitioner, which
determines the message partition.
● If the key is not specified, Partitioner chooses round-robin
algorithm to determine the partition, otherwise uses key
hash to determine the partition.
● Partitioner adds ProducerRecord to a batch, which will be
delivered to same topic and partition.
● Broker receives records and responses back with success
or failure.
● In case of failure, if retry is possible, re-sends records
otherwise throws an exception.

Kafka Producer - Concepts(2)
There are 3 ways to send message to Kafka via Producer API:
1. Fire & Forget: Using send() we emit a message,
without taking waiting broker response - if message
was sent successfully or not.
2. Sync: Using send().get() we are waiting (blocking) for
a reply from the broker. If there are any non-retriable
errors we will get them in the catch block.
3. Async: Adding a callback in send() method, we receive
an asynchronous response (non-blocking) from the
broker, if the message was sent with success or not.

Kafka Producer - Configuration (1)
Let’s explore some required producer configurations.
1. bootstrap.servers -> List of brokers, so that the producer can find the Kafka cluster. Example
host:port
2. key.serializer -> Serializer class for ProducerRecord’s key which implements the
org.apache.kafka.common.serialization.Serializer interface.
3. value.serializer -> Serializer class for ProducerRecord’s value that implements the
org.apache.kafka.common.serialization.Serializer interface.

Let’s explore some core producer configurations.
4. client.id -> Allows to easily correlate requests on the broker with the client instance which made it
5. acks -> Controls the durability of messages written to Kafka
a. acks=0: Producer will not wait for a reply from the broker if the message was sent
successfully. No guarantee that the message was successfully written to the broker.
b. acks=1 (default): Producer is waiting until receives a success response from the partition
leader that the write succeeded. Message still can get lost if leader crashes and a follower
replica which is promoted to leader does not have the message.
c. acks=all: Strongest guarantee, as producer is waiting until the message is successfully
written in the partition leader and replicated to all of the in-sync replicas. Increases latency.

6. retries -> Controls how many times Producer will retry in case of error.
7. enable.idempotence -> When set to 'true', the producer will ensure that exactly one copy of each
message is written in the stream.
8. max.in.flight.requests.per.connection -> Max number of unacknowledged requests the client will
send on a single connection before blocking. Note that if this config is set to be greater than 1 and
enable.idempotence is set to false, there is a risk of message re-ordering after a failed send due to
retries (i.e., if retries are enabled).
9. batch.size -> Controls amount in bytes for each batch.

Kafka Consumer - Concepts (1)
● Reads messages from Kafka topics.
● An offset (increasing integer value) is used to track which messages are already consumed.
● Each message in a specific partition has a unique offset.
● Consumer group is a set of consumers. Each partition is assigned and consumed only by one
consumer.
● Every consumer group has a group coordinator which is one of the brokers.
● Group coordinator is responsible for managing the members of the group and partition’s
assignments.

Kafka Consumer - Concepts (2)
● Can have multiple apps / clients read same topics. Each app gets all topic’s messages.
● When a new consumer is added to the group or a consumer crashes, then a rebalance process
starts.
● During rebalance stage, partition ownership is moving from one consumer to another and can not
consume messages.
● Sends heartbeats to Kafka broker (group’s coordinator), consumers maintain membership in group
and partition ownership.
● If consumer stops sending heartbeats, group coordinator consider it as dead and triggers rebalance.

Kafka Consumer - Poll Loop
1. When the consumer starts up, it finds the group coordinator
and requests to join the group.
2. After the consumer’s assignment to its group by the
coordinator, consumer must find the initial position (offset)
for each assigned partition.
3. After consumer starts reading messages from the assigned
partitions, it must commit the offsets for the messages it
has read.
4. The offset commit policy is important to ensure the
message delivery guarantee.
5. To maintain group membership and partition ownership, the
consumer sends heartbeats to coordinator regularly.

Kafka Consumer - Configuration (1)
Let’s explore some required consumer configurations.
1. bootstrap.servers -> List of brokers, so that the consumer can find the Kafka cluster. Example
host:port
2. key.deserializer-> Deserializer class for record’s key which implements the
org.apache.kafka.common.serialization.Deserializer interface.
3. value.deserializer -> Deserializer class for record’s value that implements the
org.apache.kafka.common.serialization.Deserializer interface.
4. group.id -> A unique string that identifies the consumer group this consumer belongs to.

Kafka Consumer - Configuration (2)
Let’s explore some core consumer configurations.
5. enable.auto.commit -> Controls if the consumer will commit offset automatically or not. Using auto
commit gives us “at least once” delivery, as Kafka guarantees that no messages will be missed, but
duplicates are possible.
6. auto.commit.interval.ms -> The frequency in milliseconds that the consumer offsets are auto-
committed to Kafka if enable.auto.commit is set to true.
7. fetch.min.bytes -> Specifies the minimum amount of data received from the broker.
8. heartbeat.interval.ms -> The expected time between heartbeats to the consumer coordinator when
using Kafka's group management facilities.
9. auto.offset.reset -> Specifies what to do when there is no initial offset in Kafka or if the current
offset does not exist any more on the server. E.g. earliest: automatically reset the offset to the
earliest offset, while latest: automatically reset the offset to the latest offset.

Kafka Consumer - Commits & Offsets (1)
● To commit an offset, the consumer emits a message to Kafka on a special topic called
__consumer_offsets, for the read messages for each partition.
● A rebalance will be triggered if a consumer joins or crashes, and the new one will read the last
committed offset for its assigned partition.
● Rebalance has two phases: partition revocation and partition assignment.
○ The revocation is always called before a rebalance and is our last chance to commit offsets
before partitions reassignment.
○ The assignment method can be used to set the initial position of the assigned partitions.
● Messages can be processed twice (duplication), if the committed offset is smaller than the last
message offset.

A consumer can commit offset using 3 different ways:
1. Automatic Commit: (at least once delivery)
a. Set enable.auto.commit = true
b. By default, every 5 secs, the largest offset received from poll() is automatically committed.
c. We can select the time-window for auto committing offset using auto.commit.interval.ms.
d. Auto-commit works as a cron job and the auto.commit.interval.ms config sets the time-
window for triggering.
e. If a consumer crashes before auto.commit.interval.ms period, a rebalance will be triggered,
and new consumer’s offset will be older than actual message offset, which leads to
duplications, as it will re-read same messages.
f. Automatic commits are convenient, but give no enough control on reducing duplications.

2. Sync Commit:
a. Set enable.auto.commit = false
b. commitSync() commits latest offset
returned by poll() synchronously.
c. This method should be called after
processing all records, to avoid missing
messages.
d. The consumer is blocked until receives
response from broker.
e. If an error occurs, consumer will retry to
commit until either succeeds or fails
with non-retriable failure.

3. Async Commit:
a. Set enable.auto.commit = false
b. commitAsync() commits latest offset
returned by poll() asynchronously.
c. This method should be called after
processing all records, to avoid missing
messages.
d. Consumer sends the request and return
immediately without blocking.
e. Using commitAsync(), we are not able to
retry if case of retryable error.

● There is still the chance of duplication when using
commitAsync() or commitSync(). E.g. if app crashes after
record’s process but before commit offset. In this case the
same messages will be re-processed.
● A way to remove duplicates, is to use a database table
which keeps track of events that have already been
consumed and avoid reprocessing them again.
● Also to reduce duplicates, we can store offset in a
database’s transaction with processed events. This way is
an atomic operation and both or neither can happen. We can
use consumer.seek() method to fetch specific offset from db
and set it on consumer.

Kafka Streams Concepts (1)
● A client library for stream processing applications built on top of Apache Kafka.
● Processes records from Kafka topics.
● Stream processing is the ongoing, and record-by-record real-time processing of data.
● Wrapper around Kafka Producer and Consumer.
● All Kafka topics are stored as a stream.
● It supports as primitives KStreams and KTables types and high level Streams DSL (KSQL).
○ KStreams: provides immutable data. It supports only inserting (appending) new events and It’s a never-ending
flow of data in a stream
○ KTables: is just an abstraction of the stream, where only the latest value is kept and allows for fast key
lookups. It always automatically updates latest state if new value arrives on steam
○ A KStream can be interpreted as a KTable, and a KTable can be interpreted as a KStream.
● Topology describes processing flow defined a graph of stream processors (nodes) that are connected by
streams (edges)

● Can be easily embedded in any Java / Kotlin / Scala application.
● Fully integrated with Kafka security.
● No external dependencies on systems other than Apache Kafka itself.
● Supports fault-tolerant local state for stateful operations (joins, window aggregations).
● One-record-at-a-time processing to achieve millisecond processing latency.
● Supports event-time based windowing operations with late arrival of records.
● Supports exactly-once processing: Each record will be processed once and only once even when there
is a failure on either Streams clients or Kafka brokers in the middle of processing. We have to set
processing.guarantee=exactly_once

● Stream: An ordered, replayable, and fault-tolerant sequence of immutable
data records (key-value pair).
● Stream Processor: Is a node in the processor topology; it represents a
processing step to transform data in streams by receiving one input record
at a time from its upstream processors in the topology, applying its
operation to it, and may subsequently produce one or more output records
to its downstream processors.
● Source Processor: Is a special type of stream processor that does not have
any upstream processors. It produces an input stream to its topology from
one or multiple Kafka topics by consuming records from these topics and
forwarding them to its down-stream processors.
● Sink Processor: Is a special type of stream processor that does not have
down-stream processors. It sends any received records from its up-stream
processors to a specified Kafka topic.

Kafka Streams Example (2)
In previous code examples we are trying to find character counts:
● Via createTopology() method, we create stream’s topology, which accepts strings and finds the
counter of its char.
● Via getStreamConfig() method, we add all the required KafkaStream configurations.
● Via startStreaming() method, we initialize KStream and start streaming.
● As we can see on the tests, our topology pipeline works as we get the expected output.

Kafka Reliable Data Delivery
● Reliability is referring to a system in which behavior is guaranteed.
○ Kafka provides order guarantee in each partition
○ Produced messages are considered committed when they are written to all in-sync replicas.
○ Messages which are committed will not be lost if one at least replica is alive.
○ Consumers can only read committed messages
● Durability ensures that once the operation has been carried out, it will persist and cannot be undone
even if the system fails.
○ Kafka with multiple replicas provides durability.

Kafka configurations for Reliable Data Delivery (1)
On broker level we can set some configuration to provide reliable message storage and delivery:
1. Replication Factor
default.replication.factor specifies each partition’s replicas. A replication factor of N allows loss N-1 brokers, but the
system will remain reliable. Higher replication factor means higher availability and reliability, but needs more disc
capacity. (tradeoff among availability / capacity)
2. Unclean Leader Election
unclean.leader.election.enable specifies if a replica which is out-of-sync can be selected as leader. Setting it to false
we ensure that the electing replica is always in-sync with the leader, preventing losing messages but we wait until
the original leader is online again, resulting in lower availability. (tradeoff among consistency / availability)
3. Minimum In-Sync Replicas
Using min.insync.replicas we specify the minimum number of replicas that must acknowledge a write to be
considered successful. Preferred value: replication.factor - 1.

On producer level we can set some configuration to provide reliable message storage and delivery:
1. Send Acknowledgments
Best approach is setting acks=all to wait until leader and all in-sync replicas get the message,
before is considered as “committed”. It reduces messages lost risk but is the slowest option, which
can be improved by combining it with async send method and batches.
2. Configuring Producer Retries
It is essential, the producer handles retryable errors. If we want to avoid losing messages, then our
producer should always retry to send them.

On consumer level we can set some configuration to provide reliable message storage and delivery:
● Manually commit offsets, instead of auto-commit them, to have more control on offsets and we can
reduce duplications.
● Reducing missing messages on consumer side via committing offset after events were processed and not
after polling.
● Design rebalances properly (e.g. commit offset before a partition is revoked).
● Enforce EoS (exactly-once delivery). We can achieve it via database table to store already processed
messages with unique id and storing offset in database in a transaction. This way manipulate records and
commit offset becomes a atomic operation. This way we avoid re-process same messages.
● Retry on errors. We can write retryable errors in a separate topic and continue. These messages can be
processed later by another service.

Kafka Use Cases
Kafka allows us to have a huge amount of messages. As a result, it's ideal to be used as the heart of our
system’s architecture, which connects different applications. Below we can see some use cases:
● Messaging
● Website Activity Tracking
● Metrics
● Real-time data processing
● Event Sourcing
● Log Aggregation

Kafka Summary
● Apache Kafka is a distributed streaming platform.
● Kafka stores messages in topics.
● Topics are partitioned and replicated across multiple brokers in a cluster.
● Producers send messages to topics from which consumers read.
● Kafka is able to process streams of events real-time.
● Kafka is capable of handling enormous amount of events.
● Kafka provides low-latency, high-throughput, fault-tolerance and great scalability.

Let’s see some code snippets!
(https://github.com/arconsis/Eshop-EDA)

References
● Kafka The Definitive Guide by Neha Narkhede, Gwen Shapira, Todd Palino:
https://www.amazon.com/Kafka-Definitive-Real-Time-Stream-Processing/dp/1491936169
● Designing event driven systems by Ben Stopford (https://www.confluent.io/designing-event-driven-
systems/)
● https://docs.confluent.io/platform/current/installation/configuration
● https://docs.confluent.io/platform/current/clients/
● https://kafka.apache.org/
● https://kafka.apache.org/26/documentation/
● https://betterprogramming.pub/thorough-introduction-to-apache-kafka-6fbf2989bbc1
● https://jaceklaskowski.gitbooks.io/apache-kafka/content/
● https://preparingforcodinginterview.wordpress.com/2020/07/18/kafka-consumer-group-
coordinator-consumer-group-leader/

Contact Info
● Arconsis:
○ Website: https://www.arconsis.com
○ Github: https://github.com/arconsis
● Dimos Botsaris:
○ Website: https://www.eldimious.com/
○ Github: https://github.com/eldimious
○ Email: botsaris.d@gmail.com
● Alexandros Koufatzis:
○ Github: https://github.com/akoufa
○ Email: akoufa@gmail.com

Introduction to Kafka and Event-Driven

Introduction to Kafka and Event-Driven

More Related Content

What's hot

Similar to Introduction to Kafka and Event-Driven

Recently uploaded

Introduction to Kafka and Event-Driven