Kafka Tutorial - Introduction to Apache Kafka (Part 1)

™
Cassandra / Kafka Support in EC2/AWS. Kafka Training, Kafka
Consulting
Cloudurable provides Cassandra and Kafka Support on
AWS/EC2
Kafka Tutorial
Kafka Tutorial
What is Kafka?
Why is Kafka important?
Kafka architecture and design
Kafka Universe
Kafka Schemas
Java Producer and Consumer
examples
Tutorial on Kafka
Streaming Platform

™
Consulting

™
Consulting
Kafka growing
Why Kafka?
Kafka adoption is on the rise
but why
What is Kafka?

Consulting
™
Kafka growth exploding
❖ Kafka growth exploding
❖ 1/3 of all Fortune 500 companies
❖ Top ten travel companies, 7 of ten top banks, 8 of ten top
insurance companies, 9 of ten top telecom companies
❖ LinkedIn, Microsoft and Netflix process 4 comma
message a day with Kafka (1,000,000,000,000)
❖ Real-time streams of data, used to collect big data or to
do real time analysis (or both)

Consulting
™
Why Kafka is Needed?
❖ Real time streaming data processed for real time
analytics
❖ Service calls, track every call, IOT sensors
❖ Apache Kafka is a fast, scalable, durable, and fault-
tolerant publish-subscribe messaging system
❖ Kafka is often used instead of JMS, RabbitMQ and
AMQP
❖ higher throughput, reliability and replication

Consulting
™
Why is Kafka needed? 2
❖ Kafka can works in combination with
❖ Flume/Flafka, Spark Streaming, Storm, HBase and Spark
for real-time analysis and processing of streaming data
❖ Feed your data lakes with data streams
❖ Kafka brokers support massive message streams for follow-
up analysis in Hadoop or Spark
❖ Kafka Streaming (subproject) can be used for real-time
analytics

Consulting
™
Kafka Use Cases
❖ Stream Processing
❖ Website Activity Tracking
❖ Metrics Collection and Monitoring
❖ Log Aggregation
❖ Real time analytics
❖ Capture and ingest data into Spark / Hadoop
❖ CRQS, replay, error recovery
❖ Guaranteed distributed commit log for in-memory computing

Consulting
™
Who uses Kafka?
❖ LinkedIn: Activity data and operational metrics
❖ Twitter: Uses it as part of Storm – stream processing
infrastructure
❖ Square: Kafka as bus to move all system events to various
Square data centers (logs, custom events, metrics, an so
on). Outputs to Splunk, Graphite, Esper-like alerting
systems
❖ Spotify, Uber, Tumbler, Goldman Sachs, PayPal, Box,
Cisco, CloudFlare, DataDog, LucidWorks, MailChimp,
NetFlix, etc.

Consulting
™
Why is Kafka Popular?
❖ Great performance
❖ Operational Simplicity, easy to setup and use, easy to reason
❖ Stable, Reliable Durability,
❖ Flexible Publish-subscribe/queue (scales with N-number of consumer groups),
❖ Robust Replication,
❖ Producer Tunable Consistency Guarantees,
❖ Ordering Preserved at shard level (Topic Partition)
❖ Works well with systems that have data streams to process, aggregate,
transform & load into other stores
Most important reason: Kafka’s great performance: throughput, latency, obtained through great engineering

Consulting
™
Why is Kafka so fast?
❖ Zero Copy - calls the OS kernel direct rather to move data fast
❖ Batch Data in Chunks - Batches data into chunks
❖ end to end from Producer to file system to Consumer
❖ Provides More efficient data compression. Reduces I/O latency
❖ Sequential Disk Writes - Avoids Random Disk Access
❖ writes to immutable commit log. No slow disk seeking. No random I/O
operations. Disk accessed in sequential manner
❖ Horizontal Scale - uses 100s to thousands of partitions for a single topic
❖ spread out to thousands of servers
❖ handle massive load

Consulting
™
Kafka Streaming Architecture

Consulting
™
Why Kafka Review
❖ Why is Kafka so fast?
❖ How fast is Kafka usage growing?
❖ How is Kafka getting used?
❖ Where does Kafka fit in the Big Data Architecture?
❖ How does Kafka relate to real-time analytics?
❖ Who uses Kafka?

™
Consulting
Cassandra / Kafka Support in EC2/AWS
What is Kafka? Kafka messaging
Kafka Overview

Consulting
™
What is Kafka?
❖ Distributed Streaming Platform
❖ Publish and Subscribe to streams of records
❖ Fault tolerant storage
❖ Replicates Topic Log Partitions to multiple servers
❖ Process records as they occur
❖ Fast, efficient IO, batching, compression, and more
❖ Used to decouple data streams

Consulting
™
Kafka helps decouple data
streams
❖ Kafka decouple data streams
❖ producers don’t know about consumers
❖ Flexible message consumption
❖ Kafka broker delegates log partition offset (location) to
Consumers (clients)

Consulting
™
Kafka messaging allows
❖ Feeding of high-latency daily or hourly data analysis into
Spark, Hadoop, etc.
❖ Feeding microservices real-time messages
❖ Sending events to CEP system
❖ Feeding data to do real-time analytic systems
❖ Up to date dashboards and summaries
❖ At same time

Consulting
™
Kafka Decoupling Data
Streams

Consulting
™
Kafka Polyglot clients / Wire
protocol
❖ Kafka communication from clients and servers wire
protocol over TCP protocol
❖ Protocol versioned
❖ Maintains backwards compatibility
❖ Many languages supported
❖ Kafka REST proxy allows easy integration (not part of core)
❖ Also provides Avro/Schema registry support via Kafka
ecosystem (not part of core)

Consulting
™
Kafka Usage
❖ Build real-time streaming applications that react to streams
❖ Real-time data analytics
❖ Transform, react, aggregate, join real-time data flows
❖ Feed events to CEP for complex event processing
❖ Feed data lakes
❖ Build real-time streaming data pipe-lines
❖ Enable in-memory microservices (actors, Akka, Vert.x, Qbit,
RxJava)

Consulting
™
Kafka Use Cases
❖ Metrics / KPIs gathering
❖ Aggregate statistics from many sources
❖ Event Sourcing
❖ Used with microservices (in-memory) and actor systems
❖ Commit Log
❖ External commit log for distributed systems. Replicated
data between nodes, re-sync for nodes to restore state
❖ Real-time data analytics, Stream Processing, Log
Aggregation, Messaging, Click-stream tracking, Audit trail,
etc.

Consulting
™
Kafka Record Retention
❖ Kafka cluster retains all published records
❖ Time based – configurable retention period
❖ Size based - configurable based on size
❖ Compaction - keeps latest record
❖ Retention policy of three days or two weeks or a month
❖ It is available for consumption until discarded by time, size or
compaction
❖ Consumption speed not impacted by size

Consulting
™
Kafka scalable message
storage
❖ Kafka acts as a good storage system for records/messages
❖ Records written to Kafka topics are persisted to disk and replicated to
other servers for fault-tolerance
❖ Kafka Producers can wait on acknowledgment
❖ Write not complete until fully replicated
❖ Kafka disk structures scales well
❖ Writing in large streaming batches is fast
❖ Clients/Consumers can control read position (offset)
❖ Kafka acts like high-speed file system for commit log storage,
replication

Consulting
™
Kafka Review
❖ How does Kafka decouple streams of data?
❖ What are some use cases for Kafka where you work?
❖ What are some common use cases for Kafka?
❖ How is Kafka like a distributed message storage
system?
❖ How does Kafka know when to delete old messages?
❖ Which programming languages does Kafka support?

™
Consulting
Kafka Architecture

Consulting
™
Kafka Fundamentals
❖ Records have a key (optional), value and timestamp; Immutable
❖ Topic a stream of records (“/orders”, “/user-signups”), feed name
❖ Log topic storage on disk
❖ Partition / Segments (parts of Topic Log)
❖ Producer API to produce a streams or records
❖ Consumer API to consume a stream of records
❖ Broker: Kafka server that runs in a Kafka Cluster. Brokers form a cluster.
Cluster consists on many Kafka Brokers on many servers.
❖ ZooKeeper: Does coordination of brokers/cluster topology. Consistent file
system for configuration information and leadership election for Broker Topic
Partition Leaders

Consulting
™
Kafka: Topics, Producers, and
Consumers
Kafka
Cluster
Topic
Producer
Producer
Producer
Consumer
Consumer
Consumer
record
record

Consulting
™
Apache Kafka - Core Kafka
❖ Kafka gets conflated with Kafka ecosystem
❖ Apache Core Kafka consists of Kafka Broker, startup scripts for
ZooKeeper, and client APIs for Kafka
❖ Apache Core Kafka does not include
❖ Confluent Schema Registry (not an Apache project)
❖ Kafka REST Proxy (not an Apache project)
❖ Kafka Connect (not an Apache project)
❖ Kafka Streams (not an Apache project)

Consulting
™
Apache Kafka

Consulting
™
Kafka needs Zookeeper
❖ Zookeeper helps with leadership election of Kafka Broker and
Topic Partition pairs
❖ Zookeeper manages service discovery for Kafka Brokers that
form the cluster
❖ Zookeeper sends changes to Kafka
❖ New Broker join, Broker died, etc.
❖ Topic removed, Topic added, etc.
❖ Zookeeper provides in-sync view of Kafka Cluster configuration

Consulting
™
Kafka Producer/Consumer
Details
❖ Producers write to and Consumers read from Topic(s)
❖ Topic associated with a log which is data structure on disk
❖ Producer(s) append Records at end of Topic log
❖ Topic Log consist of Partitions -
❖ Spread to multiple files on multiple nodes
❖ Consumers read from Kafka at their own cadence
❖ Each Consumer (Consumer Group) tracks offset from where they left off reading
❖ Partitions can be distributed on different machines in a cluster
❖ high performance with horizontal scalability and failover with replication

Consulting
™
Kafka Topic Partition, Consumers,
Producers
0 1 42 3 5 6 7 8 9 10 11
Partition
0
Consumer Group A
Producer
Consumer Group B
Consumer groups remember offset where they left off.
Consumers groups each have their own offset.
Producer writing to offset 12 of Partition 0 while…
Consumer Group A is reading from offset 6.
Consumer Group B is reading from offset 9.

Consulting
™
Kafka Scale and Speed
❖ How can Kafka scale if multiple producers and consumers read/write to
same Kafka Topic log?
❖ Writes fast: Sequential writes to filesystem are fast (700 MB or more a
second)
❖ Scales writes and reads by sharding:
❖ Topic logs into Partitions (parts of a Topic log)
❖ Topics logs can be split into multiple Partitions different
machines/different disks
❖ Multiple Producers can write to different Partitions of the same Topic
❖ Multiple Consumers Groups can read from different partitions efficiently

Consulting
™
Kafka Brokers
❖ Kafka Cluster is made up of multiple Kafka Brokers
❖ Each Broker has an ID (number)
❖ Brokers contain topic log partitions
❖ Connecting to one broker bootstraps client to entire
cluster
❖ Start with at least three brokers, cluster can have, 10,
100, 1000 brokers if needed

Consulting
™
Kafka Cluster, Failover, ISRs
❖ Topic Partitions can be replicated
❖ across multiple nodes for failover
❖ Topic should have a replication factor greater than 1
❖ (2, or 3)
❖ Failover
❖ if one Kafka Broker goes down then Kafka Broker with
ISR (in-sync replica) can serve data

Consulting
™
ZooKeeper does coordination for Kafka
Cluster

Consulting
™
Failover vs. Disaster Recovery
❖ Replication of Kafka Topic Log partitions allows for failure of
a rack or AWS availability zone
❖ You need a replication factor of at least 3
❖ Kafka Replication is for Failover
❖ Mirror Maker is used for Disaster Recovery
❖ Mirror Maker replicates a Kafka cluster to another data-center
or AWS region
❖ Called mirroring since replication happens within a cluster

Consulting
™
Kafka Review
❖ How does Kafka decouple streams of data?
❖ What are some use cases for Kafka where you work?
❖ What are some common use cases for Kafka?
❖ What is a Topic?
❖ What is a Broker?
❖ What is a Partition? Offset?
❖ Can Kafka run without Zookeeper?
❖ How do implement failover in Kafka?
❖ How do you implement disaster recovery in Kafka?

™
Consulting
Kafka versus

Consulting
™
Kafka vs JMS, SQS, RabbitMQ
Messaging
❖ Is Kafka a Queue or a Pub/Sub/Topic?
❖ Yes
❖ Kafka is like a Queue per consumer group
❖ Kafka is a queue system per consumer in consumer group so load
balancing like JMS, RabbitMQ queue
❖ Kafka is like Topics in JMS, RabbitMQ, MOM
❖ Topic/pub/sub by offering Consumer Groups which act like
subscriptions
❖ Broadcast to multiple consumer groups
❖ MOM = JMS, ActiveMQ, RabbitMQ, IBM MQ Series, Tibco, etc.

Consulting
™
Kafka vs MOM
❖ By design, Kafka is better suited for scale than traditional MOM systems due
to partition topic log
❖ Load divided among Consumers for read by partition
❖ Handles parallel consumers better than traditional MOM
❖ Also by moving location (partition offset) in log to client/consumer side of
equation instead of the broker, less tracking required by Broker and more
flexible consumers
❖ Kafka written with mechanical sympathy, modern hardware, cloud in mind
❖ Disks are faster
❖ Servers have tons of system memory
❖ Easier to spin up servers for scale out

Consulting
™
Kinesis and Kafka are similar
❖ Kinesis Streams is like Kafka Core
❖ Kinesis Analytics is like Kafka Streams
❖ Kinesis Shard is like Kafka Partition
❖ Similar and get used in similar use cases
❖ In Kinesis, data is stored in shards. In Kafka, data is stored in partitions
❖ Kinesis Analytics allows you to perform SQL like queries on data streams
❖ Kafka Streaming allows you to perform functional aggregations and mutations
❖ Kafka integrates well with Spark and Flink which allows SQL like queries on
streams

Consulting
™
Kafka vs. Kinesis
❖ Data is stored in Kinesis for default 24 hours, and you can increase that up to 7 days.
❖ Kafka records default stored for 7 days
❖ can increase until you run out of disk space.
❖ Decide by the size of data or by date.
❖ Can use compaction with Kafka so it only stores the latest timestamp per key per record
in the log
❖ With Kinesis data can be analyzed by lambda before it gets sent to S3 or RedShift
❖ With Kinesis you pay for use, by buying read and write units.
❖ Kafka is more flexible than Kinesis but you have to manage your own clusters, and requires
some dedicated DevOps resources to keep it going
❖ Kinesis is sold as a service and does not require a DevOps team to keep it going (can be
more expensive and less flexible, but much easier to setup and run)

™
Consulting
Kafka Topics
Kafka Topics
Architecture
Replication
Failover
Parallel processing
Kafka Topic Architecture

Consulting
™
Topics, Logs, Partitions
❖ Kafka Topic is a stream of records
❖ Topics stored in log
❖ Log broken up into partitions and segments
❖ Topic is a category or stream name or feed
❖ Topics are pub/sub
❖ Can have zero or many subscribers - consumer groups
❖ Topics are broken up and spread by partitions for speed and
size

Consulting
™
Topic Partitions
❖ Topics are broken up into partitions
❖ Partitions decided usually by key of record
❖ Key of record determines which partition
❖ Partitions are used to scale Kafka across many servers
❖ Record sent to correct partition by key
❖ Partitions are used to facilitate parallel consumers
❖ Records are consumed in parallel up to the number of partitions
❖ Order guaranteed per partition
❖ Partitions can be replicated to multiple brokers

Consulting
™
Topic Partition Log
❖ Order is maintained only in a single partition
❖ Partition is ordered, immutable sequence of records that is
continually appended to—a structured commit log
❖ Records in partitions are assigned sequential id number called
the offset
❖ Offset identifies each record within the partition
❖ Topic Partitions allow Kafka log to scale beyond a size that will
fit on a single server
❖ Topic partition must fit on servers that host it
❖ topic can span many partitions hosted on many servers

Consulting
™
Topic Parallelism and
Consumers
❖ Topic Partitions are unit of parallelism - a partition can only
be used by one consumer in group at a time
❖ Consumers can run in their own process or their own thread
❖ If a consumer stops, Kafka spreads partitions across
remaining consumer in group
❖ #of Consumers you can run per Consumer Group limited by
#of Partitions
❖ Consumers getting assigned partition aids in efficient
message consumption tracking

Consulting
™
Kafka Topic Partitions Layout
0 1 42 3 5 6 7 8 9 10 11
0 1 42 3 5 6 7 8
0 1 42 3 5 6 7 8 9 10
Older Newer
0 1 42 3 5 6 7
Partition
0
Partition
1
Partition
2
Partition
3
Writes

Consulting
™
Replication: Kafka Partition
Distribution
❖ Each partition has leader server and zero or more follower
servers
❖ Leader handles all read and write requests for partition
❖ Followers replicate leader, and take over if leader dies
❖ Used for parallel consumer handling within a group
❖ Partitions of log are distributed over the servers in the Kafka cluster
with each server handling data and requests for a share of
partitions
❖ Each partition can be replicated across a configurable number of
Kafka servers - Used for fault tolerance

Consulting
™
Replication: Kafka Partition
Leader
❖ One node/partition’s replicas is chosen as leader
❖ Leader handles all reads and writes of Records for
partition
❖ Writes to partition are replicated to followers
(node/partition pair)
❖ An follower that is in-sync is called an ISR (in-sync
replica)
❖ If a partition leader fails, one ISR is chosen as new
leader

Consulting
™
Kafka Replication to Partition 0
Kafka Broker 0
Partition 0
Partition 1
Partition 2
Partition 3
Partition 4
Kafka Broker 1
Partition 0
Partition 1
Partition 2
Partition 3
Partition 4
Kafka Broker 2
Partition 1
Partition 2
Partition 3
Partition 4
Client Producer
1) Write record
Partition 0
2) Replicate
record
2) Replicate
record
Leader Red
Follower Blue
Record is considered "committed"
when all ISRs for partition
wrote to their log.
Only committed records are
readable from consumer

Consulting
™
Kafka Replication to Partitions
1
Kafka Broker 0
Partition 0
Partition 1
Partition 2
Partition 3
Partition 4
Kafka Broker 1
Partition 0
Partition 1
Partition 2
Partition 3
Partition 4
Kafka Broker 2
Partition 1
Partition 2
Partition 3
Partition 4
Client Producer
1) Write record
Partition 0
2) Replicate
record
2) Replicate
record
Another partition can
be owned
by another leader
on another Kafka broker
Leader Red
Follower Blue

Consulting
™
Topic Review
❖ What is an ISR?
❖ How does Kafka scale Consumers?
❖ What are leaders? followers?
❖ How does Kafka perform failover for Consumers?
❖ How does Kafka perform failover for Brokers?

™
Consulting
Kafka Producers
Kafka Producers
Partition selection
Durability
Kafka Producer Architecture

Consulting
™
Kafka Producers
❖ Producers send records to topics
❖ Producer picks which partition to send record to per topic
❖ Can be done in a round-robin
❖ Can be based on priority
❖ Typically based on key of record
❖ Kafka default partitioner for Java uses hash of keys to
choose partitions, or a round-robin strategy if no key
❖ Important: Producer picks partition

Consulting
™
Kafka Producers and
Consumers
0 1 42 3 5 6 7 8 9 10 11
Partition
0
Producers
Consumer Group A
Producers are writing at Offset 12
Consumer Group A is Reading from Offset 9.

Consulting
™
Kafka Producers
❖ Producers write at their own cadence so order of Records
cannot be guaranteed across partitions
❖ Producer configures consistency level (ack=0, ack=all,
ack=1)
❖ Producers pick the partition such that Record/messages
goes to a given same partition based on the data
❖ Example have all the events of a certain 'employeeId' go to
same partition
❖ If order within a partition is not needed, a 'Round Robin'
partition strategy can be used so Records are evenly
distributed across partitions.

Consulting
™
Producer Review
❖ Can Producers occasionally write faster than
consumers?
❖ What is the default partition strategy for Producers
without using a key?
❖ What is the default partition strategy for Producers using
a key?
❖ What picks which partition a record is sent to?

™
Consulting
Kafka Consumers
Load balancing consumers
Failover for consumers
Offset management per consumer
group
Kafka Consumer Architecture

Consulting
™
Kafka Consumer Groups
❖ Consumers are grouped into a Consumer Group
❖ Consumer group has a unique id
❖ Each consumer group is a subscriber
❖ Each consumer group maintains its own offset
❖ Multiple subscribers = multiple consumer groups
❖ Each has different function: one might delivering records to
microservices while another is streaming records to Hadoop
❖ A Record is delivered to one Consumer in a Consumer Group
❖ Each consumer in consumer groups takes records and only one consumer
in group gets same record
❖ Consumers in Consumer Group load balance record consumption

Consulting
™
Kafka Consumer Load Share
❖ Kafka Consumer consumption divides partitions over consumers in a
Consumer Group
❖ Each Consumer is exclusive consumer of a "fair share" of partitions
❖ This is Load Balancing
❖ Consumer membership in Consumer Group is handled by the Kafka
protocol dynamically
❖ If new Consumers join Consumer group, it gets a share of partitions
❖ If Consumer dies, its partitions are split among remaining live Consumers
in Consumer Group

Consulting
™
0 1 42 3 5 6 7 8 9 10 11
Partition
0
Consumer Group A
Producers
Consumer Group B
Consumers remember offset where they left off.
Consumers groups each have their own offset per partition.

Consulting
™
Processing
❖ How does Kafka divide up topic so multiple Consumers in a Consumer
Group can process a topic?
❖ You group consumers into consumers group with a group id
❖ Consumers with same id belong in same Consumer Group
❖ One Kafka broker becomes group coordinator for Consumer Group
❖ assigns partitions when new members arrive (older clients would talk
direct to ZooKeeper now broker does coordination)
❖ or reassign partitions when group members leave or topic changes
(config / meta-data change
❖ When Consumer group is created, offset set according to reset policy of
topic

Consulting
™
Kafka Consumer Failover
❖ Consumers notify broker when it successfully processed a
record
❖ advances offset
❖ If Consumer fails before sending commit offset to Kafka
broker,
❖ different Consumer can continue from the last committed
offset
❖ some Kafka records could be reprocessed
❖ at least once behavior
❖ messages should be idempotent

Consulting
™
Kafka Consumer Offsets and
Recovery
❖ Kafka stores offsets in topic called “__consumer_offset”
❖ Uses Topic Log Compaction
❖ When a consumer has processed data, it should
commit offsets
❖ If consumer process dies, it will be able to start up and
start reading where it left off based on offset stored in
“__consumer_offset”

Consulting
™
Kafka Consumer: What can be
consumed?
❖ "Log end offset" is offset of last record written
to log partition and where Producers write to
next
❖ "High watermark" is offset of last record
successfully replicated to all partitions followers
❖ Consumer only reads up to “high watermark”.
Consumer can’t read un-replicated data

Consulting
™
Consumer to Partition
Cardinality
❖ Only a single Consumer from the same Consumer
Group can access a single Partition
❖ If Consumer Group count exceeds Partition count:
❖ Extra Consumers remain idle; can be used for failover
❖ If more Partitions than Consumer Group instances,
❖ Some Consumers will read from more than one
partition

Consulting
™
2 server Kafka cluster hosting 4 partitions (P0-P5)
Kafka Cluster
Server 2
P0 P1 P5
Server 1
P2 P3 P4
Consumer Group A
C0 C1 C3
Consumer Group B
C0 C1 C3

Consulting
™
Multi-threaded Consumers
❖ You can run more than one Consumer in a JVM process
❖ If processing records takes a while, a single Consumer can run multiple threads to process
records
❖ Harder to manage offset for each Thread/Task
❖ One Consumer runs multiple threads
❖ 2 messages on same partitions being processed by two different threads
❖ Hard to guarantee order without threads coordination
❖ PREFER: Multiple Consumers can run each processing record batches in their own thread
❖ Easier to manage offset
❖ Each Consumer runs in its thread
❖ Easier to mange failover (each process runs X num of Consumer threads)

Consulting
™
Consumer Review
❖ What is a consumer group?
❖ Does each consumer have its own offset?
❖ When can a consumer see a record?
❖ What happens if there are more consumers than
partitions?
❖ What happens if you run multiple consumers in many
thread in the same JVM?

™
Consulting
Using Kafka Single Node
Using Kafka Single
Node
Run ZooKeeper, Kafka
Create a topic
Send messages from command
line
Read messages from command
line
Tutorial Using Kafka Single Node

Consulting
™
Run Kafka
❖ Run ZooKeeper start up script
❖ Run Kafka Server/Broker start up script
❖ Create Kafka Topic from command line
❖ Run producer from command line
❖ Run consumer from command line

Consulting
™
Run ZooKeeper

Consulting
™
Run Kafka Server

Consulting
™
Create Kafka Topic

Consulting
™
List Topics

Consulting
™
Run Kafka Producer

Consulting
™
Run Kafka Consumer

Consulting
™
Running Kafka Producer and
Consumer

Consulting
™
Kafka Single Node Review
❖ What server do you run first?
❖ What tool do you use to create a topic?
❖ What tool do you use to see topics?
❖ What tool did we use to send messages on the command line?
❖ What tool did we use to view messages in a topic?
❖ Why were the messages coming out of order?
❖ How could we get the messages to come in order from the
consumer?

™
Consulting
Use Kafka to send and receive messages
Lab Use Kafka
Use single server version of
Kafka.
Setup single node.
Single ZooKeeper.
Create a topic.
Produce and consume messages
from the command line.

™
Consulting
Using Kafka Cluster
and Failover
Demonstrate Kafka Cluster
Create topic with replication
Show consumer failover
Show broker failover
Kafka Tutorial Cluster and
Failover

Consulting
™
Objectives
❖ Run many Kafka Brokers
❖ Create a replicated topic
❖ Demonstrate Pub / Sub
❖ Demonstrate load balancing
consumers
❖ Demonstrate consumer failover
❖ Demonstrate broker failover

Consulting
™
Running many nodes
❖ If not already running, start up ZooKeeper
❖ Shutdown Kafka from first lab
❖ Copy server properties for three brokers
❖ Modify properties files, Change port, Change Kafka log
location
❖ Start up many Kafka server instances
❖ Create Replicated Topic
❖ Use the replicated topic

Consulting
™
Create three new server-
n.properties files
❖ Copy existing server.properties to server-
0.properties, server-1.properties, server-2.properties
❖ Change server-1.properties to use log.dirs
“./logs/kafka-logs-0”
❖ Change server-1.properties to use port 9093, broker
id 1, and log.dirs “./logs/kafka-logs-1”
❖ Change server-2.properties to use port 9094, broker
id 2, and log.dirs “./logs/kafka-logs-2”

Consulting
™
Modify server-x.properties
❖ Each have different
broker.id
❖ Each have different
log.dirs
❖ Each had different
port

Consulting
™
Create Startup scripts for three Kafka
servers
❖ Passing properties files
from last step

Consulting
™
Run Servers

Consulting
™
Create Kafka replicated topic my-
failsafe-topic
❖ Replication Factor is set to 3
❖ Topic name is my-failsafe-topic
❖ Partitions is 13

Consulting
™
Start Kafka Consumer
❖ Pass list of Kafka servers to bootstrap-
server
❖ We pass two of the three
❖ Only one needed, it learns about the rest

Consulting
™
Start Kafka Producer
❖ Start producer
❖ Pass list of Kafka Brokers

Consulting
™
Kafka 1 consumer and 1 producer
running

Consulting
™
Start a second and third
consumer
❖ Acts like pub/sub
❖ Each consumer
in its own group
❖ Message goes to
each
❖ How do we load
share?

Consulting
™
Running consumers in same
group
❖ Modify start consumer script
❖ Add the consumers to a group called
mygroup
❖ Now they will share load

Consulting
™
Start up three consumers
again
❖ Start up producer and three consumers
❖ Send 7 messages
❖ Notice how messages are spread among 3 consumers

Consulting
™
Consumer Failover
❖ Kill one consumer
❖ Send seven more
messages
❖ Load is spread to
remaining consumers
❖ Failover WORK!

Consulting
™
Create Kafka Describe Topic
❖ —describe will show list partitions, ISRs, and
partition leadership

Consulting
™
Use Describe Topics
❖ Lists which broker owns (leader of) which partition
❖ Lists Replicas and ISR (replicas that are up to
date)
❖ Notice there are 13 topics

Consulting
™
Test Broker Failover: Kill 1st
server
se Kafka topic describe to see that a new leader was elected!
Kill the first server

Consulting
™
Show Broker Failover Worked
❖ Send two more
messages from the
producer
❖ Notice that the
consumer gets the
messages
❖ Broker Failover WORKS!

Consulting
™
Kafka Cluster Review
❖ Why did the three consumers not load share the
messages at first?
❖ How did we demonstrate failover for consumers?
❖ How did we demonstrate failover for producers?
❖ What tool and option did we use to show ownership of
partitions and the ISRs?

™
Consulting
Use Kafka to send and receive messages
Lab 2 Use Kafka
multiple nodes
Use a Kafka Cluster to
replicate a Kafka topic log

™
Consulting
Kafka Ecosystem
Kafka Connect
Kafka Streaming
Kafka Schema Registry
Kafka REST

Consulting
™
Kafka Universe
❖ Ecosystem is Apache Kafka Core plus these (and community Kafka Connectors)
❖ Kafka Streams
❖ Streams API to transform, aggregate, process records from a stream and produce
derivative streams
❖ Kafka Connect
❖ Connector API reusable producers and consumers
❖ (e.g., stream of changes from DynamoDB)
❖ Kafka REST Proxy
❖ Producers and Consumers over REST (HTTP)
❖ Schema Registry - Manages schemas using Avro for Kafka Records
❖ Kafka MirrorMaker - Replicate cluster data to another cluster

Consulting
™
What comes in Apache Kafka
Core?
Apache Kafka Core Includes:
❖ ZooKeeper and startup scripts
❖ Kafka Server (Kafka Broker), Kafka Clustering
❖ Utilities to monitor, create topics, inspect topics, replicated
(mirror) data to another datacenter
❖ Producer APIs, Consumer APIs
❖ Part of Apache Foundation
❖ Packages / Distributions are free do download with no registry

Consulting
™
What comes in Kafka Extensions?
Confluent.io:
❖ All of Kafka Core
❖ Schema Registry (schema versioning and compatibility checks) (Confluent project)
❖ Kafka REST Proxy (Confluent project)
❖ Kafka Streams (aggregation, joining streams, mutating streams, creating new
streams by combining other streams) (Confluent project)
❖ Not Part of Apache Foundation controlled by Confluent.io
❖ Code hosted on GitHub
❖ Packages / Distributions are free do download but you must register with Confluent
Community of Kafka Connectors from 3rd parties and Confluent

Consulting
™
Kafka Universe

Consulting
™
Kafka REST Proxy and Schema
Registry

Consulting
™
Kafka Stream : Stream
Processing
❖ Kafka Streams for Stream Processing
❖ Kafka enable real-time processing of streams.
❖ Kafka Streams supports Stream Processor
❖ processing, transformation, aggregation, and produces 1 to * output streams
❖ Example: video player app sends events videos watched, videos paused
❖ output a new stream of user preferences
❖ can gear new video recommendations based on recent user activity
❖ can aggregate activity of many users to see what new videos are hot
❖ Solves hard problems: out of order records, aggregating/joining across streams, stateful
computations, and more

Consulting
™
Kafka Connectors and
Streams
DB DB
App App
Connectors
Producers
Consumers
Streams
Kafka
Cluster
App
App
App
App
App
App

Consulting
™
Kafka Ecosystem review
❖ What is Kafka Streams?
❖ What is Kafka Connect?
❖ What is the Schema Registry?
❖ What is Kafka Mirror Maker?
❖ When might you use Kafka REST Proxy?

Consulting
™
References
❖ Learning Apache Kafka, Second Edition 2nd Edition by Nishant Garg (Author), 2015, ISBN 978-
1784393090, Packet Press
❖ Apache Kafka Cookbook, 1st Edition, Kindle Edition by Saurabh Minni (Author), 2015, ISBN 978-
1785882449, Packet Press
❖ Kafka Streams for Stream processing: A few words about how Kafka works, Serban Balamaci, 2017,
Blog: Plain Ol' Java
❖ Kafka official documentation, 2017
❖ Why we need Kafka? Quora
❖ Why is Kafka Popular? Quora
❖ Why is Kafka so Fast? Stackoverflow
❖ Kafka growth exploding (Tech Republic)
❖ Apache Kafka Series - Learning Apache Kafka for Beginners - great introduction to using Kafka on
Udemy by Stephane Maarek

™
Consulting
Working with Kafka Consumers – Java Example
Producer
Kafka Producer
Introduction
Java Examples
Working with producers in Java
Step by step first example
Creating a Kafka Producer in
Java

Consulting
™
Objectives Create Producer
❖ Create simple example that creates a Kafka Producer
❖ Create a new replicated Kafka topic
❖ Create Producer that uses topic to send records
❖ Send records with Kafka Producer
❖ Send records asynchronously.
❖ Send records synchronously

Consulting
™
Create Replicated Kafka
Topic

Consulting
™
Gradle Build script

Consulting
™
Create Kafka Producer to send
records
❖ Specify bootstrap servers
❖ Specify client.id
❖ Specify Record Key serializer
❖ Specify Record Value serializer

Consulting
™
Common Kafka imports and
constants

Consulting
™
Create Kafka Producer to send
records

Consulting
™
Send sync records with Kafka
Producer

Consulting
™
Running the Producer

Consulting
™
Send async records with Kafka
Producer

Consulting
™
Async Interface Callback

Consulting
™
Async Send Method
❖ Used to send a record to a topic
❖ provided callback gets called when the send is acknowledged
❖ Send is asynchronous, and method will return immediately
❖ once the record gets stored in the buffer of records waiting to post to the
Kafka broker
❖ Allows sending many records in parallel without blocking

Consulting
™
Checking that replication is
working
❖ Verify that replication is working with
❖ kafka-replica-verification
❖ Utility that ships with Kafka
❖ If lag or outage you will see it as follows:
$ kafka/bin/kafka-replica-verification.sh --broker-list localhost:9092 --topic-white-list my-example-topic
2017-05-17 14:06:46,446: verification process is started.
2017-05-17 14:07:16,416: max lag is 0 for partition [my-example-topic,12] at offset 197 among 13 partitions
…

Consulting
™
Java Kafka Simple Producer
recap
❖ Created simple example that creates a Kafka Producer
❖ Created a new replicated Kafka topic
❖ Created Producer that uses topic to send records
❖ Sent records with Kafka Producer using async and
sync send

Consulting
™
Kafka Producer Review
❖ What does the Callback lambda do?
❖ What will happen if the first server is down in the
bootstrap list? Can the producer still connect to the
other Kafka brokers in the cluster?
❖ When would you use Kafka async send vs. sync send?
❖ Why do you need two serializers for a Kafka record?

™
Consulting
Working with Kafka Consumer – Java Example
Kafka Consumer
Introduction
Java Examples
Working with consumers
Step by step first example
Creating a Kafka Java Consumer

Consulting
™
Objectives Create a Consumer
❖ Create simple example that creates a Kafka Consumer
❖ that consumes messages form the Kafka Producer
we wrote
❖ Create Consumer that uses topic from first example to
receive messages
❖ Process messages from Kafka with Consumer
❖ Demonstrate how Consumer Groups work

Consulting
™
Create Consumer using Topic to Receive
Records
❖ Specify bootstrap servers
❖ Specify Consumer Group
❖ Specify Record Key deserializer
❖ Specify Record Value deserializer
❖ Subscribe to Topic from last session

Consulting
™
Create Consumer using Topic to Receive
Records

Consulting
™
Process messages from Kafka with Consumer

Consulting
™
Consumer poll
❖ poll() method returns fetched records based on current
partition offset
❖ Blocking method waiting for specified time if no records
available
❖ When/If records available, method returns straight away
❖ Control the maximum records returned by the poll() with
props.put(ConsumerConfig.MAX_POLL_RECORDS_CON
FIG, 100);
❖ poll() is not meant to be called from multiple threads

Consulting
™
Running both Consumer then
Producer

Consulting
™
Logging
❖ Kafka uses
sl4j
❖ Set level to
DEBUG to
see what is
going on

Consulting
™
Try this: Consumers in Same
Group
❖ Three consumers and one producer sending 25 records
❖ Run three consumers processes
❖ Change Producer to send 25 records instead of 5
❖ Run one producer
❖ What happens?

Consulting
™
Outcome 3 Consumers Load
ShareConsumer 0 (key, value, partition, offset)
Consumer 1 (key, value, partition, offset)
Producer
Which consumer owns partition 10?
How many ConsumerRecords objects
did Consumer 0 get?
What is the next offset from Partition 5
that Consumer 2 should get?
Why does each consumer get unique

Consulting
™
Try this: Consumers in Different
Groups
❖ Three consumers with unique group and one producer
sending 5 records
❖ Modify Consumer to have unique group id
❖ Run three consumers processes
❖ Run one producer
❖ What happens?

Consulting
™
Pass Unique Group Id

Consulting
™
Outcome 3 Subscribers
Producer
Which consumer(s) owns partition 10?
How many ConsumerRecords objects
did Consumer 0 get?
What is the next offset from Partition 2
that Consumer 2 should get?
Why does each consumer get the same
messages?

Consulting
™
Try this: Consumers in Different
Groups
❖ Modify consumer: change group id back to non-unique
value
❖ Make the batch size 5
❖ Add a 100 ms delay in the consumer after each
message poll and print out record count and partition
count
❖ Modify the Producer to run 10 times with a 30 second
delay after each run and to send 50 messages each run
❖ Run producer

Consulting
™
Modify Consumer
❖ Change
group
name to
common
name
❖ Change
batch
size to 5

Consulting
™
Add a 100 ms delay to Consumer
after poll

Consulting
™
Modify Producer: Run 10 times, add 30
second delay
❖ Run 10
times
❖ Add 30
second
delay
❖ Send 50
records

Consulting
™
Notice one or more partitions per
ConsumerRecords

Consulting
™
Now run it again but..
❖ Run the consumers and producer again
❖ Wait 30 seconds
❖ While the producer is running kill one of the consumers and see the
records go to the other consumers
❖ Now leave just one consumer running, all of the messages should go to
the remaining consumer
❖ Now change consumer batch size to 500
props.put(ConsumerConfig.MAX_POLL_RECORDS_CONFIG, 500)
❖ and run it again

Consulting
™
Output form batch size 500

Consulting
™
Java Kafka Simple Consumer
Example Recap
❖ Created simple example that creates a Kafka Consumer to
consume messages from our Kafka Producer
❖ Used the replicated Kafka topic from first example
❖ Created Consumer that uses topic to receive messages
❖ Processed records from Kafka with Consumer
❖ Consumers in same group divide up and share partitions
❖ Each Consumer groups gets a copy of the same data
(really has a unique set of offset partition pairs per
Consumer Group)

Consulting
™
Kafka Consumer Review
❖ How did we demonstrate Consumers in a Consumer
Group dividing up topic partitions and sharing them?
❖ How did we demonstrate Consumers in different
Consumer Groups each getting their own offsets?
❖ How many records does poll get?
❖ Does a call to poll ever get records from two different
partitions?

Consulting
™
Related Content
Creating a Kafka Consumer in Java
Creating a Kafka Producer in Java
Kafka from the command line
Kafka clustering and failover basics
Kafka Architecture
What is Kafka?
Kafka Topic Architecture
Kafka Consumer Architecture
Kafka Producer Architecture
Kafka and Schema Registry
Kafka and Avro

™
Consulting
Kafka Low Level Architecture
Kafka Low-Level
Design
Design discussion of Kafka
low level design
Kafka Architecture: Low-Level
Design

Consulting
™
Kafka Design Motivation Goals
❖ Kafka built to support real-time analytics
❖ Designed to feed analytics system that did real-time processing
of streams
❖ Unified platform for real-time handling of streaming data feeds
❖ Goals:
❖ high-throughput streaming data platform
❖ supports high-volume event streams like log aggregation, user
activity, etc.

Consulting
™
Kafka Design Motivation Scale
❖ To scale Kafka is
❖ distributed,
❖ supports sharding
❖ load balancing
❖ Scaling needs inspired Kafka partitioning and consumer
model
❖ Kafka scales writes and reads with partitioned, distributed,
commit logs

Consulting
™
Kafka Design Motivation Use
Cases
❖ Also designed to support these Use Cases
❖ Handle periodic large data loads from offline systems
❖ Handle traditional messaging use-cases, low-latency.
❖ Like MOMs, Kafka is fault-tolerance for node failures
through replication and leadership election
❖ Design more like a distributed database transaction log
❖ Unlike MOMs, replication, scale not afterthought

Consulting
™
Persistence: Embrace
filesystem
❖ Kafka relies heavily on filesystem for storing and caching messages/records
❖ Disk performance of hard drives performance of sequential writes is fast
❖ JBOD with six 7200rpm SATA RAID-5 array clocks at 600MB/sec
❖ Heavily optimized by operating systems
❖ Ton of cache: Operating systems use available of main memory for disk
caching
❖ JVM GC overhead is high for caching objects OS file caches are almost
free
❖ Kafka greatly simplifies code for cache coherence by using OS page cache
❖ Kafka disk does sequential reads easily optimized by OS page cache

Consulting
™
Big fast HDDs and long sequential
access
❖ Like Cassandra, LevelDB, RocksDB, and others, Kafka uses
long sequential disk access for read and writes
❖ Kafka uses tombstones instead of deleting records right
away
❖ Modern Disks have somewhat unlimited space and are fast
❖ Kafka can provide features not usually found in a messaging
system like holding on to old messages for a really long time
❖ This flexibility allows for interesting application of Kafka

Consulting
™
Kafka Record Retention
Redux
❖ Kafka cluster retains all published records
❖ Time based – configurable retention period
❖ Size based - configurable based on size
❖ Compaction - keeps latest record
❖ Kafka uses Topic Partitions
❖ Partitions are broken down into Segment files

Consulting
™
Broker Log Config
Kafka Broker Config for Logs
NAME DESCRIPTION DEFAULT
log.dir Log Directory will topic logs will be
stored use this or log.dirs.
/tmp/kafka-logs
log.dirs The directories where the Topics
logs are kept used for JBOD.
log.flush.interval.messages Accumulated messages count on a
log partition before messages are
flushed to disk.
9,223,372,036,854,780,
000
log.flush.interval.ms Maximum time that a topic message
is kept in memory before flushed to
disk. If not set, uses
log.flush.scheduler.interval.ms.
log.flush.offset.checkpoint.interval.m
s
Interval to flush log recovery point. 60,000
log.flush.scheduler.interval.ms Interval that topic messages are
periodically flushed from memory to
log.
9,223,372,036,854,780,
000

Consulting
™
Broker Log Retention Config
Kafka Broker Config for Logs
log.retention.bytes Delete log records by size. The
maximum size of the log before
deleting its older records.
long -1
log.retention.hours Delete log records by time hours.
Hours to keep a log file before deleting
older records (in hours), tertiary to
log.retention.ms property.
int 168
log.retention.minutes Delete log records by time minutes.
Minutes to keep a log file before
deleting it, secondary to
log.retention.ms property. If not set,
use log.retention.hours is used.
int null
log.retention.ms Delete log records by time
milliseconds. Milliseconds to keep a
log file before deleting it, If not set, use
log.retention.minutes.
long null

Consulting
™
Broker Log Segment File
Config
Kafka Broker Config - Log Segments
NAME DESCRIPTION TYPE DEFAULT
log.roll.hours Time period before rolling a new topic log segment.
(secondary to log.roll.ms property)
int 168
log.roll.ms Time period in milliseconds before rolling a new log
segment. If not set, uses log.roll.hours.
long
log.segment.bytes The maximum size of a single log segment file. int 1,073,741,82
4
log.segment.delete.delay.ms Time period to wait before deleting a segment file from
the filesystem.
long 60,000

Consulting
™
Kafka Producer Load
Balancing
❖ Producer sends records directly to Kafka broker partition
leader
❖ Producer asks Kafka broker for metadata about which
Kafka broker has which topic partitions leaders - thus no
routing layer needed
❖ Producer client controls which partition it publishes
messages to
❖ Partitioning can be done by key, round-robin or using a
custom semantic partitioner

Consulting
™
Kafka Producer Record
Batching
❖ Kafka producers support record batching. by the size of records
and auto-flushed based on time
❖ Batching is good for network IO throughput.
❖ Batching speeds up throughput drastically.
❖ Buffering is configurable
❖ lets you make a tradeoff between additional latency for better
throughput.
❖ Producer sends multiple records at a time which equates to fewer
IO requests instead of lots of one by one sends
QBit a microservice library uses message batching in an identical fashion as K
to send messages over WebSocket between nodes and from client to QBit ser

Consulting
™
More producer settings for
performance
For higher throughput, Kafka Producer allows buffering based on time and size.
Multiple records can be sent as a batches with fewer network requests.
Speeds up throughput drastically.

Consulting
™
Kafka Compression
❖ Kafka provides End-to-end Batch Compression
❖ Bottleneck is not always CPU or disk but often network bandwidth
❖ especially in cloud, containerized and virtualized environments
❖ especially when talking datacenter to datacenter or WAN
❖ Instead of compressing records one at a time, compresses whole batch
❖ Message batches can be compressed and sent to Kafka broker/server in one
go
❖ Message batch get written in compressed form in log partition
❖ don’t get decompressed until they consumer
❖ GZIP, Snappy and LZ4 compression protocols supported
Read more at Kafka documents on end to end compression

Consulting
™
Kafka Compression Config
Kafka Broker Compression Config
compression.type Configures compression type for topics.
Can be set to codecs ‘gzip', 'snappy', ‘lz4' or
‘uncompressed’. If set to 'producer' then it
retains compression codec set by the
producer (so it does not have to be
uncompressed and then recompressed).
Default:
producer

Consulting
™
Pull vs. Push/Streams: Pull
❖ With Kafka consumers pull data from brokers
❖ Other systems are push based or stream data to consumers
❖ Messaging is usually a pull-based system (SQS, most MOM is pull)
❖ if consumer fall behind, it catches up later when it can
❖ Pull-based can implement aggressive batching of data
❖ Pull based systems usually implement some sort of long poll
❖ long poll keeps a connection open for response after a request for a
period
❖ Pull based systems have to pull data and then process it
❖ There is always a pause between the pull

Consulting
™
Pull vs. Push/Streams: Push
❖ Push based push data to consumers (scribe, flume, reactive streams, RxJava, Akka)
❖ push-based have problems dealing with slow or dead consumers
❖ push system consumer can get overwhelmed
❖ push based systems use back-off protocol (back pressure)
❖ consumer can indicate it is overwhelmed, (http://www.reactive-streams.org/)
❖ Push-based streaming system can
❖ send a request immediately or accumulate request and send in batches
❖ Push-based systems are always pushing data or streaming data
❖ Advantage: Consumer can accumulate data while it is processing data already sent
❖ Disadvantage: If consumer dies, how does broker know and when does data get
resent to another consumer (harder to manage message acks; more complex)

Consulting
™
MOM Consumer Message
State
❖ With most MOM it is brokers responsibility to keep track
of which messages have been consumed
❖ As message is consumed by a consumer, broker keeps
track
❖ broker may delete data quickly after consumption
❖ Trickier than it sounds (acknowledgement feature), lots
of state to track per message, sent, acknowledge

Consulting
™
Kafka Consumer Message
State
❖ Kafka topic is divided into ordered partitions - A topic partition
gets read by only one consumer per consumer group
❖ Offset data is not tracked per message - a lot less data to
track
❖ just stores offset of each consumer group, partition pairs
❖ Consumer sends offset Data periodically to Kafka Broker
❖ Message acknowledgement is cheap compared to MOM
❖ Consumer can rewind to older offset (replay)
❖ If bug then fix, rewind consumer and replay

Consulting
™
Message Delivery Semantics
❖ At most once
❖ Messages may be lost but are never redelivered
❖ At least once
❖ Messages are never lost but may be redelivered
❖ Exactly once
❖ this is what people actually want, each message is
delivered once and only once

Consulting
™
Consumer: Message Delivery
Semantics
❖ "at-most-once" - Consumer reads message, save offset, process message
❖ Problem: consumer process dies after saving position but before processing
message - consumer takes over starts at last position and message never
processed
❖ “at-least-once" - Consumer reads message, process messages, saves offset
❖ Problem: consumer could crash after processing message but before saving
position - consumer takes over receives already processed message
❖ “exactly once” - need a two-phase commit for consumer position, and message
process output - or, store consumer message process output in same location
as last position
❖ Kafka offers the first two and you can implement the third

Consulting
™
Kafka Producer
Acknowledgement
❖ Kafka's offers operational predictable semantics
❖ When publishing a message, message get committed to the
log
❖ Durable as long as at least one replica lives
❖ If Producer connection goes down during of send
❖ Producer not sure if message sent; resends until message
sent ack received (log could have duplicates)
❖ Important: use message keys, idempotent messages
❖ Not guaranteed to not duplicate from producer retry

Consulting
™
Producer Durability Levels
❖ Producer can specify durability level
❖ Producer can wait on a message being committed. Waiting for
commit ensures all replicas have a copy of the message
❖ Producer can send with no acknowledgments (0)
❖ Producer can send with acknowledgment from partition leader (1)
❖ Producer can send and wait on acknowledgments from all replicas
(-1) (default)
❖ As of June 2017: producer can ensure a message or group of
messages was sent "exactly once"

Consulting
™
Improved Producer (coming
soon)
❖ New feature:
❖ exactly once delivery from producer, atomic write
across partitions, (coming soon),
❖ producer sends sequence id, broker keeps track if
producer already sent this sequence
❖ if producer tries to send it again, it gets ack for duplicate,
but nothing is save to log
❖ NO API changed

Consulting
™
Coming Soon: Kafka Atomic Log
Writes
Consumer only see
committed logs
Marker written to log to signi
what has been
successful transacted
Transaction coordinator and
transaction log maintain stat
New producer API
for transactions

Consulting
™
Kafka Replication
❖ Kafka replicates each topic's partitions across a configurable number of Kafka brokers
❖ Kafka is replicated by default not a bolt-on feature
❖ Each topic partition has one leader and zero or more followers
❖ leaders and followers are called replicas
❖ replication factor = 1 leader + N followers
❖ Reads and writes always go to leader
❖ Partition leadership is evenly shared among Kafka brokers
❖ logs on followers are in-sync to leader's log - identical copy - sans un-replicated
offsets
❖ Followers pull records in batches records from leader like a regular Kafka consumer

Consulting
™
Kafka Broker Failover
❖ Kafka keeps track of which Kafka Brokers are alive (in-sync)
❖ To be alive Kafka Broker must maintain a ZooKeeper session (heart beat)
❖ Followers must replicate writes from leader and not fall "too far" behind
❖ Each leader keeps track of set of "in sync replicas" aka ISRs
❖ If ISR/follower dies, falls behind, leader will removes follower from ISR set -
falling behind replica.lag.time.max.ms > lag
❖ Kafka guarantee: committed message not lost, as long as one live ISR -
"committed" when written to all ISRs logs
❖ Consumer only reads committed messages

Consulting
™
Replicated Log Partitions
❖ A Kafka partition is a replicated log - replicated log is a distributed data
system primitive
❖ Replicated log useful for building distributed systems using state-machines
❖ A replicated log models “coming into consensus” on ordered series of values
❖ While leader stays alive, all followers just need to copy values and
ordering from leader
❖ When leader does die, a new leader is chosen from its in-sync followers
❖ If producer told a message is committed, and then leader fails, new elected
leader must have that committed message
❖ More ISRs; more to elect during a leadership failure

Consulting
™
Kafka Consumer Replication
Redux
❖ What can be consumed?
❖ "Log end offset" is offset of last record written
to log partition and where Producers write to
next
❖ "High watermark" is offset of last record
successfully replicated to all partitions followers
❖ Consumer only reads up to “high watermark”.
Consumer can’t read un-replicated data

Consulting
™
Kafka Broker Replication
Config
Kafka Broker Config
auto.leader.rebalance.enable Enables auto leader balancing. boolean TRUE
leader.imbalance.check.interval.seconds The interval for checking for partition
leadership balancing.
long 300
leader.imbalance.per.broker.percentage Leadership imbalance for each broker. If
imbalance is too high then a rebalance is
triggered.
int 10
min.insync.replicas When a producer sets acks to all (or -1), This
setting is the minimum replicas count that
must acknowledge a write for the write to be
considered successful. If not met, then the
producer will raise an exception (either
NotEnoughReplicas or
NotEnoughReplicasAfterAppend).
int 1
num.replica.fetchers Replica fetcher count. Used to replicate
messages from a broker that has a
leadership partition. Increase this if followers
are falling behind.
int 1

Consulting
™
Kafka Replication Broker
Config 2
Kafka Broker Config
NAME DESCRIPTION
replica.high.watermark.checkpoint.inter
val.ms
The frequency with which the high watermark is saved out to disk used for
knowing what consumers can consume. Consumer only reads up to “high
watermark”. Consumer can’t read un-replicated data.
replica.lag.time.max.ms Determines which Replicas are in the ISR set and which are not. ISR is is
important for acks and quorum.
replica.socket.receive.buffer.bytes The socket receive buffer for network requests
replica.socket.timeout.ms The socket timeout for network requests. Its value should be at least
replica.fetch.wait.max.ms
unclean.leader.election.enable What happens if all of the nodes go down?
Indicates whether to enable replicas not in the ISR. Replicas that are not in-
sync. Set to be elected as leader as a last resort, even though doing so may
result in data loss. Availability over Consistency. True is the default.

Consulting
™
Kafka and Quorum
❖ Quorum is number of acknowledgements required and
number of logs that must be compared to elect a leader
such that there is guaranteed to be an overlap
❖ Most systems use a majority vote - Kafka does not use a
majority vote
❖ Leaders are selected based on having the most complete
log
❖ Problem with majority vote Quorum is it does not take
many failure to have inoperable cluster

Consulting
™
Kafka and Quorum 2
❖ If we have a replication factor of 3
❖ Then at least two ISRs must be in-sync before the
leader declares a sent message committed
❖ If a new leader needs to be elected then, with no
more than 3 failures, the new leader is guaranteed to
have all committed messages
❖ Among the followers there must be at least one
replica that contains all committed messages

Consulting
™
Kafka Quorum Majority of
ISRs
❖ Kafka maintains a set of ISRs
❖ Only this set of ISRs are eligible for leadership election
❖ Write to partition is not committed until all ISRs ack write
❖ ISRs persisted to ZooKeeper whenever ISR set
changes

Consulting
™
Kafka Quorum Majority of
ISRs 2
❖ Any replica that is member of ISRs are eligible to be
elected leader
❖ Allows producers to keep working with out majority
nodes
❖ Allows a replica to rejoin ISR set
❖ must fully re-sync again
❖ even if replica lost un-flushed data during crash

Consulting
™
All nodes die at same time. Now
what?
❖ Kafka's guarantee about data loss is only valid if at least one replica
being in-sync
❖ If all followers that are replicating a partition leader die at once, then
data loss Kafka guarantee is not valid.
❖ If all replicas are down for a partition, Kafka chooses first replica (not
necessarily in ISR set) that comes alive as the leader
❖ Config unclean.leader.election.enable=true is default
❖ If unclean.leader.election.enable=false, if all replicas are down for
a partition, Kafka waits for the ISR member that comes alive as new
leader.

Consulting
™
Producer Durability Acks
❖ Producers can choose durability by setting acks to - 0, 1 or all replicas
❖ acks=all is default, acks happens when all current in-sync replicas (ISR) have
received the message
❖ If durability over availability is prefer
❖ Disable unclean leader election
❖ Specify a minimum ISR size
❖ trade-off between consistency and availability
❖ higher minimum ISR size guarantees better consistency
❖ but higher minimum ISR reduces availability since partition won't be
unavailable for writes if size of ISR set is less than threshold

Consulting
™
Quotas
❖ Kafka has quotas for Consumers and Producers
❖ Limits bandwidth they are allowed to consume
❖ Prevents Consumer or Producer from hogging up all
Broker resources
❖ Quota is by client id or user
❖ Data is stored in ZooKeeper; changes do not
necessitate restarting Kafka

Consulting
™
Kafka Low-Level Review
❖ How would you prevent a denial of service attack from a poorly
written consumer?
❖ What is the default producer durability (acks) level?
❖ What happens by default if all of the Kafka nodes go down at
once?
❖ Why is Kafka record batching important?
❖ What are some of the design goals for Kafka?
❖ What are some of the new features in Kafka as of June 2017?
❖ What are the different message delivery semantics?

™
Consulting
Kafka Log Compaction Architecture
Kafka Log
Compaction
Design discussion of Kafka Log
Compaction
Kafka Architecture: Log
Compaction

Consulting
™
Kafka Log Compaction
Overview
❖ Recall Kafka can delete older records based on
❖ time period
❖ size of a log
❖ Kafka also supports log compaction for record key
compaction
❖ Log compaction: keep latest version of record and
delete older versions

Consulting
™
Log Compaction
❖ Log compaction retains last known value for each record key
❖ Useful for restoring state after a crash or system failure, e.g., in-
memory service, persistent data store, reloading a cache
❖ Data streams is to log changes to keyed, mutable data,
❖ e.g., changes to a database table, changes to object in in-memory
microservice
❖ Topic log has full snapshot of final values for every key - not just
recently changed keys
❖ Downstream consumers can restore state from a log compacted
topic

Consulting
™
Log Compaction Structure
❖ Log has head and tail
❖ Head of compacted log identical to a traditional Kafka log
❖ New records get appended to the head
❖ Log compaction works at tail of the log
❖ Tail gets compacted
❖ Records in tail of log retain their original offset when
written after compaction

Consulting
™
Compaction Tail/Head

Consulting
™
Log Compaction Basics
❖ All offsets remain valid, even if record at offset has been
compacted away (next highest offset)
❖ Compaction also allows for deletes. A message with a key and a
null payload acts like a tombstone (a delete marker for that key)
❖ Tombstones get cleared after a period.
❖ Log compaction periodically runs in background by recopying
log segments.
❖ Compaction does not block reads and can be throttled to avoid
impacting I/O of producers and consumers

Consulting
™
Log Compaction Cleaning

Consulting
™
Log Compaction Guarantees
❖ If consumer stays caught up to head of the log, it sees every record that is
written.
❖ Topic config min.compaction.lag.ms used to guarantee minimum period
that must pass before message can be compacted.
❖ Consumer sees all tombstones as long as the consumer reaches head of log
in a period less than the topic config delete.retention.ms (the default is 24
hours).
❖ Compaction will never re-order messages, just remove some.
❖ Offset for a message never changes.
❖ Any consumer reading from start of the log, sees at least final state of all
records in order they were written

Consulting
™
Log Cleaner
❖ Log cleaner does log compaction.
❖ Has a pool of background compaction threads that recopy log
segments, removing records whose key appears in head of log
❖ Each compaction thread works as follows:
❖ Chooses topic log that has highest ratio: log head to log tail
❖ Recopies log from start to end removes records whose keys occur later
❖ As log partition segments cleaned, they get swapped into log partition
❖ Additional disk space required: only one log partition segment
❖ not whole partition

Consulting
™
Topic Config for Log
Compaction
❖ To turn on compaction for a topic
❖ topic config log.cleanup.policy=compact
❖ To start compacting records after they are written
❖ topic config log.cleaner.min.compaction.lag.ms
❖ Records wont be compacted until after this period

Consulting
™
Broker Config for Log
CompactionKafka Broker Log Compaction Config
log.cleaner.backoff.ms Sleep period when no logs need cleaning long 15,000
log.cleaner.dedupe.buffer.size The total memory for log dedupe process for all
cleaner threads
long 134,217,728
log.cleaner.delete.retention.ms How long record delete markers (tombstones) are
retained.
long 86,400,000
log.cleaner.enable Turn on the Log Cleaner. You should turn this on if
any topics are using clean.policy=compact.
boolean TRUE
log.cleaner.io.buffer.size Total memory used for log cleaner I/O buffers for all
cleaner threads
int 524,288
log.cleaner.io.max.bytes.per.second This is a way to throttle the log cleaner if it is taking
up too much time.
double 1.7976931348623157E3
08
log.cleaner.min.cleanable.ratio The minimum ratio of dirty head log to total log
(head and tail) for a log to get selected for cleaning.
double 0.5
log.cleaner.min.compaction.lag.ms Minimum time period a new message will remain
uncompacted in the log.
long 0
log.cleaner.threads Threads count used for log cleaning. Increase this if you have a lot
of log compaction going on across many topic log partitions.
int 1
log.cleanup.policy The default cleanup policy for segment files that are beyond their
retention window. Valid policies are: "delete" and “compact”. You
could use log compaction just for older segment files. instead of
deleting them, you could just compact them.
list [delete]

Consulting
™
Log Compaction Review
❖ What are three ways Kafka can delete records?
❖ What is log compaction good for?
❖ What is the structure of a compacted log? Describe the
structure.
❖ After compaction, do log record offsets change?
❖ What is a partition segment?

Consulting
™
References
❖ Learning Apache Kafka, Second Edition 2nd Edition by Nishant Garg (Author),
2015, ISBN 978-1784393090, Packet Press
❖ Apache Kafka Cookbook, 1st Edition, Kindle Edition by Saurabh Minni (Author),
2015, ISBN 978-1785882449, Packet Press
❖ Kafka Streams for Stream processing: A few words about how Kafka works,
Serban Balamaci, 2017, Blog: Plain Ol' Java
❖ Kafka official documentation, 2017
❖ Why we need Kafka? Quora
❖ Why is Kafka Popular? Quora
❖ Why is Kafka so Fast? Stackoverflow
❖ Kafka growth exploding (Tech Republic)

™
Consulting
Working with Kafka Producers
Kafka Producer
Advanced
Working with producers in
Java
Details and advanced topics

Consulting
™
Objectives Create Producer
❖ Cover advanced topics regarding Java Kafka
Consumers
❖ Custom Serializers
❖ Custom Partitioners
❖ Batching
❖ Compression
❖ Retries and Timeouts

Consulting
™
Kafka Producer
❖ Kafka client that publishes records to Kafka cluster
❖ Thread safe
❖ Producer has pool of buffer that holds to-be-sent
records
❖ background I/O threads turning records into request
bytes and transmit requests to Kafka
❖ Close producer so producer will not leak resources

Consulting
™
Kafka Producer Send, Acks and
Buffers
❖ send() method is asynchronous
❖ adds the record to output buffer and return right away
❖ buffer used to batch records for efficiency IO and compression
❖ acks config controls Producer record durability. ”all" setting
ensures full commit of record, and is most durable and least fast
setting
❖ Producer can retry failed requests
❖ Producer has buffers of unsent records per topic partition (sized at
batch.size)

Consulting
™
Kafka Producer: Buffering and
batching
❖ Kafka Producer buffers are available to send immediately as fast as broker can
keep up (limited by inflight max.in.flight.requests.per.connection)
❖ To reduce requests count, set linger.ms > 0
❖ wait up to linger.ms before sending or until batch fills up whichever comes
first
❖ Under heavy load linger.ms not met, under light producer load used to
increase broker IO throughput and increase compression
❖ buffer.memory controls total memory available to producer for buffering
❖ If records sent faster than they can be transmitted to Kafka then this buffer
gets exceeded then additional send calls block. If period blocks
(max.block.ms) after then Producer throws a TimeoutException

Consulting
™
Producer Acks
❖ Producer Config property acks
❖ (default all)
❖ Write Acknowledgment received count required from
partition leader before write request deemed complete
❖ Controls Producer sent records durability
❖ Can be all (-1), none (0), or leader (1)

Consulting
™
Acks 0 (NONE)
❖ acks=0
❖ Producer does not wait for any ack from broker at all
❖ Records added to the socket buffer are considered sent
❖ No guarantees of durability - maybe
❖ Record Offset returned is set to -1 (unknown)
❖ Record loss if leader is down
❖ Use Case: maybe log aggregation

Consulting
™
Acks 1 (LEADER)
❖ acks=1
❖ Partition leader wrote record to its local log but responds
without followers confirmed writes
❖ If leader fails right after sending ack, record could be lost
❖ Followers might have not replicated the record
❖ Record loss is rare but possible
❖ Use Case: log aggregation

Consulting
™
Acks -1 (ALL)
❖ acks=all or acks=-1
❖ Leader gets write confirmation from full set of ISRs before
sending ack to producer
❖ Guarantees record not be lost as long as one ISR remains alive
❖ Strongest available guarantee
❖ Even stronger with broker setting min.insync.replicas
(specifies the minimum number of ISRs that must acknowledge
a write)
❖ Most Use Cases will use this and set a min.insync.replicas >
1

Consulting
™
KafkaProducer config Acks

Consulting
™
Producer Buffer Memory Size
❖ Producer config property: buffer.memory
❖ default 32MB
❖ Total memory (bytes) producer can use to buffer records
to be sent to broker
❖ Producer blocks up to max.block.ms if buffer.memory
is exceeded
❖ if it is sending faster than the broker can receive,
exception is thrown

Consulting
™
Batching by Size
❖ Producer config property: batch.size
❖ Default 16K
❖ Producer batch records
❖ fewer requests for multiple records sent to same partition
❖ Improved IO throughput and performance on both producer and server
❖ If record is larger than the batch size, it will not be batched
❖ Producer sends requests containing multiple batches
❖ batch per partition
❖ Small batch size reduce throughput and performance. If batch size is too big,
memory allocated for batch is wasted

Consulting
™
Batching by Time and Size - 1
❖ Producer config property: linger.ms
❖ Default 0
❖ Producer groups together any records that arrive before
they can be sent into a batch
❖ good if records arrive faster than they can be sent out
❖ Producer can reduce requests count even under
moderate load using linger.ms

Consulting
™
Batching by Time and Size - 2
❖ linger.ms adds delay to wait for more records to build up so
larger batches are sent
❖ good brokers throughput at cost of producer latency
❖ If producer gets records who size is batch.size or more for a
broker’s leader partitions, then it is sent right away
❖ If Producers gets less than batch.size but linger.ms interval
has passed, then records for that partition are sent
❖ Increase to improve throughput of Brokers and reduce broker
load (common improvement)

Consulting
™
Compressing Batches
❖ Producer config property: compression.type
❖ Default 0
❖ Producer compresses request data
❖ By default producer does not compress
❖ Can be set to none, gzip, snappy, or lz4
❖ Compression is by batch
❖ improves with larger batch sizes
❖ End to end compression possible if Broker config “compression.type” set to
producer. Compressed data from producer sent to log and consumer by broker

Consulting
™
Batching and Compression
Example

Consulting
™
Custom Serializers
❖ You don’t have to use built in serializers
❖ You can write your own
❖ Just need to be able to convert to/fro a byte[]
❖ Serializers work for keys and values
❖ value.serializer and key.serializer

Consulting
™
Custom Serializers Config

Consulting
™
Custom Serializer

Consulting
™
StockPrice

Consulting
™
Broker Follower Write Timeout
❖ Producer config property: request.timeout.ms
❖ Default 30 seconds (30,000 ms)
❖ Maximum time broker waits for confirmation from
followers to meet Producer acknowledgment
requirements for ack=all
❖ Measure of broker to broker latency of request
❖ 30 seconds is high, long process time is indicative of
problems

Consulting
™
Producer Request Timeout
❖ Producer config property: request.timeout.ms
❖ Default 30 seconds (30,000 ms)
❖ Maximum time producer waits for request to complete to
broker
❖ Measure of producer to broker latency of request
❖ 30 seconds is very high, long request time is an
indicator that brokers can’t handle load

Consulting
™
Producer Retries
❖ Producer config property: retries
❖ Default 0
❖ Retry count if Producer does not get ack from Broker
❖ only if record send fail deemed a transient error (API)
❖ as if your producer code resent record on failed attempt
❖ timeouts are retried, retry.backoff.ms (default to 100 ms)
to wait after failure before retry

Consulting
™
Retry, Timeout, Back-off
Example

Consulting
™
Producer Partitioning
❖ Producer config property: partitioner.class
❖ org.apache.kafka.clients.producer.internals.DefaultPartitioner
❖ Partitioner class implements Partitioner interface
❖ Default Partitioner partitions using hash of key if record
has key
❖ Default Partitioner partitions uses round-robin if record
has no key

Consulting
™
Configuring Partitioner

Consulting
™
StockPricePartitioner

Consulting
™
StockPricePartitioner
partition()

Consulting
™
Producer Interception
❖ Producer config property: interceptor.classes
❖ empty (you can pass an comma delimited list)
❖ interceptors implementing ProducerInterceptor interface
❖ intercept records producer sent to broker and after acks
❖ you could mutate records

Consulting
™
KafkaProducer - Interceptor
Config

Consulting
™
KafkaProducer
ProducerInterceptor

Consulting
™
ProducerInterceptor onSend
ic=stock-prices2 key=UBER value=StockPrice{dollars=737, cents=78, name='
Output

Consulting
™
ProducerInterceptor onAck
onAck topic=stock-prices2, part=0, offset=18360
Output

Consulting
™
ProducerInterceptor the rest

Consulting
™
KafkaProducer send() Method
❖ Two forms of send with callback and with no callback both return Future
❖ Asynchronously sends a record to a topic
❖ Callback gets invoked when send has been acknowledged.
❖ send is asynchronous and return right away as soon as record has added to
send buffer
❖ Sending many records at once without blocking for response from Kafka broker
❖ Result of send is a RecordMetadata
❖ record partition, record offset, record timestamp
❖ Callbacks for records sent to same partition are executed in order

Consulting
™
KafkaProducer send()
Exceptions
❖ InterruptException - If the thread is interrupted while
blocked (API)
❖ SerializationException - If key or value are not valid
objects given configured serializers (API)
❖ TimeoutException - If time taken for fetching metadata or
allocating memory exceeds max.block.ms, or getting acks
from Broker exceed timeout.ms, etc. (API)
❖ KafkaException - If Kafka error occurs not in public API.
(API)

Consulting
™
Using send method

Consulting
™
KafkaProducer flush() method
❖ flush() method sends all buffered records now (even if
linger.ms > 0)
❖ blocks until requests complete
❖ Useful when consuming from some input system and
pushing data into Kafka
❖ flush() ensures all previously sent messages have been
sent
❖ you could mark progress as such at completion of flush

Consulting
™
KafkaProducer close()
❖ close() closes producer
❖ frees resources (threads and buffers) associated with producer
❖ Two forms of method
❖ both block until all previously sent requests complete or duration
passed in as args is exceeded
❖ close with no params equivalent to close(Long.MAX_VALUE,
TimeUnit.MILLISECONDS).
❖ If producer is unable to complete all requests before the timeout
expires, all unsent requests fail, and this method fails

Consulting
™
Orderly shutdown using close

Consulting
™
Wait for clean close

Consulting
™
KafkaProducer partitionsFor()
method
❖ partitionsFor(topic) returns meta data for partitions
❖ public List<PartitionInfo> partitionsFor(String topic)
❖ Get partition metadata for give topic
❖ Produce that do their own partitioning would use this
❖ for custom partitioning
❖ PartitionInfo(String topic, int partition, Node leader,
Node[] replicas, Node[] inSyncReplicas)
❖ Node(int id, String host, int port, optional String rack)

Consulting
™
KafkaProducer metrics()
method
❖ metrics() method get map of metrics
❖ public Map<MetricName,? extends Metric> metrics()
❖ Get the full set of producer metrics
MetricName(
String name,
String group,
String description,
Map<String,String> tags
)

Consulting
™
Metrics producer.metrics()
❖ Call producer.metrics()
❖ Prints out metrics to log

Consulting
™
Metrics producer.metrics()
output
Metric producer-metrics, record-queue-time-max, 508.0,
The maximum time in ms record batches spent in the record accumulator.
17:09:22.721 [pool-1-thread-9] INFO c.c.k.p.MetricsProducerReporter -
Metric producer-node-metrics, request-rate, 0.025031289111389236,
The average number of requests sent per second.
Metric producer-metrics, records-per-request-avg, 205.55263157894737,
The average number of records per request.
Metric producer-metrics, record-size-avg, 71.02631578947368,
The average record size
Metric producer-node-metrics, request-size-max, 56.0,
The maximum size of any request sent in the window.
Metric producer-metrics, request-size-max, 12058.0,
The maximum size of any request sent in the window.
Metric producer-metrics, compression-rate-avg, 0.41441360272859273,
The average compression rate of record batches.

Consulting
™
Metrics via JMX

Kafka Tutorial - Introduction to Apache Kafka (Part 1)

Recommended

More Related Content

What's hot (20)

Similar to Kafka Tutorial - Introduction to Apache Kafka (Part 1) (20)

More from Jean-Paul Azar (7)

Recently uploaded (20)

Kafka Tutorial - Introduction to Apache Kafka (Part 1)