Design Patterns for working with Fast Data

© 2016 MapR Technologies 1© 2016 MapR Technologies 1© 2016 MapR Technologies
Design Patterns for working with Fast Data

© 2016 MapR Technologies 2© 2016 MapR Technologies 2
Abstract
• This presentation provides an introduction to Apache Kafka,
describes how to use the Kafka API, and illustrates strategies for
maximizing the throughput of Kafka pipelines to accommodate
fast data.
• The code examples used during this talk are available
at github.com/iandow/design-patterns-for-fast-data.

• Ian Downard is a technical evangelist for MapR where he is
focused on creating developer-friendly ways to use the MapR
Converged Data Platform.
• Personal Blog: http://www.bigendiandata.com
• Twitter: https://twitter.com/iandownard
• GitHub: https://github.com/iandow
• LinkedIn: https://www.linkedin.com/in/iandownard
Author

What is Kafka?
• A distributed publish-subscribe messaging system.
• Persists on disk (not “in-memory”).
• Supports true streaming (not just fast batching).
• Can be used for data storage AND data processing (not just
storage)
• Oriented around small messages and dynamic data sets (i.e.
“streams”).

Service A
Service B
Service C
Logs
Service 1
Service 2
Service 3
n Producers
m Consumers
How many connections?

• Elastic, fault tolerant, and low latency.
– Fits well into microservice architectures.
• Low CapEx message bus for Big Data.
• Supports both real-time processing and short-term storage.
• Kafka is becoming the default for inter-service communication.
Kafka Strengths

Who is streaming data?
• Retail: orders, sales, shipments, price adjustments
• Financial Services: stock prices, credit card fraud
• Web sites: clicks, impressions, searches, web click fraud
• Operating Systems: machine metrics, and logs

Some examples of streaming datasets
• Social Media
– https://dev.twitter.com/streaming/overview
• IoT
– Lots of examples at https://data.sparkfun.com/streams/
– http://dweet.io is pretty cool
• Servers:
– /var/log/syslog
– tcpdump
• Banking
– Yahoo Finance API: https://github.com/cjmatta/TwitterStream,
http://meumobi.github.io/stocks%20apis/2016/03/13/get-realtime-stock-quotes-
yahoo-finance-api.html
– NYSE markets: http://www.nyxdata.com/Data-Products/Daily-TAQ
• Other
– screen scrape anything

“Text mining is the application of natural language processing
techniques and analytical methods to text data in order to derive
relevant information.”
http://adilmoujahid.com/posts/2014/07/twitter-analytics/

https://heroku.github.io/kafka-demo/

● Messages contain <Key, Value>
● delivered in order
● transferred as bytes (you specify serializer / deserializer)
● remain available for TTL
● with at least once semantics
1 2 3 4 5 6 7 8
topic “topic1”
ConsumerProducer
ConsumerProducer
Message Log

● Producers send messages one at a time
● Consumers subscribe to topics, poll for messages,
and receive records many at a time
“topic1”
ConsumerProducer
ConsumerProducer
1 2 3 4 5 6 7 8
Message Log

● Producer sees a topic as one logic object.
● The order of messages within a single partition is guaranteed
1 2 3 4
“topic1”, partition 1
Producer
1 2 3 4
Partitioning

● A single partition can only be read by one consumer.
● Partitioning facilitates concurrent consumers.
● New consumers are automatically assigned a partition.
Producer Load balanced
Consumer(s)
1 2 3 4
1 2 3 4
Partitioning

● Calling commit marks where a consumer has finished
processing.
1 2 3 4 5 6 7 8
Consumer
commit (p1, o6)
Commit

● Thanks to commit, when a consumer crashes, we let it.
1 2 3 4 5 6 7 8
Consumer
Commit

● When another consumer takes its place, it continues where
the old consumer left off.
1 2 3 4 5 6 7 8
New Consumer
Commit
cursor

Kafka depends on Zookeeper
• Leader election – each partition has a leader
• Cluster membership – broker up/down status
• Topic Configurations – topic names, partitions, replicas, TTLs…

DEMO: Kafka CLI
• Create, alter, pub/sub, and delete topic.
• See where topics are saved.

DEMO: Java API
• Basic producer/consumer example

DEMO: Heroku Kafka-as-a-Service
• Demo app:
– https://github.com/iandow/heroku-kafka-demo-java
• Docs and Pricing:
– https://devcenter.heroku.com/articles/kafka-on-heroku
– $1,500/month !!!
– Oh, but rounding credits.

DEMO: Kafka in a web app
• How to send a message
– How to ensure delivery guarantee
– How to handle failure
• Demonstrate it:
– Change 20 second timeout in DemoResources.java
– Observe the java.util.concurrent.Timeout in the user specified callback
– Also observe that the message may ultimately get there, (if its in transit)
• To get rid of Dropwizard metrics, change timeout in DemoApplication.java
– reporter.start(1, TimeUnit.MINUTES);

© 2016 MapR Technologies 25© 2016 MapR Technologies 25© 2016 MapR Technologies© 2016 MapR Technologies
Consumer Groups, Cursors, and Partitions

ProducerRecord vs ConsumerRecord
• Producers send messages one at a time as ProducerRecord
• Consumers subscribe to topics and poll for messages and
receive them many-at-a-time as ConsumerRecords

Consumer Groups, Cursors, and Partitions
• Consumer Groups – groups of consumers working together
• Cursors – track positions of consumers
• Partitions – subdivide a topic for scale
• Replication Factor – replicate partitions across nodes for fault
tolerance

Partition Balancing
• Available partitions are balanced among consumers for a given group
– Single partition can be read by only one consumer
– A single consumer can read multiple partitions
– Examples:
• 3 partitions, 2 consumers: 2 & 1
• 5 partitions, 1 consumer: 5
• 3 partitions, 4 consumers: 1, 1, 1, 0 (4th consumer gets no messages)
• When new consumers join or new partitions added, rebalance occurs
– Some consumers will gain or lose partitions and there may be a slight pause in
message delivery while this happens
– Can cause duplicates if consumers don’t properly commit cursors

Consumer Groups
• Only one consumer IN A GROUP can consume one topic/partition
simultaneously
• Consumers in different groups will see the same messages if they
read the same topic/partition
http://kafka.apache.org/documentation.html

Cursor Persistence
• Ensure consumers pick up from where they left off after a stop
• Tracks last known commit point for message reads
• By default, auto commit after 5s (auto.commit.interval.ms)
– If slow consumer code, auto commit() might go out before processed
– You can explicitly commit with consumer.commit()
• If you need to concurrent consumers, would you use consumer groups
or topic partitions?
8 7 6 5 4 3 2 1
topic “topic1”
Consumer
Group A
Consumer
Group B
Cursor A
Cursor B

At Least Once semantics
• Messages may be seen multiple times
– Producer side
• Retransmit on network
• Producer might fail and resend previously sent messages on restart
– Consumer side
• Client might crash after get but before cursor.commit()
• Applications must be idempotent message processors

Batching Behavior
• Producer.send() places messages in client-side memory buffer
• That buffer is not flushed until one of three criteria met:
– Timeout (linger.ms)
– Max memory (buffer.memory)
– Producer.flush() called

Serialization
• Each Kafka message is a byte array.
• You must convert <K,V> to/from bytes with a serializer
• Two provided out of the box
– ByteArraySerializer/ByteArrayDeserializer
– StringSerializer/StringDeserializer
• You can write your own serializer.

Key Concept - Serialization
• If you’re streaming data that is a byte array, use the byte array
serializer.
• If you’re streaming data that is a String, use the String serializer.
• If you’re streaming data that is a POJO, use ???

How to stream a POJO
1. Make it serializable
– Otherwise you’ll get a java.io.NotSerializableException later when we
write the object to a byte array
2. Write the object to a byte array
3. Send that byte array
– Why can’t we just send the Person
object?
4. In the consumer, poll the topic, then convert record.value() to
your POJO.

Problems with streaming POJOs
• What if your raw data is dirty, and casting fails because you
streamed a record with missing POJO fields?
• The overhead of encoding/decoding a POJO to/from bytes is
redundant, especially if this is done at multiple stages in the
Kafka pipeline.
• It’s better to postpone parsing/casting downstream.

How would you stream this?
080449201DAA T 00000195700000103000N0000000000000004CT1000100710071007
080449201DAA T 00000131000000107000N0000000000000005CT10031007
080449201DAA T 00000066600000089600N0000000000000006CT10041005
080449201DAA T 00000180200000105100N0000000000000007CT100310051009
080449201DAA T 00000132200000089700N0000000000000008CT100410051005
080449201DAA T 00000093500000089400N0000000000000009CT10031007
080449201DAA T 00000075100000105400N0000000000000010CT100410081006
080449201DAA T 00000031300000088700N0000000000000011CT1004100810081007
080449201DAA T 00000021800000105100N0000000000000012CT100410091005
080449201DAA T 00000191300000104500N0000000000000013CT10041006
080449201DAA T 00000124500000105100N0000000000000014CT1001100610071008

Option A?
• Parse each record into a JSON object,
• Publish json_object.toString()
• Stream with with String serializer

Option B?
• Create a POJO (“Tick”) with
attributes for each field
• Parse each record and create a
Tick object
• Stream Tick objects using a
custom serializer

Option C?
• Create a POJO “Tick”
with a single byte[]
attribute, and getters
that return fields by
directly indexing the
byte[]
• Annotate getters with
@JsonProperty

A
B
C

Major Performance Factors
• Replication turned on?
• Streaming large messages?
• Producers sending synchronously or async.?
• Send buffer flushed until one of three criteria met:
– Timeout (linger.ms)
– Max memory (buffer.memory)
– Producer.flush() called
• Topics / Producer affinity

• What affects producer latency?
– Using send futures? (i.e. synchronous sends)
– acks=all,1,or 0 (ack after leader logs / after all replicas log / or not at all)
– ”Nagle’s algorithm” and the linger.ms config
• What affects producer throughput?
– Producer batch.size
– Concurrent producer.send()
– Number of topics
– Message size
• What affects consumer throughput?
– Partitions

• What affects Kafka server stability?
– Garbage Collection: Long GCs from sending large messages can break
the kafka/zookeeper connection.
– Topics can easily consume disk space
– May need to increase file handles
• Topic data and index files are in log.dir (e.g. /tmp/kafka-logs)
• Kafka keeps an open file for each topic.
– Reference:
http://www.michael-noll.com/blog/2013/03/13/running-a-multi-broker-apache-kafka-cluster-on-a-single-node/

Read vs Processing Parallelism
• Consumer read parallelism
– Concurrent reads can be achieved via topic partitioning and consumer
groups
– If your consumer logic is CPU-bound, use partitions or consumer groups
• Downstream Processing Parallelism
– If you need more processing threads than read threads, then “fan-out”
the data to more topics.

What I mean by, “Fan-out”:
vv
SQL
15 min. roll
Intra-day
Trades
Ingest trades
archived
trades
Microservice
Index by
sender/receiver
Tick
stream
Microservice
STREAMING
https://www.mapr.com/appblueprint/overview

Performance Anti-Pattern
• Problem: I want to count all the messages read in each consumer
thread.
• Anti-pattern: I’ll just increment a counter in a synchronized block.

• I want to count all the messages read in each consumer thread.
• Fine, I’ll just increment a counter in a synchronized block.
– Lesson learned: synchronization really hurts throughput.
– Solution: I decreased the cost of metric collection to near zero by giving
each thread its own metrics structure, and every second or so, updating a
shared metric structure with locks.

• I want to be sure never to read a duplicate message.
• Fine, I’ll just commit the read point for every message I consume.

• I want to be sure never to read a duplicate message.
• Fine, I’ll just commit the read point for every message I consume.
– Lesson learned, you’re still not guaranteed not to get a duplicate
message. Kafka consumers must be idempotent.
– Solution: Postpone deduplication downstream. (i.e. Give up till’ later.)

• I want to stream JSON strings to simplify DB persistence and
columnar selections in Spark.
• Fine, I’ll just convert every message to a JSON object at the first
step of my pipeline and use a JSON serializer.

• I want to stream JSON strings to simplify DB persistence and
columnar selections in Spark.
• Fine, I’ll just convert every message to a JSON object at the first step
of my pipeline and use a JSON serializer.
– Lesson learned: parsing is expensive! Handling parsing errors is expensive.
Custom serializers are not as easy I as thought.
– This anti-pattern is illustrated here: https://github.com/mapr-demos/finserv-ticks-
demo/commit/35a0ddaad411bf7cec34919a4817482df2bf6462
– Solution: Stream a data structure containing a byte[] to hold raw data, and
annotate it with com.fasterxml.jackson.annotation.JsonProperty.
– Example: https://github.com/mapr-demos/finserv-ticks-
demo/blob/35a0ddaad411bf7cec34919a4817482df2bf6462/src/main/java/com/mapr/demo/finserv/Tick2.java

Good Practices
• Design your data structures to prefer arrays of objects, and
primitive types, instead of the standard Java or Scala collection
classes (e.g. HashMap).
• Maintain affinity between topic and producer thread.
– Kafka will batch multiple sends together. If a send hits only one topic, it
will benefit from writes to sequential memory locations.
– Reference code:
https://github.com/mapr-demos/finserv-application-
blueprint/blob/master/src/test/java/com/mapr/demo/finserv/ThreadCountS
peedIT.java
• Consider using numeric IDs or enumeration objects instead of
strings for keys.

Three Performance Takeaways
1. Process as little as possible in the pipeline.
– Transfer byte arrays. It’s faster than serializers based on strings, POJOs,
or custom serialization.
2. Use JSON annotation. It makes persistence easier in the final
stages. It also makes data analysis easier because it gives you
an easy schema.
3. Categorize raw data into multiple topics (aka “fan-out”). It makes
real-time analysis easier.
4. Use parameterized tests in Junit to tune Kafka.

DEMO: Tuning Kafka with Junit
https://github.com/iandow/kafka_junit_tests

How to create Junit parameterized tests
1. Annotate test class with @RunWith(Parameterized.class).
2. Create a public static method annotated with @Parameters that
returns a Collection of Objects (as Array) as test data set.
3. Create a public constructor that takes in one "row" of test data.

Benchmarks
• Lazy benchmark, by Jay Kreps at LinkedIn (2014):
– https://engineering.linkedin.com/kafka/benchmarking-apache-kafka-2-
million-writes-second-three-cheap-machines
• Answers questions like these:
– Will Kafka slow down if I never consumer messages?
– Should I stream small messages or large messages
(in order to maximize Throughput)?
– Should I stream small messages or large messages?

Kafka MapR Streams
Complies with standard Kafka API ✓ ✓
Scale to 1000s of topics ✗ ✓
Automatically span large partitions
across cluster nodes
✗ ✓
Failover consumers & producers
across mirrored clusters
✗ ✓
Real-time cluster replication across
data centers
✗ ✓
Kafka vs MapR Streams

• Inherits efficient I/O patterns from MapR-FS, allowing topics to
save much more data and replicate much faster
• Load balances topics and partitions more effectively
• Span topics across clusters in multi-datacenters and synchronize
offsets across clusters
• Inherits efficient I/O patterns from the core MapR platform
MapR Streams is a more scalable Kafka

Blueprint of a Converged Application
for real-time tick streaming
vv
SQL
15 min. roll
Intra-day
Trades
Ingest trades
archived
trades
Microservice
Index by
sender/receiver
Tick
stream
Microservice
STREAMING
https://www.mapr.com/appblueprint/overview

Open Source Engines & Tools Commercial Engines & Applications
Enterprise-Grade Platform Services
DataProcessing
Web-Scale Storage
MapR-FS MapR-DB
Search and
Others
Real Time Unified Security Multi-tenancy Disaster Recovery Global NamespaceHigh Availability
MapR Streams
Cloud and
Managed
Services
Search and
Others
UnifiedManagementandMonitoring
Search and
Others
Event StreamingDatabase
Custom
Apps
HDFS API POSIX, NFS HBase API JSON API Kafka API
MapR Converged Data Platform

• Read about Kafka and Spark Streaming
• https://www.mapr.com/blog/real-time-streaming-data-pipelines-
apache-apis-kafka-spark-streaming-and-hbase
• Read the Streaming Architecture ebook
• https://www.mapr.com/streaming-architecture-using-apache-kafka-
mapr-streams
• Check out this presentation and demo code
• https://github.com/iandow/design-patterns-for-fast-data
Next Steps:

Free training  http://learn.mapr.com

Q&A
@mapr
@iandownard
Engage with us!
mapr-technologies

Design Patterns for working with Fast Data

More Related Content

What's hot (19)

Viewers also liked (11)

Similar to Design Patterns for working with Fast Data (20)

More from MapR Technologies (18)

Recently uploaded (20)

Design Patterns for working with Fast Data

Editor's Notes