Overview of Apache Fink: the 4 G of Big Data Analytics Frameworks

Overview of Apache Flink:
the 4 G of Big Data Analytics Frameworks
Hadoop Summit Europe,
Dublin, Ireland.
April 13th, 2016
Slim Baltagi
Director, Enterprise Architecture
Capital One Financial Corporation

2
Agenda
1. How Apache Flink is a multi-purpose Big
Data Analytics Framework?
2. Why streaming analytics are emerging?
3. Why Flink is suitable for real-world
streaming analytics?
4. What are some novel use cases enabled by
Flink?
5. Who is using Flink?
6. Where do you go from here?

3
1. How Apache Flink is a multi-purpose Big Data
Analytics Framework?
1.1. What is Apache Flink Stack?
1.2. Why Apache Flink is the 4G of Big Data
Analytics?
1.3. What are Apache Flink Innovations?

4
1.1. What is Apache Flink Stack?
Gelly
Table
HadoopM/R
Storm
DataSet (Java/Scala/Python)
Batch Processing
DataStream (Java/Scala)
Stream Processing
FlinkML
Local
• Single JVM
• Embedded
• Docker
Cluster
• Standalone
• YARN,
• Mesos (WIP)
Cloud
• Google’s GCE
• Amazon’s EC2
• IBM Docker Cloud, …
ApacheBeam
Cascading
Table
MRQL
Distributed Streaming
Dataflow Engine
Zeppelin
DEPLOYSYSTEMAPIs&LIBRARIESSTORAGE
Files
• Local
• HDFS
• S3, Azure
• Alluxio
Databases
• MongoDB
• HBase
• SQL
…
Streams
• Flume
• Kafka, MapR Streams
• RabbitMQ
…
Batch Optimizer Stream Builder
SAMOA
FlinkCEP
Gelly-Stream
ApacheBeam

5
1.2. Why Apache Flink is the 4G of Big Data Analytics?
 Batch  Batch
 Interactive
 Batch
 Interactive
 Near-Real
Time Streaming
(micro-batches)
 Iterative
processing
 Hybrid
 Interactive
 Real-Time
Streaming +
Real-World
Streaming (out of
order streams,
windowing,
backpressure,
CEP, …)
 Native Iterative
processing
MapReduce Direct Acyclic
Graphs (DAG)
Dataflows
RDD: Resilient
Distributed Datasets
Cyclic Dataflows
1G 2G 3G 4G

6
Apache Flink came with many innovations.
Some of these innovations are influencing quite a few
features in other frameworks such as:
1. Custom memory management and binary
processing in Flink from day one inspired Apache
Spark to so so for its project Tungsten since
version 1.6
• https://flink.apache.org/news/2015/05/11/Juggling-with-Bits-and-Bytes.html
• https://databricks.com/blog/2015/04/28/project-tungsten-bringing-spark-
closer-to-bare-metal.html
2. DataSet API is in Flink since its early days and
inspired Apache Spark to come with its Dataset
API in version 1.6
• https://ci.apache.org/projects/flink/flink-docs-master/apis/batch/index.html
• https://databricks.com/blog/2016/01/04/introducing-spark-datasets.html

7
3. Flink’s rich windowing semantics for streaming
Flink supports windows over time, count, or
sessions
Windows can be customized with flexible triggering
conditions, to support sophisticated streaming
patterns.
Flink inspired both Apache Storm (1.0.0 was
released on April 12th , 2016) and Spark streaming
(version 2.0 is expected in May 2016) to start
supporting rich windowing
• https://storm.apache.org/2016/04/12/storm100-released.html
• http://www.slideshare.net/databricks/2016-spark-summit-east-keynote-
matei-zaharia/15

8
Some of Flink innovations are not available in other
open source tools such as:
1. The only hybrid (Real-Time Streaming + Batch)
distributed data processing engine natively
supporting many use cases: Batch, Real-Time
streaming, Machine learning, Graph processing
and Relational queries
2. Native iterations ( Iterate and DeltaIterate)
dramatically boost the performance of Machine
learning and Graph analytics requiring iterations.

9
The only hybrid (Real-Time Streaming + Batch)
open source distributed data processing engine
natively supporting many use cases:
Real-Time stream processing Machine Learning at scale
Graph AnalysisBatch Processing

10
3. Simplicity of configuration: Flink requires no
memory thresholds to configure, no complicated
network configurations, no serializers to be
configured, …
4. Little tuning required: Flink’s optimizer can
choose execution strategies automatically in any
environment.
 According to Mike Olsen, Chief Strategy Officer of
Cloudera Inc. “Spark is too knobby — it has too
many tuning parameters, and they need constant
adjustment as workloads, data volumes, user
counts change.”
Reference: http://vision.cloudera.com/one-platform/

11
5. Full support of Apache Beam (for combination of
Batch and Stream) : event time, sessions, …
References:
• The Dataflow Model: A Practical Approach to Balancing Correctness,
Latency, and Cost in Massive-Scale, Unbounded, Out-of-Order Data
Processing, 2015 http://research.google.com/pubs/pub43864.html
• Dataflow/Beam & Spark: A Programming Model Comparison, February
3rd, 2016https://cloud.google.com/dataflow/blog/dataflow-beam-and-spark-
comparison
6. Innovations in stream processing: event
time, rich streaming window operations,
savepoints, …
• http://data-artisans.com/how-apache-flink-enables-new-streaming-applications-
part-1/
• http://data-artisans.com/how-apache-flink-enables-new-streaming-applications/

12
7. FlinkCEP is the Complex Event Processing library for
Flink. It allows you to easily detect complex event
patterns in a stream of endless data to support better
insight and decision making.
• Introducing Complex Event Processing (CEP) with Apache Flink, Till Rohrmann
April 6, 2016 http://flink.apache.org/news/2016/04/06/cep-monitoring.html
• FlinkCEP - Complex event processing for
Flinkhttps://ci.apache.org/projects/flink/flink-docs-
master/apis/streaming/libs/cep.html
8. Run Legacy Big Data applications on Flink: Preserve
your investment in your legacy Big Data applications by
currently running your legacy code on Flink’s powerful
engine using Hadoop and Storm compatibility layers,
Cascading adapter and probably a Spark adapter in the
future.

13
Run your legacy Big Data applications on Flink
Flink’s MapReduce compatibility layer allows to run legacy Hadoop
MapReduce jobs, reuse Hadoop input and output formats and reuse
functions like Map and Reduce. https://ci.apache.org/projects/flink/flink-docs-
master/apis/batch/hadoop_compatibility.html
Cascading on Flink allows to port existing Cascading-MapReduce
applications to Apache Flink with virtually no code changes.
Expected advantages are performance boost and less resources
consumption. https://github.com/dataArtisans/cascading-flink/tree/release-0.2
Flink is compatible with Apache Storm interfaces and therefore
allows reusing code that was implemented for Storm: Execute
existing Storm topologies using Flink as the underlying engine.
Reuse legacy application code (bolts and spouts) inside Flink
programs. https://ci.apache.org/projects/flink/flink-docs-
master/apis/streaming/storm_compatibility.html

14
Agenda
Flink?

15
Stonebraker et al. predicted in 2005 that stream
processing is going to become increasingly important
and attributed this to the ‘sensorization of the real
world: everything of material significance on the planet
get ‘sensor-tagged’ and report its state or location in
real time’. Reference: http://cs.brown.edu/~ugur/8rulesSigRec.pdf
I think stream processing is becoming important not only
because of this sensorization of the real world but also
because of the following factors:
1. Data streams
2. Technology
3. Business
4. Customers

16
CustomersData Streams
Technology Business1
2 3
4
Emergence of
Streaming Analytics

17
1 Data Streams
 Real-world data is available as series of events that
are continuously produced by a variety of
applications and disparate systems inside and
outside the enterprise. Examples:
• Sensor networks data
• Web logs
• Database transactions
• System logs
• Tweets and social media data in general
• Click streams
• Mobile apps data

18
2 Technology
Simplified data architecture with Apache Kafka as a
major innovation and backbone of streaming
architectures.
Rapidly maturing open source streaming analytics
tools: Apache Flink, Apache Spark’s Streaming module, Kafka
Streams, Apache Samza, Apache Storm, Apache Nifi…
Cloud services for streaming processing: Google Cloud
Dataflow, Azure Stream Analytics, Amazon Kinesis Streams, IBM
InfoSphere Streams, …
Vendors innovating in this space: Data Artisans,
DataTorrent, Striim, Databricks, MapR, Hortonworks, Confluent,
StreamSets, …
More mobile devices than human beings!

19
3 Business
Challenges:
 Lag between data creation and actionable insights.
 Web and mobile application growth, new types/sources of data.
 Need of organizations to shift from reactive approach to a more of
a proactive approach to interactions with customers, suppliers
and employees.
Opportunities:
Embracing streaming analytics helps organizations with faster
time to insight, competitive advantages and operational efficiency
in a wide range of verticals.
With streaming analytics, new startups are/will be challenging
established companies. Example: Pay-As-You-Go insurance or
Usage-Based Auto Insurance
Speed is said to have become the new currency of business.

20
4 Customers
Customers are becoming more and more demanding
for instant responses in the way they are used to in
social networks: Twitter, Facebook, Linkedin, …
Younger generation who grow up with video gaming
and accustomed to real-time interaction are now
themselves a growing class of customers

21
Agenda
Flink?

22
3. Why Flink is suitable for real-world streaming
analytics?
3.1. Flink’s streaming analytics features
3.2. What are some streaming analytics use
cases suitable for Flink?

23
Apache Flink 1.0, which was released on March 8th
2016, comes with a competitive set of streaming
analytics features, some of which are unique in the
open source domain.
Apache Flink 1.0.1 was released on April 6th 2016.
The combination of these features makes Apache
Flink a unique choice for real-world streaming
analytics.
Let’s discuss some of Apache Flink features for real-
world streaming analytics.

24
1. Pipelined processing engine
2. Stream abstraction: DataStream as in the real-world
3. Performance: Low latency and high throughput
4. Support for rich windowing semantics
5. Support for different notions of time
6. Stateful stream processing
7. Fault tolerance and correctness
8. High Availability
9. Backpressure handling
10. Expressive and easy-to-use APIs in Scala and Java
11. Support for batch
12. Integration with the Hadoop ecosystem

25
1. Pipelined processing engine
 Flink is a pipelined (streaming) engine akin to parallel
database systems, rather than a batch engine as
Spark.
 ‘Flink’s runtime is not designed around the idea that
operators wait for their predecessors to finish before
they start, but they can already consume partially
generated results.’
 ‘This is called pipeline parallelism and means that
several transformations in a Flink program are
actually executed concurrently with data being
passed between them through memory and network
channels.’ http://data-artisans.com/apache-flink-new-kid-on-the-
block/

26
2. Stream abstraction: DataStream as in the real-
world
 Real world data is a series of events that are
continuously produced by a variety of applications and
disparate systems inside and outside the enterprise.
 Flink, as a stream processing system, models streams
as what they are in the real world, a series of events
and use DataStream as an abstraction.
 Spark, as a batch processing system, approximates
these streams as micro-batches and uses DStream as
an abstraction. This adds an artificial latency!

27
Pipelined processing engine enable true low latency
streaming applications with fast results in milliseconds
High throughput: efficiently handle high volume of
streams (millions of events per second)
Tunable latency / throughput tradeoff: Using a tuning
knob to navigate the latency-throughput trade off.
Yahoo! benchmarked Storm, Spark Streaming and Flink
with a production use-case (counting ad impressions
grouped by campaign).
Full Yahoo! Article, benchmark stops at low write
throughput and programs are not fault tolerant.
https://yahooeng.tumblr.com/post/135321837876/benchmarking-streaming-
computation-engines-at

28
Full Data Artisans article, extends the Yahoo!
benchmark to high volumes and uses Flink’s built-in
state http://data-artisans.com/extending-the-yahoo-streaming-benchmark/
Flink outperformed both Spark Streaming and Storm
in this benchmark modeled after a real-world
application:
• Flink achieves throughput of 15 million messages/second on a
10 machines cluster. This is 35x higher throughput compared to
Storm (80x compared to Yahoo’s runs)
• Flink ran with exactly once guarantees, Storm with at least
once.
Ultimately, you need to test the performance of your
own streaming analytics application as it depends on
your own logic and the version of your preferred
stream processing tool!

29
Flink provides rich windowing semantics. A window is
a grouping of events based on some function of time
(all records of the last 5 minutes), count (the last 10
events) or session (all the events of a particular web
user ).
Window types in Flink:
• Tumbling windows ( no overlap)
• Sliding windows (with overlap)
• Session windows ( gap of activity)
• Custom windows (with assigners, triggers and
evictors)

30
In many systems, these windows are hard-coded and
connected with the system’s internal checkpointing
mechanism. Flink is the first open source streaming
engine that completely decouples windowing from
fault tolerance, allowing for richer forms of windows,
such as sessions.
Further reading:
• http://flink.apache.org/news/2015/12/04/Introducing-windows.html
• http://beam.incubator.apache.org/beam/capability/2016/03/17/capability-matrix.html
• https://www.oreilly.com/ideas/the-world-beyond-batch-streaming-101
• https://www.oreilly.com/ideas/the-world-beyond-batch-streaming-102

31
In a streaming program with Flink, for example to define
windows in respect to time, one can refer to different
notions of time:
• Event Time: when an event did happen in the real
world.
• Ingestion time: when data is loaded into Flink, from
Kafka for example.
• Processing Time: when data is processed by Flink
In the real word, streams of events rarely arrive in the
order that they are produced due to distributed sources,
non-synced clocks, network delays… They are said to be
“out of order’ streams.
Flink is the first open source streaming engine that
supports out of order streams and which is able to
consistently process events according to their event
time.

32
http://beam.incubator.apache.org/beam/capability/2016/03/17/capability-matrix.html
https://ci.apache.org/projects/flink/flink-docs-master/concepts/concepts.html#time
https://ci.apache.org/projects/flink/flink-docs-master/apis/streaming/event_time.html
http://data-artisans.com/how-apache-flink-enables-new-streaming-applications-part-1/

33
Many operations in a dataflow simply look at one
individual event at a time, for example an event parser.
Some operations called stateful operations are defined as
the ones where data is needed to be stored at the end of a
window for computations occurring in later windows.
Now, where the state of these stateful operations is
maintained?

34
 The state can be stored in memory in the File System
or in RocksDB which is an embedded key value data
store and not an external database.
 Flink also supports state versioning through
savepoints which are checkpoints of the state of a
running streaming job that can be manually triggered
by the user while the job is running.
 Savepoints enable:
• Code upgrades: both application and framework
• Cluster maintenance and migration
• A/B testing and what-if scenarios
• Testing and debugging.
• Restart a job with adjusted parallelism
Further reading: http://data-artisans.com/how-apache-flink-enables-new-streaming-
applications/
 https://ci.apache.org/projects/flink/flink-docs-master/apis/streaming/savepoints.html

35
How to ensure that the state is correct after failures?
Apache Flink offers a fault tolerance mechanism to
consistently recover the state of data streaming
applications.
This ensures that even in the presence of failures, the
operators do not perform duplicate updates to their
state (exactly once guarantees). This basically means
that the computed results are the same whether there
are failures along the way or not.
There is a switch to downgrade the guarantees to at
least once if the use case tolerates duplicate updates.

36
Further reading:
• High-throughput, low-latency, and exactly-once stream
processing with Apache Flinkhttp://data-artisans.com/high-
throughput-low-latency-and-exactly-once-stream-processing-with-apache-
flink/
• Data Streaming Fault Tolerance document:
http://ci.apache.org/projects/flink/flink-docs-
master/internals/stream_checkpointing.html
• ‘Lightweight Asynchronous Snapshots for Distributed
Dataflows’ http://arxiv.org/pdf/1506.08603v1.pdf June 28, 2015
• Distributed Snapshots: Determining Global States of
Distributed Systems, February 1985, Chandra-Lamport
algorithm http://research.microsoft.com/en-
us/um/people/lamport/pubs/chandy.pdf

37
8. High Availability
In the real world, streaming analytics applications need
to be reliable and capable of running jobs for months
and remain resilient in the event of failures.
The JobManager (Master) is responsible for scheduling
and resource management. If it crashes, no new
programs can be submitted and running program will
fail.
Flink provides a High Availability (HA) mode to recover
from JobManager crash, to eliminate the Single Point
Of Failure (SPOF)
Further reading: JobManager High Availability
https://ci.apache.org/projects/flink/flink-docs-
master/setup/jobmanager_high_availability.html

38
9. Backpressure handling
In the real world, there are situations where a system is
receiving data at a higher rate than it can normally
process. This is called backpressure.
Flink handles backpressure implicitly through its
architecture without user interaction while
backpressure handling in Spark is through manual
configuration: spark.streaming.backpressure.enabled.
Flink provides backpressure monitoring to allow users
to understand bottlenecks in streaming applications.
Further reading:
• How Flink handles backpressure? by Ufuk Celebi, Kostas Tzoumas and
Stephan Ewen, August 31, 2015. http://data-artisans.com/how-flink-handles-
backpressure/

39
 High level, expressive and easy to use DataStream API
with flexible window semantics results in significantly
less custom application logic compared to other open
source stream processing solutions.
 Flink's DataStream API ports many operators from its
DataSet batch processing API such as map, reduce, and
join to the streaming world.
 In addition, it provides stream-specific operations such
as window, split, connect, …
 Its support for user-defined functions eases the
implementation of custom application behavior.
 The DataStream API is available in Scala and Java.

40
case class Word (word: String, frequency: Int)
val env = StreamExecutionEnvironment.getExecutionEnvironment()
val lines: DataStream[String] = env.fromSocketStream(...)
lines.flatMap {line => line.split(" ")
.map(word => Word(word,1))}
.window(Time.of(5,SECONDS)).every(Time.of(1,SECONDS))
.keyBy("word").sum("frequency")
.print()
env.execute()
val env = ExecutionEnvironment.getExecutionEnvironment()
val lines: DataSet[String] = env.readTextFile(...)
lines.flatMap {line => line.split(" ")
.map(word => Word(word,1))}
.groupBy("word").sum("frequency")
.print()
env.execute()
DataSet API (batch): WordCount
DataStream API (streaming): Window WordCount

41
11. Support for batch
 In Flink, batch processing is a special case of stream
processing, as finite data sources are just streams that
happen to end.
 Flink offers a full toolset for batch processing with a
dedicated DataSet API and libraries for machine learning
and graph processing.
 In addition, Flink contains several batch-specific
optimizations such as for scheduling, memory
management, and query optimization.
 Flink out-performs dedicated batch processing engine
such as Spark and Hadoop MapReduce in batch use
cases.

42
12. Integration with the Hadoop ecosystem
POSIX Java/Scala
Collections
POSIX

43
3.2 What are some streaming analytics use cases
suitable for Flink?
1. Financial services
2. Telecommunications
3. Online gaming systems
4. Security & Intelligence
5. Advertisement serving
6. Sensor Networks
7. Social Media
8. Healthcare
9. Oil & Gas
10. Retail & eCommerce
11. Transportation and logistics

44
Agenda
Flink?

45
Flink?
4.1. Flink as an imbedded key/value data store
4.2. Flink as a distributed CEP engine

46
4.1. Flink as an imbedded key/value data store
 The stream processor as a database: a new design pattern for data
streaming applications, using Apache Flink and Apache Kafka:
Building applications directly on top of the stream processor, rather
than on top of key/value databases populated by data streams.
 The stateful operator features in Flink allow a streaming application
to query state in the stream processor instead of a key/value store
often a bottleneck http://data-artisans.com/extending-the-yahoo-streaming-benchmark/

47
“State querying” feature is expected in upcoming Flink 1.1
http://www.slideshare.net/JamieGrier/stateful-stream-processing-at-inmemory-speed/38

48
4.2. Flink as a distributed CEP engine
Flink stream processor as CEP (Complex Event
Processing) engine. Example: an application that
ingests network monitoring events, identifies access
patterns such as intrusion attempts using FlinkCEP, and
analyzes and aggregates identified access patterns.
Upcoming Talk: Streaming analytics and CEP - Two sides of the
same coin’ by Till Rohrmann and Fabian Hueske at the Berlin
Buzzwords on June 05-07 2016.
http://berlinbuzzwords.de/session/streaming-analytics-and-cep-two-sides-same-coin
Further reading:
– Introducing Complex Event Processing (CEP) with Apache Flink,
Till Rohrmann April 6, 2016 http://flink.apache.org/news/2016/04/06/cep-
monitoring.html
– FlinkCEP - Complex event processing for
Flinkhttps://ci.apache.org/projects/flink/flink-docs-master/apis/streaming/libs/cep.html

49
Agenda
Flink?

50
5. Who is using Flink? . Who is using Apache
Flink?
Some companies using Flink for streaming analytics:
[Telecommunications] [Retail] [Financial Services]
Gaming Security
[Gaming] [Security]
Powered by Flink
pagehttps://cwiki.apache.org/confluence/display/FLINK/Powered+by+Flink

51
 has its hack week and the winner, announced
on December 18th 2015, was a Flink based streaming project!
Extending the Yahoo! Streaming Benchmark and Winning Twitter
Hack-Week with Apache Flink. Posted on February 2, 2016 by
Jamie Grier http://data-artisans.com/extending-the-yahoo-streaming-benchmark/
http://www.slideshare.net/JamieGrier/stateful-stream-processing-at-inmemory-speed
 did some benchmarks to compare
performance of one of their use case originally implemented on
Apache Storm against Spark Streaming and Flink. Results posted
on December 18, 2015
• http://yahooeng.tumblr.com/post/135321837876/benchmarking-streaming-computation-engines-
at
• http://data-artisans.com/extending-the-yahoo-streaming-benchmark/
• https://github.com/dataArtisans/yahoo-streaming-benchmark
• http://www.slideshare.net/JamieGrier/extending-the-yahoo-streaming-benchmark

52
Generic Streaming Analytics Architectural pattern:
This is changing with Flink’s alerts, StreamSQL, state
querying, FlinkCEP, …
Event
Producers
Collector
Broker
Processor
Indexer
Visualizer/Search
• Kafka
• RabitMQ
• JMS
• Amazon
Kinesis
• Google Cloud
Pub/Sub
• MapR Streams
• Flink
• Spark
• Storm
• Samza
• Kafka
streams
• ElasticSearch
• Solr
• Cassandra
• HBase
• MapR DB
• MongoDB
• Apache Geode
• Kibana
• Custom
GUI
• Flume
• SpringXD
• Logstash
• Nifi
• Fluentd
• Apps
• Devices
• Sensors

53
Agenda
Flink?

54
 A few resources for you:
• Flink Knowledge Base: One-Stop for everything
related to Apache Flink. By Slim
Baltagihttp://sparkbigdata.com/component/tags/tag/27-flink
• Flink at the Apache Software Foundation: flink.apache.org/
• Free Apache Flink training from data Artisans
http://dataartisans.github.io/flink-training
• Flink Forward Conference, 12-14 September 2016,
Berlin, Germany http://flink-forward.org/ (call for submissions
announced today April 13th , 2016!)
• Free ebook from MapR: Streaming Architecture: New
Designs Using Apache Kafka and MapR Streams
https://www.mapr.com/streaming-architecture-using-apache-kafka-mapr-
streams

55
 A few takeaways:
• Apache Flink unique capabilities enable new and
sophisticated use cases especially for real-world
streaming analytics.
• Customers demand will push major Hadoop distributors
to package Flink and support it.
• What would be the 5G of Big Data Analytics platforms?
Guiding principles would be Unification, Simplification
and Ease of use:
GUI to build batch and streaming applications
Unified API for batch and streaming
Single engine for batch and streaming
Unified storage layer (files, streams, NoSQL)
Unified query engine for SQL, NoSQL and structured
streams

56
Thanks!
To all of you for attending!
Let’s keep in touch!
• sbaltagi@gmail.com
• @SlimBaltagi
• https://www.linkedin.com/in/slimbaltagi
Any questions?

Overview of Apache Fink: the 4 G of Big Data Analytics Frameworks

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Viewers also liked

Viewers also liked (16)

Similar to Overview of Apache Fink: the 4 G of Big Data Analytics Frameworks

Similar to Overview of Apache Fink: the 4 G of Big Data Analytics Frameworks (20)

Recently uploaded

Recently uploaded (20)

Overview of Apache Fink: the 4 G of Big Data Analytics Frameworks