Real-time Stream Processing with Apache Flink

Marton Balassi – data Artisans
Gyula Fora - SICS
Flink committers
mbalassi@apache.org / gyfora@apache.org
Real-time Stream Processing
with Apache Flink

Stream Processing
2
 Data stream: Infinite sequence of data arriving in a continuous fashion.
 Stream processing: Analyzing and acting on real-time streaming data,
using continuous queries

Streaming landscape
3
Apache Storm
•True streaming, low latency - lower throughput
•Low level API (Bolts, Spouts) + Trident
Spark Streaming
•Stream processing on top of batch system, high throughput - higher latency
•Functional API (DStreams), restricted by batch runtime
Apache Samza
•True streaming built on top of Apache Kafka, state is first class citizen
•Slightly different stream notion, low level API
Apache Flink
•True streaming with adjustable latency-throughput trade-off
•Rich functional API exploiting streaming runtime; e.g. rich windowing semantics

Apache Storm
4
 True streaming, low latency - lower throughput
 Low level API (Bolts, Spouts) + Trident
 At-least-once processing guarantees Issues
 Costly fault tolerance
 Serialization
 Low level API

Spark Streaming
5
 Stream processing emulated on a batch system
 High throughput - higher latency
 Functional API (DStreams)
 Exactly-once processing guarantees Issues
 Restricted streaming
semantics
 Windowing
 High latency

Apache Samza
6
 True streaming built on top of Apache Kafka
 Slightly different stream notion, low level API
 At-least-once processing guarantees with state
Issues
 High disk IO
 Low level API

Apache Flink
7
 True streaming with adjustable latency and throughput
 Rich functional API exploiting streaming runtime
 Flexible windowing semantics
 Exactly-once processing guarantees with (small) state
Issues
 Limited state size
 HA issue

What is Flink
9
A "use-case complete" framework to unify
batch and stream processing
Event logs
Historic data
ETL
Relational
Graph analysis
Machine learning
Streaming analysis

Flink
Historic data
Kafka, RabbitMQ, ...
HDFS, JDBC, ...
ETL, Graphs,
Machine Learning
Relational, …
Low latency
windowing,
aggregations, ...
Event logs
Real-time data
streams
What is Flink
An engine that puts equal emphasis
to streaming and batch
10

Flink stack
11
Python
Gelly
Table
FlinkML
SAMOA
Batch Optimizer
DataSet (Java/Scala) DataStream (Java/Scala)Hadoop
M/R
Flink Runtime
Local Remote Yarn Tez Embedded
Dataflow
Dataflow
*current Flink master + few PRs
Streaming Optimizer

Overview of the API
 Data stream sources
• File system
• Message queue connectors
• Arbitrary source functionality
 Stream transformations
• Basic transformations: Map, Reduce, Filter, Aggregations…
• Binary stream transformations: CoMap, CoReduce…
• Windowing semantics: Policy based flexible windowing (Time, Count, Delta…)
• Temporal binary stream operators: Joins, Crosses…
• Native support for iterations
 Data stream outputs
 For the details please refer to the programming guide:
• http://flink.apache.org/docs/latest/streaming_guide.html
13
Reduce
Merge
Filter
Sum
Map
Src
Sink
Src

Use-case: Financial analytics
14
 Reading from multiple inputs
• Merge stock data from various sources
 Window aggregations
• Compute simple statistics over windows of data
 Data driven windows
• Define arbitrary windowing semantics
 Combine with sentiment analysis
• Enrich your analytics with social media feeds (Twitter)
 Streaming joins
• Join multiple data streams
 Detailed explanation and source code on our blog
• http://flink.apache.org/news/2015/02/09/streaming-example.html

Reading from multiple inputs
case class StockPrice(symbol : String, price : Double)
val env = StreamExecutionEnvironment.getExecutionEnvironment
val socketStockStream = env.socketTextStream("localhost", 9999)
.map(x => { val split = x.split(",")
StockPrice(split(0), split(1).toDouble) })
val SPX_Stream = env.addSource(generateStock("SPX")(10) _)
val FTSE_Stream = env.addSource(generateStock("FTSE")(20) _)
val stockStream = socketStockStream.merge(SPX_Stream, FTSE_STREAM) 15
(1)
(2)
(4)
(3)
(1)
(2)
(3)
(4)
"HDP, 23.8"
"HDP, 26.6"
StockPrice(SPX, 2113.9)
StockPrice(FTSE, 6931.7)
StockPrice(HDP, 23.8)

Window aggregations
val windowedStream = stockStream
.window(Time.of(10, SECONDS)).every(Time.of(5, SECONDS))
val lowest = windowedStream.minBy("price")
val maxByStock = windowedStream.groupBy("symbol").maxBy("price")
val rollingMean = windowedStream.groupBy("symbol").mapWindow(mean _)
16
(1)
(2)
(4)
(3)
(1)
(2)
(4)
(3)

Data-driven windows
case class Count(symbol : String, count : Int)
val priceWarnings = stockStream.groupBy("symbol")
.window(Delta.of(0.05, priceChange, defaultPrice))
.mapWindow(sendWarning _)
val warningsPerStock = priceWarnings.map(Count(_, 1)) .groupBy("symbol")
.window(Time.of(30, SECONDS))
.sum("count") 17
(1)
(2) (4)
(3)
(1)
(2)
(4)
(3)
Count(HDP, 1)StockPrice(HDP, 23.8)

Combining with a Twitter stream
val tweetStream = env.addSource(generateTweets _)
val mentionedSymbols = tweetStream.flatMap(tweet => tweet.split(" "))
.map(_.toUpperCase())
.filter(symbols.contains(_))
val tweetsPerStock = mentionedSymbols.map(Count(_, 1)).groupBy("symbol")
.sum("count")
18
"hdp is on the rise!"
"I wish I bought more
YHOO and HDP stocks"
Count(HDP, 2)
Count(YHOO, 1)(1)
(2)
(4)
(3)
(1)
(2)
(4)
(3)

Streaming joins
val tweetsAndWarning = warningsPerStock.join(tweetsPerStock)
.onWindow(30, SECONDS)
.where("symbol")
.equalTo("symbol"){ (c1, c2) => (c1.count, c2.count) }
val rollingCorrelation = tweetsAndWarning
.mapWindow(computeCorrelation _)
19
Count(HDP, 2)
Count(YHOO, 1)
Count(HDP, 1)
(1,2)
(1) (2)
(1)
(2)
0.5

Fault tolerance
 Exactly once semantics
• Asynchronous barrier snapshotting
• Checkpoint barriers streamed from the sources
• Operator state checkpointing + source backup
• Pluggable backend for state management
20
1
1
2 3
JM
SM
State manager
Job manager
Operator
Snapshot barrier
Event channel
Data channel
Checkpoint
JM
SM

Performance
21
 Performance optimizations
• Effective serialization due to strongly typed topologies
• Operator chaining (thread sharing/no serialization)
• Different automatic query optimizations
 Competitive performance
• ~ 1.5m events / sec / core
• As a comparison Storm promises ~ 1m tuples / sec / node

Roadmap
22
 Persistent, high-throughput state backend
 Job manager high availability
 Application libraries
• General statistics over streams
• Pattern matching
• Machine learning pipelines library
• Streaming graph processing library
 Integration with other frameworks
• Zeppelin (Notebook)
• SAMOA (Online ML)

Summary
 Flink is a use-case complete framework to unify batch
and stream processing
 True streaming runtime with high-level APIs
 Flexible, data-driven windowing semantics
 Competitive performance
 We are just getting started!
23

Flink Community
24
0
20
40
60
80
100
120
Jul-09 Nov-10 Apr-12 Aug-13 Dec-14 May-16
Unique git contributors

Real-time Stream Processing with Apache Flink

More Related Content

What's hot (20)

Viewers also liked (16)

Similar to Real-time Stream Processing with Apache Flink (20)

More from DataWorks Summit (20)

Real-time Stream Processing with Apache Flink

Editor's Notes