Stream processing with Apache Flink - Maximilian Michels Data Artisans

Stream Processing with 
Apache Flink
Maximilian Michels
Flink PMC member
mxm@apache.org
@stadtlegende

The Agenda
▪ What is Apache Flink?
▪ Streaming 101
▪ The Flink Engine
▪ A Quick Look at the API
2

Apache Flink
▪ A distributed open-source data analysis
framework
▪ True streaming at its core
▪ Streaming & Batch API
3
Historic data
Kafka, RabbitMQ, ...
HDFS, JDBC, ...
Event logs
ETL, Graphs, 
Machine Learning 
Relational, …
Low latency, 
windowing,
aggregations, ...

Organizations at Flink Forward
4

Flink Community
Top 5 Apache Big Data project in the Apache
Software Foundation
500+ messages/month on the mailing list
8400+ commits
1500+ pull requests merged
950+ stars
510+ forks

Use Case: Log File Analysis
▪ Load log files from a distributed file system
▪ Process them, sessionize according to the user id
▪ Write a view to the database or dump more data
for further processing
8
• Process
• Analyze
• Aggregate

Use Case: Tweet Impressions
9
Continuous Stream of Tweets
(each with a timestamp)
▪ How do we measure the importance of Tweets?
• Total number of views
• Views within a time period
▪ We need to process and aggregate Tweets!
Max Marie Jonas Tim are tweeting.

Use Case: Tweet Impressions
10
Max Marie Jonas Tim are tweeting.
Last minute
Last hour
Last day
Impressions
Impression Events Aggregation of Impressions Output
More at: http://data-artisans.com/extending-the-yahoo-streaming-benchmark/

Why Stream Processing?
▪ Most problems have streaming nature
▪ Stream processing gives lower latency
▪ Data volumes more easily tamed
▪ More predictable resource consumption
12
Event stream
batch
(solved)
event
based

Challenges in Streaming
▪ Latency
▪ Throughput
▪ Fault-Tolerance
▪ Correctness
▪ Elements may be out-of-order
▪ Elements may be processed more than
once
13

Windows
▪ A grouping of records according to time,
count, or session, e.g.
• Count: The last 100 records
• Session: All records for user X
• Time: All records of the last 2 minutes
14

Event Time
▪ Processing time: when data is processed
▪ Ingestion time: when data is loaded
▪ Event time: when data is generated
▪ Almost always, the three are different
▪ Event time helps to process out-of-order or
to replay elements as they occurred
15

Event Time & Watermarks
▪ Elements arrives: How do we know what time it
is?
▪ Processing time: take the hardware clock
▪ Event time: Watermarks
▪ Watermarks are timestamps
▪ No elements later than the timestamp are
expected to arrive
16

17
0
0
0 0
Watermark. Event Timewindow operator

17
0
0 0
1

171
0
0 0
1

17
1
0
0 0
1

17
0
0 0
1
2

17
0
0 0
1
2
2

17
0
0 0
1
2
1
2

17
0
0 0
1
1
2

17
0
0 0
1
1
2
1

17
0
0 0
1
1
2
1
1

17
0
0 0
1
1
2
1
2

17
0
0 0
1
1
2
1
2
2

17
0
0 0
1
1
2
1
2
2
2

17
0
0 0
1
1
2
1
2
2
2
2

17
0
0 0
1
1
2
1
2
2 2

18
Tumbling Windows of 4 Seconds
123412
4
59
9 0
20
20
22212326323321
26
35

18
123412
4
59
9
0
20
20
22212326323321
26
35

18
123412
4
59
9
20
20
22212326323321
26
35

0-3
18
123412
4
59
9
20
20
22212326323321
26
35

0-3
18
1
23412
4
59
9
20
20
22212326323321
26
35

0-3
18
12
3412
4
59
9
20
20
22212326323321
26
35

0-3
18
123
412
4
59
9
20
20
22212326323321
26
35

0-3
4-7
18
123
412
4
59
9
20
20
22212326323321
26
35

0-3
4-7
18
123
4
12
4
59
9
20
20
22212326323321
26
35

0-3
4-7
18
123
4
1
2
4
59
9
20
20
22212326323321
26
35

4-7
18
4
4
59
9
20
20
22212326323321
26
35

4-7
18
4
59
9
20
20
22212326323321
26
35

4-7
18
45
9
9
20
20
22212326323321
26
35

8-11
4-7
18
45
9
9
20
20
22212326323321
26
35

8-11
18
9
9
20
20
22212326323321
26
35

8-11
18
9
20
20
22212326323321
26
35

20-23
8-11
18
9
20
20
22212326323321
26
35

20-23
18
20
20
22212326323321
26
35

20-23
18
20
22212326323321
26
35

20-23
18
20222123
26323321
26
35

24-27
20-23
18
20222123
26323321
26
35

24-27
20-23
18
20222123
26
323321
26
35

32-35
24-27
20-23
18
20222123
26
323321
26
35

32-35
24-27
20-23
18
20222123
26
3233
21
26
35

32-35
24-27
18
26
3233
26
35

32-35
24-27
18
26
3233
35

32-35
24-27
18
26
323335

From Program to Execution
case class Path (from: Long, to: Long)
val tc = edges.iterate(10) {
paths: DataSet[Path] =>
val next = paths
.join(edges)
.where("to")
.equalTo("from") {
(path, edge) =>
Path(path.from, edge.to)
}
.union(paths)
.distinct()
next
}
Cost-based
optimizer
Type extraction
stack
Task
scheduling
Recovery
metadata
Pre-flight (Client)
Master
Workers
DataSource
orders.tbl
Filter
Map DataSource
lineitem.tbl
Join
Hybrid Hash
buildHT probe
hash-part [0] hash-part [0]
GroupRed
sort
forward
Program
Dataflow 
Graph
Memory
manager
Out-of-core
algorithms
Batch &
Streaming
State &
Checkpoints
deploy 
operators
track 
intermediate 
results

Flink Applications
21
Streaming
topologies
Heavy
Batch jobs
Machine Learning at scale
Graph processing at scale

E.g.: Non-Native Iterations
22
Step Step Step Step Step
Client
for (int i = 0; i < maxIterations; i++) {
// Execute MapReduce job
}

Iterative Processing in Flink
▪ Built-in iterations and delta iterations
▪ Executes machine learning and graph
algorithms efficiently
23

E.g.: Non-Native Streaming
24
discretize
stream
Job Job Job Job
while (true) {
// get next few records
// issue batch job
}

Pipelining
25
Basic building block to “keep data moving”
• Low latency
• Operators push data
forward
• Data shipping as
buffers, not tuple-
wise
• Natural handling of

Flink Engine
1. Execute everything as streams

Flink Engine
2. Iterative (cyclic) dataflows

Flink Engine
3. Mutable state in operators State +
Computation

Flink Engine
3. Mutable state in operators
4. Operate on managed memory
State +
Computation

Flink Engine
5. Special code paths for batch
State +
Computation

Flink Engine
6. HA mode – no single point of failure
State +
Computation

Flink Engine
6. HA mode – no single point of failure
7. Checkpointing of operator state
State +
Computation

Flink Eco System
Gelly
Table
ML
SAMOA
DataSet (Java/Scala/Python) DataStream
HadoopM/R
Local Cluster Yarn
Dataflow
Dataflow
MRQL
Table
Cascading
Streaming dataflow runtime
Storm
Zeppelin

Flink Eco System
Gelly
Table
ML
SAMOA
DataSet (Java/Scala/Python) DataStream
HadoopM/R
Local Cluster Yarn
Dataflow
Dataflow
MRQL
Table
Cascading
Streaming dataflow runtime
Storm
Zeppelin
HDFS
HBase
Kafka
RabbitMQ
Flume
HCatalog
JDBC

A Quick Look at the DataStream API
28

API Structure
// Create Environment
StreamExecutionEnvironment env =
StreamExecutionEnvironment.getExecutionEnvironment();
// Add Source
DataStream<Type> source = env.addSource(…);
// Perform transformations
DataStream<Type2> trans = source.keyBy(“field”).map(…).timeWindow(...)
// Add Sink
trans.addSink(…);
// Execute!
env.execute();
29

Hourly Impressions
// read from Kafka Tweet Impressions topic 
DataStream<Tweet> tweets = 
env.addSource(new FlinkKafkaConsumer<>(...)); 
// count total number of tweets 
DataStream<Tweet> summaryStream = tweets
.filter(tweet -> tweet.tweetId != null) 
.keyBy(tweet -> tweet.tweetId) 
.window(TumblingTimeWindows.of(Time.hours(1))) 
.sum("impressions"); 
 
// output to Kafka
summaryStream.addSink(
new FlinkKafkaProducer<Tweet>(...));
30
class Tweet { 
String tweetId; 
String userId; 
String text; 
long impressions; 
}

Up-to-date Daily Impressions
// read from Kafka Tweet Impressions topic 
DataStream<Tweet> tweets = 
env.addSource(new FlinkKafkaConsumer<>(...)); 
// count total number of tweets 
DataStream<Tweet> summaryStream = tweets
.filter(tweet -> tweet.tweetId != null) 
.window(SlidingTimeWindows.of(
Time.days(1), Time.minutes(1))) 
.sum("impressions"); 
 
// output to database or Kafka
summaryStream.addSink(
new FlinkKafkaProducer<Tweet>(...));
31
class Tweet { 
String tweetId; 
String userId; 
String text; 
}

Hourly Impression Summary
DataStream<Summary> summaryStream = tweets 
.window(TumblingTimeWindows.of(Time.hours(1))) 
.apply(new WindowFunction<>() { 
public void apply(String tweetId,
TimeWindow window, 
Iterable<Tweet> impressions, 
Collector<Summary> out) { 
long count = 0; Tweet tweet = null; 
for (Tweet val : impressions) { 
tweet = val; count++; 
} 
// output summary 
out.collect(new Summary(tweet, count, 
window.getStart(), 
window.getEnd()));
} 
});
32
class Tweet { 
String tweetId; 
String userId; 
String text; 
}
class Summary { 
Tweet tweet; 
long beginTime; 
long endTime; 
}

Apache Flink
▪ A powerful framework with stream
processor at its core
▪ Features
• True Streaming with great Batch support
• Easy to use APIs, library ecosystem
• Fault-tolerant and Consistent
• Low latency - High throughput
• Growing community

I ♥ , do you?
35
▪ More information on flink.apache.org
▪ Flink Training at data-artisans.com
▪ Subscribe to the mailing lists
▪ Follow @ApacheFlink
▪ Next: 1.0.0 release
▪ Soon: Stream SQL, Mesos, Dynamic scaling

Thank you for your attention!
36

Stream processing with Apache Flink - Maximilian Michels Data Artisans

Recommended

More Related Content

What's hot (20)

Similar to Stream processing with Apache Flink - Maximilian Michels Data Artisans (20)

More from Evention (20)

Recently uploaded (20)

Stream processing with Apache Flink - Maximilian Michels Data Artisans