Complex event processing (CEP) and stream analytics are commonly treated as distinct classes of stream processing applications. While CEP workloads identify patterns from event streams in near real-time, stream analytics queries ingest and aggregate high-volume streams. Both types of use cases have very different requirements which resulted in diverging system designs. CEP systems excel at low-latency processing whereas engines for stream analytics achieve high throughput. Recent advances in open source stream processing yielded systems that can process several millions of events per second at a sub-second latency. One of these systems is Apache Flink and it enables applications that include typical CEP features as well as heavy aggregations. Guided by examples, I will demonstrate how Apache Flink enables the user to process CEP and stream analytics workloads alike. Starting from aggregations over streams, we will next detect temporal patterns in our data triggering alerts and finally aggregate these alerts to gain more insights from our data. As an outlook, I will present Flink's CEP-enriched StreamSQL interface providing a declarative way to specify temporal patterns in your SQL query.

1. Till Rohrmann
trohrmann@apache.org
@stsffap
Unifying Stream SQL and CEP
for Declarative Stream
Processing with Apache Flink

2. 2
Original creators of Apache
Flink®
Providers of the
dA Platform, a supported
Flink distribution

3. Streams are Everywhere
 Most data is continuously produced as stream
 Processing data as it arrives
is becoming very popular
 Many diverse applications
and use cases
3

4. Batch Analytics
4
 The batch approach to data analytics

5. Streaming Analytics
 Online aggregation of streams
• No delay – Continuous results
 Stream analytics subsumes batch analytics
• Batch is a finite stream
 Demanding requirements on stream processor
• High throughput
• Exactly-once semantics & event-time support
• Advanced window support
5

6. Complex Event Processing
 Analyzing a stream of events and drawing conclusions
• Detect patterns and assemble new events
 Applications
• Network intrusion
• Process monitoring
• Algorithmic trading
 Demanding requirements on stream processor
• Low latency!
• Exactly-once semantics & event-time support
6

7. Apache Flink®
 Platform for scalable stream processing
 Meets requirements of CEP and stream analytics
• Low latency and high throughput
• Exactly-once semantics
• Event-time support
• Advanced windowing
 Core DataStream API available for Java & Scala
7

8. Tracking an Order Process
Use Case
8

9. Order Process
9

10. Order Events
 Process is reflected in a stream of order events
 Order(orderId, tStamp, “received”)
 Shipment(orderId, tStamp, “shipped”)
 Delivery(orderId, tStamp, “delivered”)
 orderId: Identifies the order
 tStamp: Time at which the event happened
10

11. Aggregating Massive Streams
Stream Analytics
11

12. Stream Analytics
 Traditional batch analytics
• Repeated queries on finite and changing data sets
• Queries join and aggregate large data sets
 Stream analytics
• “Standing” query produces continuous results
from infinite input stream
• Query computes aggregates on high-volume streams
 How to compute aggregates on infinite streams?
12

13. Compute Aggregates on Streams
 Split infinite stream into finite “windows”
• Split usually by time
 Tumbling windows
• Fixed size & consecutive
 Sliding windows
• Fixed size & may overlap
 Event time mandatory for correct & consistent results!
13

14. Example: Count Orders by Hour
14

15. Example: Count Orders by Hour
15
SELECT
TUMBLE_START(tStamp, INTERVAL ‘1’ HOUR) AS hour,
COUNT(*) AS cnt
FROM events
WHERE
status = ‘received’
GROUP BY
TUMBLE(tStamp, INTERVAL ‘1’ HOUR)

16. Stream SQL Architecture
 Flink features SQL on static
and streaming tables
 Parsing and optimization by
Apache Calcite
 SQL queries are translated
into native Flink programs
16

17. Pattern Matching on Streams
Complex Event Processing
17

18. Real-time Warnings
18

19. CEP to the Rescue
 Define processing and delivery intervals (SLAs)
 ProcessSucc(orderId, tStamp, duration)
 ProcessWarn(orderId, tStamp)
 DeliverySucc(orderId, tStamp, duration)
 DeliveryWarn(orderId, tStamp)
 orderId: Identifies the order
 tStamp: Time when the event happened
 duration: Duration of the processing/delivery
19

20. CEP Example
20

21. Processing: Order  Shipment
21
val processingPattern = Pattern
.begin[Event]("received").subtype(classOf[Order])
.followedBy("shipped").where(_.status == "shipped")
.within(Time.hours(1))
val processingPatternStream = CEP.pattern(
input.keyBy("orderId"),
processingPattern)
val procResult: DataStream[Either[ProcessWarn, ProcessSucc]] =
processingPatternStream.select {
(pP, timestamp) => // Timeout handler
ProcessWarn(pP("received").orderId, timestamp)
} {
fP => // Select function
ProcessSucc(
fP("received").orderId, fP("shipped").tStamp,
fP("shipped").tStamp – fP("received").tStamp)
}

22. … and both at the same time!
Integrated Stream Analytics with CEP
22

23. Count Delayed Shipments
23

24. Compute Avg Processing Time
24

25. CEP + Stream SQL
25
// complex event processing result
val delResult: DataStream[Either[DeliveryWarn, DeliverySucc]] = …
val delWarn: DataStream[DeliveryWarn] = delResult.flatMap(_.left.toOption)
val deliveryWarningTable: Table = delWarn.toTable(tableEnv)
tableEnv.registerTable(”deliveryWarnings”, deliveryWarningTable)
// calculate the delayed deliveries per day
val delayedDeliveriesPerDay = tableEnv.sql(
"""SELECT
| TUMBLE_START(tStamp, INTERVAL ‘1’ DAY) AS day,
| COUNT(*) AS cnt
|FROM deliveryWarnings
|GROUP BY TUMBLE(tStamp, INTERVAL ‘1’ DAY)""".stripMargin)

26. CEP-enriched Stream SQL
26
SELECT
TUMBLE_START(tStamp, INTERVAL '1' DAY) as day,
AVG(duration) as avgDuration
FROM (
// CEP pattern
SELECT duration, tStamp
FROM inputs MATCH_RECOGNIZE (
PARTITION BY orderId ORDER BY tStamp
MEASURES END.tStamp – START.tStamp as duration, END.tStamp as tStamp
PATTERN (START OTHER* END)
INTERVAL '1' HOUR
DEFINE
START AS START.status = ’received’,
END AS END.status = ‘shipped’
)
)
GROUP BY
TUMBLE(tStamp, INTERVAL '1' DAY)

27. Conclusion
 Apache Flink handles CEP and analytical
workloads
 Apache Flink offers intuitive APIs
 New class of applications by CEP and
Stream SQL integration 
27

28. 2
Thank you!
@stsffap
@ApacheFlink
@dataArtisans

29. 29
Stream Processing
and Apache Flink®'s
approach to it
@StephanEwen
Apache Flink PMC
CTO @ data ArtisansFLINKFORWARD IS COMING BACKTO BERLIN
SEPTEMBER11-13, 2017
BERLIN.FLINK-FORWARD.ORG -

30. We are hiring!
data-artisans.com/careers