Ts 4783 1

Design Patterns for
Complex Event
Processing (CEP)
Alexandre Alves
Shailendra Mishra
Oracle Corporation

Agenda
> Building blocks
> Design Patterns
> Conclusion

2

Building blocks
> EVENT
Defined by a schema (i.e. event type)
Tuple of event properties

Event properties
StockEventType
symbol string
lastBid float
lastAsk float

3

Building blocks
> STREAM
Time ordered sequence of events in time

APPEND-only
One cannot remove events, just add them to the sequence

Unbounded
There is no end to the sequence
{event1, event2, event3, event4, …, eventN}

4

Building blocks
> STREAM
Examples:

{{1s, event1}, {2s, event2}, {4s, event3}}
Valid STREAM

{{1s, event1}, {4s, event2}, {2s, event3}}
Not a STREAM, this is a EVENT CLOUD.

5

Building blocks
> RELATION
Bag of events at some instantaneous time T
Allow for INSERT, DELETE, and UPDATE
Example:
At T=1: {{event1}, {event2}, {event3}}
At T=2: {{event1}, {event3}, {event4}}
No changes to event1 and event3
Event2 was deleted
Event4 was inserted

6

Building blocks
> OPERATORS
Transform STREAMS and RELATIONS
Types of operators:
RELATION to RELATION
Well-known from RDBMS (e.g. project, filter, join)
STREAM to RELATION
WINDOW operator bounds STREAMS into RELATIONS (e.g.
NOW)
RELATION TO STREAM
STREAM to STREAM

7

Summary
> CEP is about continuous processing of online
streaming events

> STREAM is a APPEND-only time-ordered
sequence of events

> RELATION is a INSERT/DELETE/UPDATE-able
bag of events at some time t

8

Design Patterns
> Event Filtering
> New event detection
> Event partitioning
> Event enrichment
> Event aggregation
> Event correlation
> Application-time events
> Missing event detection

9

1. Event filtering
> Problem:
Look for specific data on a stream, dropping the
unwanted events.

> Scenario:
Consider STOCKSTREAM a stream of stock events
(symbol, lastBid, lastAsk)

Look for all stocks whose symbol is ‘AAA’

10

1. Event filtering

Time Input Output

1 {“AAA”, 10.0, 10.5} {“AAA”, 10.0, 10.5}

2 {“AAA”, 10.0, 10.5} {“AAA”, 10.0, 10.5}

3 {“BBB”, 11.0, 12.5}

11

1. Event filtering
Specify STREAM source
> Solution:
Specify a WINDOW operator
SELECT *
FROM stockstream [NOW]
WHERE symbol = ‘AAA’

Define predicate for filtering

12

1. Event filtering
> Why do we need the window operator [NOW]?
Filtering works on relations, hence we need to convert the
stock stream into a relation.
Logically this makes sense, as you cannot work on a
unbound set of things, you need to constraint it to a
bounded set.

Note that in some cases short-cuts are allowed; for
example, where [NOW] is assumed if not specified.

13

2. New event detection
> Problem:
Look for specific data on a stream, notify only if it is
new data

> Scenario:
Report only those stocks whose price have
changed since the last event

14


Time Input Output

1 {“AAA”, 10.0, 10.5} {“AAA”, 10.0, 10.5}

2 {“AAA”, 10.0, 10.5}

3 {“BBB”, 11.0, 12.5} {“BBB”, 11.0, 12.5}

15

> SELECT * FROM stockstream [NOW]
Generates RELATION or STREAM?
Generates a RELATION, which includes all events
at time t
What we need are the events that exist at time t,
but do not exist at time (t - 1)
Use RELATION-STREAM operator ISTREAM
(insert stream)
As we are interested on the last event, use the
window operator [ROWS 1] instead of [NOW].

16

> Solution

ISTREAM(
SELECT *
FROM stockstream [ROWS 1]
)

17

Although similar to event at t=2,
3. Event partitioning it is != then event at t=3…

> There is a problem with the current solution…

Time Input Output

1 {“AAA”, 10.0, 10.5} {“AAA”, 10.0, 10.5}

2 {“AAA”, 10.0, 10.5}

3 {“BBB”, 11.0, 12.5} {“BBB”, 11.0, 12.5}

4 {“AAA”, 10.0, 10.5} {“AAA”, 10.0, 10.5}

18

3. Event partitioning
> We need to partition the window by stock symbol

ISTREAM (
SELECT *
FROM stockstream
[PARTITION BY symbol ROWS 1]
)

19

3. Event partitioning

Time Input Output

1 {“AAA”, 10.0, 10.5} {“AAA”, 10.0, 10.5}

2 {“AAA”, 10.0, 10.5}

3 {“BBB”, 11.0, 12.5} {“BBB”, 11.0, 12.5}

4 {“AAA”, 10.0, 10.5}

20

4. Event enrichment
> Problem:
Enrich events from a stream with (somewhat) static data from
a relation.

> Scenario:
Consider STOCKRELATION (relation):
symbol,
company full-name,
headquarters' location

For every stock event from STOCKSTREAM whose lastBid is
greaten than 5.0, find its company’s location.

21

4. Event enrichment Query is trigged by stream,
and not by relation
STOCKRELATION

Symbol Full-name Location
AAA The AAA company San Francisco

BBB The BBB company San Jose

Time Input (STOCKSTREAM) Output (QUERY)
1
{“AAA”, 10.0, 10.5} {“AAA”, “San Francisco”}
2
{“BBB”, 11.0, 12.5} {“BBB”, “San Jose”}
3
{“AAA”, 4.0, 4.5}

22

4. Event enrichment
> Solution:
SELECT event.symbol, location
FROM stockstream [NOW] AS event,
stockrelation AS data
WHERE event.symbol = data.symbol AND
event.lastBid > 5.0
Why do we need to specify a WINDOW operator for
the STREAM and not for the RELATION?
Because a RELATION is instantaneous, it is already
bound to time t.

5. Event aggregation
> Problem:
Aggregate several (simple) events from a stream
into a single new (complex) event summarizing
event properties of the simple events

> Scenario:
Output the average bid and ask price of a stock of
the last 10 seconds.

24

5. Event aggregation The event for AAA is also
included in the result

Time Input Output

1s {“AAA”, 10.0, 12.0} {“AAA”, 10.0, 12.0}

4s {“AAA”, 12.0, 14.0} {“AAA”, 11.0, 13.0}

{“AAA”, 11.0, 13},
9s {“BBB”, 4.0, 5.0}
{“BBB”, 4.0, 5.0}
15s {“BBB”, 8.0, 10.0} {“BBB”, 6.0, 7.5}
> 10s
20s {“BBB”, 8.0, 10.0}

25

5. Event aggregation
> Solution:
SELECT symbol, AVG(lastBid), AVG(lastAsk)
FROM stockstream [RANGE 10 seconds]
GROUP BY symbol

The RANGE WINDOW operator is a STREAM to RELATION
operator
It defines a range of time in which events are kept, events older
than range are removed from window.

As we are interested on the average price per symbol, we
need to use the GROUP BY operator.

6. Event correlation
> Problem:
Correlate events from one or several streams on a common
set of values, that is, a common pattern.

> Scenario:
Consider a BIDSTREAM and ASKSTREAM, respectively
containing bid (symbol, bidPrice, customer) and ask requests
(symbol, askPrice, customer).
Correlate bids with asks that are related to same symbol and
occurring within 10 seconds of each other; anything older is
considered stale.

27


Time Input (BIDS) Input (ASKS) Output

1s {“AAA”, 12.0, cust1} {“BBB”, 9.0, cust2}

{“AAA”, 10.0, cust3, 10.0, cust2}
5s {“AAA”, 10.0, cust3} {“AAA”, 10.0, cust2}
{“AAA”, 12.0, cust1, 10.0, cust2}
> 10s
15s {“BBB”, 10.0, cust4} {“CCC”, 11.0, cust5} {“AAA”, 10.0, cust3, 10.0, cust2}

20s {“BBB”, 10.0, cust6} {“BBB”, 10.0, cust4, 10.0, cust6}

28

> Solution:
SELECT bid.symbol, bidPrice, bid.cust, askPrice, ask.cust,
FROM bidstream [RANGE 10 seconds] AS bid,
askstream [RANGE 10 seconds] AS ask
WHERE bid.symbol = ask.symbol

You can perform regular join operations between streams, including
outer joins.

7. Application time
> Problem:
Need to use application’s view of time, instead of
CPU wall-clock.
This is particularly useful when events originate
from different machines and need some way of
synchronizing.
> Scenario:
Again, consider bid and ask stream…
However, bid and ask requests are time-stamped
on trader at the time that the order is placed.

7. Application time
> Seller places two ask requests respectively at time 8:00:00
and 8:00:12
> Buyer places one bid request at time 8:00:11

7. Application time
> Remember that we want to correlate using a 10 seconds
window, as anything older is stale…
> Hence first event from seller should not be considered,
and we should correlate the ask price of 11.0 with the bid
price of 9.5

7. Application time
> However, the reality is that the first two events could arrive
together in a burst in the exchange, and the third could be
delayed in the cloud…

7. Application time
> In this case, bid price of 9.5 would correlate to ask price of
10.0 and the exchange would loose money as the spread
is lower…

7. Application time
> Solution:
The query does not change…
However STREAMS must be configured to use application
time-stamps based upon some event property, instead of
having events being system-timestamped…

BidEventType
symbol string
customer string
bidPrice float
orderTime dateTime

8. Missing event detection
> Problem:
Alert if an expected event is missing after some
amount of time.

> Scenario:
Consider a SALESTREAM stream, containing
transaction type (customer order or shipment), and
order id.
Alert if an order is not followed by a shipment within
10 seconds.

36

Time Input Output
1s {1, “ORDER”}

5s {2, “ORDER”}

10s {1, “SHIPMENT”} 10s
15s {3, “ORDER”}
15+t
{“DELAYED”, 2}
20s {3, “SHIPMENT”}

37

> Solution:
SELECT "DELAYED" as alertType, orders.orderId,
FROM salestream MATCH_RECOGNIZE (
PARTITION BY orderId

MEASURES
CustOrder.orderId AS orderId

PATTERN (CustOrder NotTheShipment*) DURATION 10 SECONDS

DEFINE
CustOrder AS (type = ‘ORDER’),
NotTheShipment AS ((NOT (eventType = ‘ SHIPMENT’)))

) AS orders

Conclusion
> Online processing, that is, no need to store to disk first
> Declarative approach
> Able to deal with unbounded data-sets (streams)
> Powerful temporal constructs
> Able to find complex relationships using pattern matching
> Leverage Relational algebra from DBMS by means of
converting STREAM-RELATION and RELATION-
STREAM
> Rich library of event processing design patterns

Alexandre Alves
Shailendra Mishra
Oracle Corporation

40

Ts 4783 1

More Related Content

Viewers also liked

Similar to Ts 4783 1

More from Alexandre de Castro Alves

Ts 4783 1