Ahg microsoft stream_insight_queries

A Hitchhiker’s Guide to Microsoft
StreamInsight Queries
Ramkumar (Ram) Krishnan, Jonathan Goldstein, Alex Raizman
Summary:This paper is a developer’s introduction to the Microsoft StreamInsight Queries. The paper
has two goals: (1) To help you think through stream processing in simple, layered levels of
understanding, complementing the product documentation. (2) To reinforce this learning through
examples of various use cases, so that you can design the query plan for a particular problem and
compose the LINQ query. This ability for top-down composition of a LINQ query, combined with
bottom-up understanding of the query model, will help you build rich and powerful streaming
applications. The more advanced sections of this paper provide an overview of a query plan, the
operators that constitute the query, and, where appropriate, the foundational query algebra itself. The
paper is not a feature reference or a substitute for MSDN documentation.
Category: Step-by-Step
Applies to:StreamInsight
Source: MSDN Library (link to source content, linkto code)
E-book publication date: November 2012 (Document Version: Jun 12 2012)

Copyright © 2012 by Microsoft Corporation
All rights reserved. No part of the contents of this book may be reproduced or transmitted in any form or by any means without the written
permission of the publisher.
Microsoft and the trademarks listed at http://www.microsoft.com/about/legal/en/us/IntellectualProperty/Trademarks/EN-US.aspx are
trademarks of the Microsoft group of companies. All other marks are property of their respective owners.
The example companies, organizations, products, domain names, email addresses, logos, people, places, and events depicted herein are
fictitious. No association with any real company, organization, product, domain name, email address, logo, person, place, or event is
intended or should be inferred.
This book expresses the author’s views and opinions. The information contained in this book is provided without any express, statutory,
or implied warranties. Neither the authors, Microsoft Corporation, nor its resellers, or distributors will be held liable for any damages
caused or alleged to be caused either directly or indirectly by this book.

Contents
1.Introduction...........................................................................................................................................................................................................................................................4
2.Tutorial.....................................................................................................................................................................................................................................................................5
2.1 “Hello, Toll!” – Programming a StreamInsight application........................................................................................................................................................5
2.2Query Development Process...................................................................................................................................................................................................................7
2.3Understanding the Query Output.......................................................................................................................................................................................................12
3.More Windows – Hop, Skip and Jump.....................................................................................................................................................................................................14
4.Partitioning Event Streams for Scale out computations...................................................................................................................................................................17
4.1Snapshot Window and AlterEventDuration to form Sliding Windows................................................................................................................................20
5.Multicast................................................................................................................................................................................................................................................................24
6.Join..........................................................................................................................................................................................................................................................................24
6.1Alerting and Anomalies – Left Anti Join (LAJ)................................................................................................................................................................................27
7.Let’s Play Ball – Heavy Hitters......................................................................................................................................................................................................................30
8. Filter and Project..............................................................................................................................................................................................................................................31
9.Operators as Legos – building up Outer Join........................................................................................................................................................................................33
10.Union....................................................................................................................................................................................................................................................................38
11.Extending the Query with Smarts............................................................................................................................................................................................................38
11.1User Defined Function...........................................................................................................................................................................................................................38
11.2User Defined Aggregates.....................................................................................................................................................................................................................39
11.3User Defined Operator..........................................................................................................................................................................................................................41
12.Notes on Events and Event Modeling....................................................................................................................................................................................................44
12.1Event Shapes..............................................................................................................................................................................................................................................44
12.2 Event Kinds.......................................................................................................................................................................................................................................46
12.3Event “CRUD”.............................................................................................................................................................................................................................................47
13.Step 5 – Time and Commitment..............................................................................................................................................................................................................48
13.1Why does order matter?.......................................................................................................................................................................................................................48
13.2 Current Time Increment (CTI) Events.....................................................................................................................................................................................51
13.3AdvanceTime Settings...........................................................................................................................................................................................................................52
14.Conclusion and Future Work.....................................................................................................................................................................................................................52
15.About the Sample Code...............................................................................................................................................................................................................................52
Feedback and Further Information................................................................................................................................................................................................................53

1.Introduction
Enterprises have always been under pressure to reduce the lag between data acquisition and acting on the acquired data. Over the past
decade, the expansion of free markets has created a massive user base for consumer and business technology –everyone from a
billionaire in Seattle to a street hawker in Bangalore has a mobile phone. The Web owns our eyeballs – in six short years, online
advertising and its adjoining business models have overtaken traditional media spend. The next generation of productivity and media
products will be a seamless mash-up between a user’s office and personal life with ubiquitous access to any content from any device.
Competitive mandates in manufacturing and worldwide initiatives like Green IT have led to massive volumes of sensor-driven, machine-
born data. Utility computing from platforms like Windows Azure aim to eliminate technology barriers of scale and economics. These
irreversible trends portend a scenario where the success of an enterprise depends on its capability to efficiently respond to every
stimulus from its customers and partners. Such a competitive enterprise that seeks a 24x7, 360
o
view of its opportunities will face a
torrent of high volume, high velocity business and consumer data.
The singular impact that these trends will have on traditional IT applications is to elevate the awareness of time. In this attention-deficit
world described above, every arriving data point is fungible – i.e., it represents a customer touch-point, an abnormal sensor reading, or a
trend point whose business value rapidly diminishes with time. For a growing number of applications, this necessitates a shift in mindset
where we want to respond to an opportunity as soon as it presents itself, based on insight built in an incremental manner over time, with
the response having an optimal, if not always the best, fidelity or returns.In other words, it necessitates a shift towards event-driven
processing, where you deal with each arriving data point as an event – as something that happened at a point, or over a period, in time –
and apply your business logic on this event to respond in a time-sensitive manner to the opportunity that this event represents.
This model suits a growing number of applications with demands of low latency (sub-second response time) and high throughput
(handling 100x K events/sec). This includes financial services (risk analytics, algorithmic trading), industrial automation (process control
and monitoring, historian based process analysis), security (anomaly/fraud detection), web analytics (behavioral targeting, clickstream
analytics, customer relationship management), and business intelligence (reporting, predictive analytics, data cleansing) and more.
StreamInsight is a .NET-based platform for the continuous and incremental processing of unending sequences of such events from
multiple sources with near-zero latency. It is a temporal query processing engine that enables an application paradigm where a standing
(i.e., potentially infinitely running) query processes these moving events over windows of time. Processing can range from simple
aggregations, to correlation of events across multiple streams, to detection of event patterns, to building complex time series and
analytical models over streaming data. The StreamInsight programming model enables you to define these queries in declarative LINQ,
along with the ability to seamlessly integrate the results into your procedural logic written in C#. The goal of this paper is to help you
learn, hands-on, how to write declarative queries in LINQ with a clear understanding of reasoning with windows of time.
In this paper, we provide a hands-on, developer’s introduction to the Microsoft StreamInsight Queries. The paper has two goals:
(1) To help you think through stream processing in simple, layered levels of understanding, complementing product documentation. (2)
To reinforce this learning through examples of various use cases – so that you can design the query plan for a particular problem, and
compose the LINQ query. This ability for top-down composition of a LINQ query, combined with a bottom-up understanding of the
query model, will help you build rich and powerful streaming applications. As you read this paper, please study the actual code examples
provided to you in the form of a Visual Studio 2010 solution (HitchHiker.sln).
StreamInsight is an in-memory event processing engine that ships as part SQL Server 2012 and is available separately in the Microsoft
Download Center.It requires .NET 4.0 Full Edition, a C# 3.0 compiler, and Windows7. The engine leverages CLR for its type system and
.NET for runtime, application packaging, and deployment. The choice of Visual Studio gives you the benefit of productivity features like
IntelliSense. An event flow debugger provides a complete developer experience. You are assumed to have a working knowledge of C#
and LINQ.

Part I Understanding Query Development
2.Tutorial
We will start this section with a tutorial example of an application that shows the basic scaffolding required around a StreamInsight
query.Keeping this application relatively constant, we will progress through query capabilities – from simple to the more complex ones.
To get started, install/enable .NET 4.0 Full Edition, StreamInsight, and Visual Studio 2010 on your Windows 7machine.Download the code
from the same location as this paper, and load up HitchHiker.sln in Visual Studio. Besides the code, keep the documentation on
StreamInsight, C#, and LINQ within reach. Keep a checkered notepad and a pencil handy to draw time plots.
2.1“Hello, Toll!” – Programming a StreamInsight application
A tolling station is a common phenomenon – we encounter them in many expressways, bridges, and tunnels across the world. Each toll
station has multiple toll booths, which may be manual – meaning that you stop to pay the toll to an attendant, or automated – where a
sensor placed on top of the booth scans an RFID card affixed to the windshield of your vehicle as you pass the toll booth. It is easy to
visualize the passage of vehicles through these toll stations as an event stream over which interesting operations can be performed.
In HelloToll.cs, we show an application built around a set of demo queries which we will go over in more details later in this paper.The
“Program plumbing” region demonstrates basic steps needed to run a query in StreamInsight.We will walk through this example step by
step. Some code regions that are not immediately relevant to understand how a query fits into an application are collapsed for brevity
and appear as ... in the example below.
...
staticvoidTumblingCount(Application app)
{
var inputStream = app.GetTollReadings();
var query = from win in inputStream.TumblingWindow(TimeSpan.FromMinutes(3))
select win.Count();❸
app.DisplayIntervalResults(query);
}
...
#region Program plumbing
...
staticIQStreamable<TollReading> GetTollReadings(thisApplication app)
{
return app.DefineEnumerable(() =>
// Simulated readings data defined as an array.
// IntervalEvent objects are constructed directly to avoid copying.
new[]
{
IntervalEvent.CreateInsert(

newDateTime(2009, 06, 25, 12, 01, 0),
newDateTime(2009, 06, 25, 12, 03, 0),
newTollReading❶
{
TollId = "1",
LicensePlate = "JNB7001",
State = "NY",
Make = "Honda",
Model = "CRV",
VehicleType = 1,
VehicleWeight = 0,
Toll = 7.0f,
Tag = ""
}),
...
})
// Predefined AdvanceTimeSettings.IncreasingStartTime is used
// to insert CTIs after each event, which occurs later than the previous one.
.ToIntervalStreamable(e => e, AdvanceTimeSettings.IncreasingStartTime);❷
}
staticvoid DisplayPointResults<TPayload>(
thisApplication app,
IQStreamable<TPayload> resultStream)
{
// Define observer that formats arriving events as points to the console window.
var consoleObserver = app.DefineObserver(() =>
Observer.Create<PointEvent<TPayload>>(ConsoleWritePoint));
// Bind resultStream stream to consoleObserver.
var binding = resultStream.Bind(consoleObserver);❹
// Run example query by creating a process from the binding we’ve built above.
using (binding.Run("ExampleProcess"))❺
{
...
}
}
We implement a running StreamInsight query essentially in fivelogical steps. For now, please focus on the code fragments – we will
explain the rationale behind the various choices made in this application as a next step.
❶Each event in StreamInsight consists of two components – the event payload, which reflects the data values carried by the event, and
the event shape – i.e., the temporal nature of the event as it models an actual occurrence. The event shape defines the lifetime of the
event – i.e., the duration for which the values in the payload last along the forward progression of time. In this example, TollReadingis a
C# class that represents the payload for events input into the query. The query outputis of primitive event type long.
❷ Define the input stream of events as a function of event payload and shape. We defined the payload component in the above step.
Next, we model the act of a vehicle crossing the toll booth as an interval of time (via combination of ToIntervalStreamable function
and IntervalEvent type). We complete the input stream definition by wrapping sample data array into enumerable sequence
usingDefineEnumerable function. StreamInsight allows the query engine to interface with practically any input or output device using
IEnumerable, IObservable, or StreamInsight specific units called adapters.
❸ Define the query logic itself, which is expressed as a LINQ statement that declaratively describes the processing on events flowing
into the engine throughinputStream. The query produces an output stream of events with payload of type long. We will discuss the
query logic in the next section. Learning how to write such queries is the central focus of this paper. For now, the key takeaway is that

query logic allows you to define the event processing intent purely based on your knowledge of input and output event payload and
shape definitions – isolated from the physical aspects of event delivery and output.
❹ “Bind” the query logic to the consumer. The consumer in our example just outputs formatted events to the console using the simple
observer consoleObserver defined using DefineObserver method. The query logic result is then bound to it using theBind
method.The output events are interpreted as point events due to usage of PointEvent type in theDefineObserver and
ConsoleWritePoint methods. TheConsoleWritePoint method provides actual implementation for outputting the events to the
console. All that is remaining is to create a process.
❺ Running the query. In order to start evaluation of the query logic we need to create a process by calling theRun method on the
binding.From this point on, unless it is stopped by a subsequent call to Dispose on this process, or abnormally terminated by some
runtime condition, this standing query will continue to run in perpetuity.
This programming model enables a flexible and rapid Dev-Deploy cycle. First, the concept of a query template allows you to declaratively
express your business logic, isolated from concerns of event delivery and output. The combination of LINQ with C# allows you to code
both the declarative and procedural aspects in a single programmatic environment. The Visual Studio IDE gives you the convenience of
IntelliSense for code completion, code analysis, refactoring and other features for rapid development. All the steps leading up to query
start are resolved at compile time, minimizing the number of surprises to deal with at runtime.
While this simple example shows you anonymous entities, the full Object Model API allows you to register each discrete entity that is part
of a process in the StreamInsight metadata. Given a shared knowledge of event payload and shape, query and source/sink developers
can work independently to build and ship their modules as .NET assemblies. The product is based on a .NET platform, so deployment
simply entails locating these assemblies with the StreamInsight libraries. An application integrator can write a wrapper program like
HelloToll.cs that connects to a server, binds a given query logic with sources and sinks, and createsa process.
2.2Query Development Process
In this section, we present a five step process to be used as a general guideline to implement your StreamInsight queries. The key goals
of this process are to help you develop the right mindset when approaching a real-world problem, and to illustrate the use of
StreamInsight query features in arriving at the solution. The steps are:
Step 1 – Model the payload and shape of input and output events from the application’s perspective
Step 2 – Understand the required query semantics by building sample input and output event tables
Step 3 – Gather elements of the query logic to develop an event flow graph and compute the output
Step 4 – Compose the query as a streaming transformation of the input to output
Step 5 – Specify the timeliness of query output in consideration of its correctness
We will repeat this process for each query we develop in this example.
Let'stry to give a definition for the first query in the HelloToll.csexample:
[Tumbling Count] Every 3 minutes, count the number of vehicles processed at the toll station since the last result. Report the result at a
point in time, at the end of the 3 minute window.
2.2.1 Step 1 – Model input and output events from the application’s perspective
StreamInsight allows you to model the event shape to reflect happenings in real-life – as required by the application. As we discussed
in❷, we model the input event as an interval of time based on the assumption that many toll stations are manned, and we want to

account for the few seconds to a minute1
2.2.2 Step 2 – Understand required query semantics by building sample input and output event tables
of interaction between an attendant and a motorist. The StartTime and EndTime of each event
mark a vehicle’s arrival at, and departure from, a toll booth respectively.
The output events containing the counts are modeled as Point events,reflecting the fact that we would know the count value only at the
end of each 3 minute interval.
StreamInsight queries function based on a strong foundational query algebra that treats every event as an interval event, irrespective of
the shape defined at the programmatic level. In this step, we tabulate a representative sample of input events as the query engine would
see it in order to understand the query semantics.
A canonical history table (or CHT) is a simple tabulation of a set of sample events, with each event showing the StartTime and the
EndTime of the event, along with the relevant payload fields. Here, given that the output is just a simple count of events (rather than
some aggregate computation based on payload values), we skip showing the payload values. The variety of input interval durations helps
in building an understanding of the query’s semantics. To further simplify our example we will consider only events generated by toll
booth number 1.
Table 1. Input events canonical history table for example [Tumbling Count] for toll booth 1.
Event StartTime EndTime Payload
TollId
Payload
VehicleId
e1 12:01 12:03
e2 12:02 12:03
e3 12:03 12:08
e4 12:07 12:10
e5 12:10 12:14
e6 12:11 12:13
e7 12:20 12:22
e8 12:22 12:25
It is useful to graph CHTs on a time plot. A time plot is a simple X/Y plot where the X axis marks the progression of time, and Y is simply a
space to place the events as they arrive. The markings on the X axis denote minutes past 12:00, e.g., X=4 corresponds to 12:04.
1
For simplicity we constructed a sample dataset to be rounded to minutes, i.e., StartTime and EndTime always have 0
seconds.

Figure 1. Time plot for input events, a tumbling window of size 3, and result counting vehicles being processed
With the input events mapped on the time plot, we will build an output CHT that reflects the query requirements stated in [Tumbling
Count] above. The problem requires a count to be reported once every three minutes. We therefore partition time into 3 minute
intervals as shown on the time plot above. The green lines depict a tumblingwindow of size 3 minutes. Now, the problem statement is a
bit ambiguous. Are we counting the number of vehicles which arrivedduring this period? Or perhaps the number of vehicles which were
being processed at some point during that period? Or is it the number of vehicles which completed processing during that period? Such
subtle ambiguities will be brought to light when we construct example output CHTs.
All intervals and windows in StreamInsight are closed at the left, open on the right – meaning that an event’s value at the start time
of the interval or window is included in every computation, but the value at the end time is excluded. Since the result of a computation
over a window can be known only at the end of the window interval we want the output as of a point in time, showing the events
processed over the window interval. From the query engine’s perspective, a point event is nothing but an interval event with a lifetime of
one chronon (one .NET tick, equivalent to 100 nanoseconds, which we will represent as ε). In other words, assume we had input events as
of a point in time. Then we can represent them as interval events in the CHT with each event having an EndTime that is StartTime +
ε.Results depicted on Figure 1 below the axis are points. We depict ε as a small line smaller than a unit of time represented on the time
plot. The output events are annotated with a set of events over which this output is computed for easier cross referencing.
Let’s revisit the earlier problem definition to remove the ambiguity (note the emphasized part).
[Tumbling Count] Every 3 minutes, report the number of vehicles processed that were being processed at some point during that period
at the toll station since the last result. Report the result at a point in time, at the end of the 3 minute window.
The comments in the table on the right hand side show the events that contributed to the result during the associated window of time.

Table 2. Output CHT for [Tumbling Count]
Event StartTime EndTime Payload Events Comment
Count
o1 12:03 12:03 +ε 2 e1,e2 e3 excluded because window interval is open at the right
o2 12:06 12:06 +ε 1 e3 e1, e2 are excluded, since event interval is open at the right
o3 12:09 12:09 +ε 2 e3, e4
o4 12:12 12:12 +ε 3 e4, e5, e6
o5 12:15 12:15 +ε 2 e5, e6
12:18 12:18+ε Note that no event is produced for this window.
o6 12:21 12:21+ε 1 e7
o7 12:24 12:24+ε 2 e7, e8
o8 12:27 12:27+ε 1 e8
To provide some perspective, if we had decided to output the number of vehicles which arrived during a given 3 minute period, output
events would have the following count values, taking only the event start times into account. The query required to generate this output
will look very different.
Figure 2. Time plot for input, a tumbling window of size 3, and result counting arrival of the vehicles
Note that even though e3’s StartTime aligns with the window’s end time, since the window interval is open-ended, e3 cannot be added
to this window count. Also, result o3 does not include e3, even though e3 overlaps with this window, because this window doesn’t
contain e3’s StartTime.
If our interest was in the count of events that completed processing during a window, the output CHT would have the following counts.

Again, the query that needs to be composed to get this output would be very different. In fact let's consider two variations here.
Figure 3. Time plot for input, a tumbling window of size 3, and two variations of results counting completion of
the vehicle processing
If we are to consider the EndTime of each event as a point and count these points, we end up with the result shown right below the input
events on the time plot, which may seem odd for the problem definition. The vehicle processing represented by e1 and e2 are accounted
for only at time 12:06. The more plausible results can be obtained if EndTimes were to be shifted back by ε. The latter result is shown on
the bottom-most track on the time plot. Notice that the only difference is in the very first result o1. Let’s trace what happens to e1 and
e2 here: If we consider the EndTime as a point representing the event,then we consider the StartTime of these points to be 12:03.As we
have discussed,these will not be accounted for in the first window, but rather they will be picked up in the second window causing the
result to appear at 12:06. On the other hand,if the EndTime is pushed back by ε,these events would be accounted for in the first window.
Shifting event time by ε only affects computation for events that align to the window boundaries.
Now that we have decided how we want the output to look, the next step is …

2.2.3 Step 3 – Gather elements of query logic to compute the output and develop an event flow graph
There are three main elements for this query:
Window – Windows are the mechanism by which events are grouped together into sets. In this case, we
chose a 3 minute tumbling window, which means that our windows consist of consecutive non-overlapping 3
minute periods of time, as described above.
Count() – Computation done within a window – we count the number of events in each window using this
built-in aggregate.
Project – We project the results out to the output stream, in this case, without any further transformations
to the value.
The query graph is a simple sketch of the event flow through these operators. Chart the query graph from top
to bottom to align yourself with the query graphs displayed by the Event Flow Debugger.
2.2.4 Step 4 – Compose the query as a streaming transformation of input into the output
Here is how the query reads: For each tumbling window win of size 3 minutes defined on inputStream, compute the count of events
and project it out as an event with payload of type long.
varquery = from win in inputStream.TumblingWindow(TimeSpan.FromMinutes(3))
select win.Count();
2.2.5 Step 5 – Consider the timeliness balanced against correctness of query output
This is a critical step which requires specification on the output delivery based on an understanding of the behavior of the input event
stream(s). We will discuss this step in detail under the section “13.Step 5 – Time and Commitment” later in the document. For now, please
overlook the event entries marked “CTI” in the output display in the examples below.
2.3Understanding the Query Output
The input events streamed into the HelloToll query are shown in Figure 5. Each line has StartTime and EndTime values, followed by the
values for the six payload fields. In the example program,input data is emulated as an array. The output events generated from an actual
run are shown in Figure 6. On the output, each INSERT signifies an insertion of an event from the query into the output stream.Let's
analyze the output: The first INSERT event signifies the output for the first tumbling window, reported with the EndTime of the window
(12:03), and reflects a count value of 3, computed based on the inputs:
The second INSERT event with a count of 4 is the output for the second tumbling window, and includes the following events:
INSERT 6/25/2009 12:01:00 06/25/2009 12:03:00 1 JNB 7001 NY Honda CRV 1 0 7.0
INSERT 6/25/2009 12:02:00 06/25/2009 12:03:00 1 YXZ 1001 NY Toyota Camry 1 0 4.0 123456789
INSERT 6/25/2009 12:02:00 06/25/2009 12:04:00 3 ABC 1004 CT Ford Taurus 1 0 5.0 456789123
INSERT 6/25/2009 12:03:00 06/25/2009 12:07:00 2 XYZ 1003 CT Toyota Corolla 1 0 4.0
INSERT 6/25/2009 12:03:00 06/25/2009 12:08:00 1 BNJ 1007 NY Honda CRV 1 0 5.0 789123456
INSERT 6/25/2009 12:05:00 06/25/2009 12:07:00 2 CDE 1007 NJ Toyota 4x4 1 0 6.0 321987654
Figure 4. Query Graph
for the [Tumbling
Count] query

Figure 5. Input events in HelloToll.cs
Figure 6. Output events from [Tumbling Count] in HelloToll.cs
Event StartTime EndTime
Toll
Id
License
Plate State Make Model
Vehicle
Type
Vehicle
Weight Toll Tag
INSERT 6/25/2009 12:01:00 06/25/2009 12:03:00 1 JNB 7001 NY Honda CRV 1 0 7.0
INSERT 6/25/2009 12:02:00 06/25/2009 12:03:00 1 YXZ 1001 NY Toyota Camry 1 0 4.0 123456789
INSERT 6/25/2009 12:03:00 06/25/2009 12:07:00 2 XYZ 1003 CT Toyota Corolla 1 0 4.0
INSERT 6/25/2009 12:03:00 06/25/2009 12:08:00 1 BNJ 1007 NY Honda CRV 1 0 5.0 789123456
INSERT 6/25/2009 12:05:00 06/25/2009 12:07:00 2 CDE 1007 NJ Toyota 4x4 1 0 6.0 321987654
INSERT 6/25/2009 12:06:00 06/25/2009 12:09:00 2 BAC 1005 NY Toyota Camry 1 0 5.5 567891234
INSERT 6/25/2009 12:07:00 06/25/2009 12:10:00 1 ZYX 1002 NY Honda Accord 1 0 6.0 234567891
INSERT 6/25/2009 12:07:00 06/25/2009 12:10:00 2 ZXY 1001 PA Toyota Camry 1 0 4.0 987654321
INSERT 6/25/2009 12:08:00 06/25/2009 12:10:00 3 CBA 1008 PA Ford Mustang 1 0 4.5 891234567
INSERT 6/25/2009 12:09:00 06/25/2009 12:11:00 2 DCB 1004 NY Volvo S80 1 0 5.5 654321987
INSERT 6/25/2009 12:09:00 06/25/2009 12:16:00 2 CDB 1003 PA Volvo C30 1 0 5.0 765432198
INSERT 6/25/2009 12:09:00 06/25/2009 12:10:00 3 YZX 1009 NY Volvo V70 1 0 4.5 912345678
INSERT 6/25/2009 12:10:00 06/25/2009 12:12:00 3 BCD 1002 NY Toyota Rav4 1 0 5.5 876543219
INSERT 6/25/2009 12:10:00 06/25/2009 12:14:00 1 CBD 1005 NY Toyota Camry 1 0 4.0 543219876
INSERT 6/25/2009 12:11:00 06/25/2009 12:13:00 1 NJB 1006 CT Ford Focus 1 0 4.5 678912345
INSERT 6/25/2009 12:12:00 06/25/2009 12:15:00 3 PAC 1209 NJ Chevy Malibu 1 0 6.0 219876543
INSERT 6/25/2009 12:15:00 06/25/2009 12:22:00 2 BAC 1005 PA Peterbilt 389 2 2.675 15.5 567891234
INSERT 6/25/2009 12:15:00 06/25/2009 12:18:00 3 EDC 3109 NJ Ford Focus 1 0 4.0 198765432
INSERT 6/25/2009 12:18:00 06/25/2009 12:20:00 2 DEC 1008 NY Toyota Corolla 1 0 4.0
INSERT 6/25/2009 12:20:00 06/25/2009 12:22:00 1 DBC 1006 NY Honda Civic 1 0 5.0 432198765
INSERT 6/25/2009 12:20:00 06/25/2009 12:23:00 2 APC 2019 NJ Honda Civic 1 0 4.0 345678912
INSERT 6/25/2009 12:22:00 06/25/2009 12:25:00 1 EDC 1019 NJ Honda Accord 1 0 4.0

Summary: In this section, we covered the following concepts:
1. The anatomy of a StreamInsight application and simple query.
2. A five step process for query development.
3. The use of canonical history tables to visualize the input and output, clarify a problem statement, and write the
correct query.
4. An application can model an event as an interval or point event.
5. The query engine internally treats every event as an interval event with a start and end time.
6. Understanding the output of a simple query.
3.More Windows – Hop, Skip and Jump
A tumbling window has the advantage of simplicity – it splits time into buckets of equal width over which
aggregates can be easily computed, and it is easy to reason about the events “belonging” to a particular
bucket. But it is good to be aware of some of the limitations of this windowing construct.
There is an inherent latency in reporting of the results, equal to the size of the window. In the above example,
assume that a burst of 50 vehicles arrives at the tolls at 12:09. This indicator of impending congestion will be
reported at time 12:12 – three minutes later.
To track the movement of vehicles at a finer granularity of time, we can use a generalized version of tumbling
window, called Hopping Window. It is a fixed sized window which hops along the timeline with a particular
duration. The tumbling window is a special case of a hopping window where the hop size is equal to the
window size.
Let's recast the first query in the HelloToll.cs example, [Tumbling Count], to use hopping windows instead. We will incorporate some
precision into the problem statement (emphasized) based on what we have learned from Step 2 of the query writing process above.
Rather than show the result every 3 minutes, we will show the result every minute, but still compute the count over a 3 minute window.
So the new problem statement becomes:
[Hopping Count] Report the count of vehicles being processed at some time over a 3 minute window, with the window moving in one
minute hops. Provide the counts as of the last reported result as of a point in time, reflecting the vehicles processed over the last 3
minutes.
Let's walk through the query development process again.
Step 1 Model input and output events from the application’s perspective. No change from [Tumbling Count].
Step 2 Understand the desired query semantics by constructing sample input and output CHTs.
The sample input events are the same as for [Tumbling Count].The time plotfor this input is shown below. The green lines above the time
axis show the progression of a hopping window of size 3 minutes, with a hop size of 1 minute. For convenience, each window is
annotated with the output event name associated with the window. The annotation is placed right above the window start.
For the output, we want the query to determine the output as of a point in time, showing the events processed over the last 3 minutes.
We will specify this interval-to-point transformation in the query logic.
The output events for this sample input are shown in the bottom track of the time plot. Just like in previous example, the output events
are annotated with a set of events over which this output is computed.

Figure 7. Time Plot for Hopping Window of size 3 and hop size 1 [Hopping Count]
The result can be reasoned about in the same terms as we did for the tumbling window. Let's reiterate
a few observations: There are no events produced for windows ending at 12:17 to 12:20, since there
are no input events overlapping those and we do not generate events for 0 counts. Just like in the
case of tumbling windows, e1 and e2 are not included in the window starting at 12:03, since intervals
are open at the end.
Step 3 Gather elements of query logic to compute the output and develop an event flow graph.
The new elements in the query compared to [Tumbling Count] are:
- Window – we choose a hopping window of size 3 minutes, with a hop size of 1 minute – as per
requirement.
- Interval to Point Event Conversion – Recall that the query engine always deals with interval events.
We are required to report the output as of a point in time at the end of a window.
Figure 8. Query Graph for
the [Hopping Count] query

Step 4 Compose the query as a streaming transformation of input to output.
The resulting query is an example of compile time query composition in StreamInsight. First, countStream is composed over
events provided by inputStream – it defines a Count() computation over hopping windows of size of 3 minutes, and a hop size of 1
minute. Next, the result is composed over countStreamby applying ToPointEventStream which converts each event in the stream
from an interval to a point event. This has the effect of truncating the EndTime of the output event to StartTime + ε (one chronon).
varcountStream = from winininputStream.HoppingWindow(TimeSpan.FromMinutes(3), TimeSpan.FromMinutes(1)
select win.Count();
varquery = countStream.ToPointEventStream();
ToPointEventStream is syntactic sugar on top of an operator called AlterEventDuration, which is based on a more general operator
called AlterEventLifeTime. We will see these operators in the upcoming examples.
The output from the complete query in HelloToll.cs is shown in Figure 9.
For now, it is clear that a hopping window finds use in scenarios where you need
to segment time into equal size windows, determine the membership of events in
these windows, and output the result in granules of time (hops) that are less than
the window size. It should also be clear that a tumbling window is a special case
of hopping window – where the window size and the hop size are the same.
Hopping windows have the advantage of predictable memory consumption over
the stream’s lifetime, given that output is released at a fixed frequency. Hopping
windows are the principal mechanism in the StreamInsight release for time-
weighted computations (discussed later). HoppingWindow supports more
overloads to specify alignment of the window to a reference timeline, and input
and output policies for treating event membership in the windows.
On the flip side, hopping windows, like tumbling windows, are chatty in
situations where the window or event intervals are long, and/or when events
arrive in a sporadic manner. This is because it is the window’s properties – rather
than that of the incoming event – that controls the output behavior. For
example, in the output above, {o5, o6} and {o12, o13, o14} output the occurrence
of the same event. A smaller hop size only exacerbates the problem.
More importantly, at high event rates, hopping windows defined for aggregations over groups of events can tend to be expensive in
terms of state maintenance and computation time. In StreamInsight, only the built-in aggregates such as Sum, Count, and Average are
incremental in nature. But any custom aggregation (via a feature called UDA – user defined aggregate)defined by you could be
expensive, by virtue of having to repeat the computation over the same set of events that have already been seen in a previous window.
Future support for incremental aggregates will remove this limitation – but for StreamInsight, it is good to be aware of this constraint.
Figure 9. Output from [Hopping Count]

4.Partitioning Event Streams for Scale out computations
So far, we have based our examples on a single toll booth. A toll station has several toll booths, and most highways or urban areas will
have multiple toll stations. For example, there are 20 such toll stations in a 20 mile radius around New York City that collectively handle
1.4 million crossings a day. (((1.4x10
6
/20)/24)/3600) implies an average of a vehicle per minute at a toll station – which is not much. But
if you assume that most toll stations are integrated with sensor-based traffic monitoring systems on main arteries and roadways, the
volume of events and the kinds of interesting queries that you can ask will demand a scalable querying mechanism. Some examples:
• Which toll stations are congested – now, over the last N minutes/hours? (integration with traffic management)
• Which toll booths are most active in a given station? (this helps in planning road work, shifts, routing)
• What is the inbound (from suburbs to city) versus outbound (city to suburbs) volume of vehicles at a toll station?
• What is the count of transgressions (vehicles with no/expired tags) for a given automated booth at a given station?
These questions require grouping and partitioning of vehicles across different toll stations and possibly other logical hierarchies of
related objects in the application. The computations themselves may be far more sophisticated than the simple count aggregate we have
seen so far. The Group and Apply operator in StreamInsight addresses this need for partitioned, scaled-out processing. We will explain
this feature and introduce a new windowing construct using a simple grouping query:
[Partitioned Hopping window] Find the toll generated from vehicles being processed at each toll boothat some time over a 3 minute
window, with the window advancing in one minute hops. Provide the value as of the last reported result.

Figure 10. Hopping windows with Group and Apply
Step 1 Model input and output events from the
application’s perspective.
We’ll have interval events for input and point events for
the output.
Step 2 Understand the desired query semantics by
constructing sample input and output CHTs.
The time plot for this input is shown in Figure 10. For this
example we will extend the time plot adding events for
toll booth 2, colored red, and add toll amounts, presented
next to the event name in parenthesis. For simplicity we
limit to just showing events from these two booths. The
green lines at the bottom of the time plot show the
progress of a hopping window of size 3 minutes, with a
hop size of 1 minute.

In this query, the output is the sum of the toll generated in each hopping window and is computed on a per-booth basis. As we seen
earlier, hopping windows are very chatty and generate a lot of output, so for brevity, we will not show the complete result here.A partial
result is shown on the bottom track of the time plot in Figure 12. For clarity we colored events according to the booth group from which
they were produced. They are also annotated with the resulting toll sum and the set of events that contributed to it, as we did in
previous examples.
The important takeaways from the time plot are:
• An event output is generated for each hopping window for each group of events (in this case 2 groups).
• The aggregate operation (Sum) is repeated for each hopping window.
• The same set of events repeat themselves in some windows. For example, events o11, o13, and o15
compute the Sum() operation over the same set of two events {e4, e7 }.
Step 3 Gather elements of
query logic to compute the
output and develop an event
flow graph.
• Group and Apply – we need a grouping operator to group the events in the stream based on Booth ID.
• Window – we choose a hopping window of size 3 minutes, with a hop size of 1 minute, as per the
requirements.
• Sum() – we use this aggregation operator to return the toll amount corresponding to a window
In [Partitioned Hopping window], the grouping clause groups events by TollId into groups per toll booth. A hopping window is
defined on each group, and within each hopping window a Sum is computed against the value of the Toll. As you can see, the query
formulation is very intuitive and follows the English description to a tee.
varquery = from e ininputStream
group e by e.TollId into perTollBooth
from winin perTollBooth.HoppingWindow(TimeSpan.FromMinutes(3), TimeSpan.FromMinutes(1))
selectnewToll
{
TollId = perTollBooth.Key,
TollAmount = win.Sum(e => e.Toll)
};
In this example, we use Sum as the aggregate. Built-in aggregates are incremental, so this is not much of a problem. While this may
seem trivial in this example, aggregates in real-life scenarios can be fairly complex. Each computation has to load up the same set of
events in its state. High volume streams in applications like trading platforms can have 10’s of 1000’s, if not millions, of such groups. The
choice of hopping window in such scenarios can become expensive.
Figure 12. Partial output events from [Partitioned Hopping
window] query
Figure 11. Query Graph
for [Partitioned Hopping
window] query

4.1Snapshot Window and AlterEventDuration to form Sliding Windows
A windowing construct whose size and forward progression is truly influenced by the incoming events, rather than a static window
definition, is the Snapshot Window. Rather than partitioning time into fixed-size buckets, a snapshot window segments time according
to the start and end timestamps of all occurring changes. The duration of a snapshot window is defined by a pair of closest event end
points – in terms of the event StartTime and EndTime timestamps – from the same or different events. To clarify, consider two events
with start and end timestamps {A, 12:10, 12:30} and {B, 12:15, 12:25}. This yields three snapshot windows {w1, 12:10, 12:15}, {w2, 12:15,
12:25} and {w3, 12:25, 12:30}. Thus, unlike a hopping window, the time span of a snapshot window will not contain the start time or end
time stamps for any individual event – the events themselves define the window.
To explain this, we will recast [Partitioned Hopping Window] to use the snapshot window instead.
[Partitioned Sliding window] Find the most recent toll generated from vehicles being processed at each station over a one minute
windowreporting the result every time a change occurs in the input.
Step 1 Model input and output events from the application’s perspective. No change from [Partitioned Hopping window].
Step 2 Understand the desired query semantics by constructing sample input and output CHTs
The phrase “every time a change occurs” in the above statement points us towards snapshot, rather than hopping, windows. Along with
this, we are required to report the most recent toll amount collected over a one minute window. In other words, we need to find the
aggregate toll amount over the last snapshot window lasting one minute from any given point in time along the time axis. What we have
just done is to essentially request a Sliding Window - a window that slides along the time axis, but with the output being reported
whenever an event arrives into the system.
To compute the aggregate, you want to pick a point on the time axis - say, “Now” - and look back one minute. But recall that it is the
events that define the snapshot. So how do we simulate the idea of “look for events back one minute”? The answer is simple – for each
event that arrives, you make the events look forward one minute. In other words, you change the duration (i.e., the lifetime) of the event
ahead by one minute.

Figure 13. Group and Apply with Snapshot windows
The time plot for this input is shown in Figure 13, based on the same input considered for hopping windows. We have split the input
stream into two groups based on the TollId as a partitioning key, and they are shown in blue (for TollId 1) and red (for TollId 2) on
separate tracks.As before, windows are depicted as green line segments, but this time we had to draw windows for each group
separately since they constitute a snapshot in time for a particular group and so are different across groups. The dotted blue and red
extensions of the original events show how the event duration (i.e., the lifetime of the event) has been altered or extended by one minute
into the future. The intended results of the query are shown in the time plot in Figure 14. And just like before, result events are color
coded according to the toll booth to which they belong, as well as annotated with the resulting toll sum and set of events contributing to
it.
Some observations:
• The first thing you’ll notice is that the number of output events is reduced by a third compared to hopping windows. The smaller
window size of 1 minute compared to 3 minutes in hopping windows does not matter since this event duration is being uniformly
applied to all events. The output stream is event driven – i.e., generated only at changes in the input. This implies that at any point in
time, you see the most recent aggregate value.

Figure 14. Output events from Group and Apply with snapshot windows
• Each window has at least one new event, rather than a repetition of the same set of events from the past window. This implies that
each aggregate computation retains just enough and the most current state – which is significant when you consider thousands or
millions of groups (we discussed this in the introduction to Group and Apply)

• The output reflects the most current event-driven state of the stream by looking back over a
specified time period. From this perspective, snapshot window combined with AlterEventDuration
is the nearest equivalent to a sliding window in StreamInsight, in that it represents the latest
output of the stream at any point in time. We say “nearest equivalent” because a pure sliding window
returns an output at every tick of advancing time.
• Group and Apply – we need a grouping operator to group the events in the stream based on Booth ID.
• Window – we choose a snapshot window – as per requirement.
• Sum() – to return the summed toll amount within a window
• Alter Event Duration – to extend the EndTime stamp of the event by one minute
The event flow graph showing the query with the above components is shown in Figure 15.
By now, the query should be very intuitive to understand. The input stream is grouped by TollId into per-TollBooth streams. On each of
these streams, every event has its end time advanced by a minute, and then a snapshot window is applied on the stream. Within each
snapshot window, we compute the aggregates Sum() and Count(). The second statement transforms the output events into Point
events.
varquery = from e in inputStream.AlterEventDuration(e =>e.EndTime - e.StartTime +TimeSpan.FromMinutes(1))
group e by e.TollId into perTollBooth
from winin perTollBooth.SnapshotWindow()
selectnewToll
{
TollId = perTollBooth.Key,
TollAmount = win.Sum(e => e.Toll),
VehicleCount = win.Count() // computed as a bonus, not asked in the query
};
One can now easily compute a moving average. First transform the output events into Point events and then,given the sum and the
count, compute the moving average of toll collected through a given tollbooth, as follows:
varpartitionedMovingAverage = from e inquery.ToPointEventStream()
selectnewTollAverage
{
TollId = e.TollId,
AverageToll = e.TollAmount / e.VehicleCount
};
Figure 15. Query graph
for [Partitioned Snapshot
window] query

5.Multicast
A CEP stream is an unidirectional flow of events from input to the output with events participating in incremental
operations as they flow through the query. There are many instances where we need to replicate a stream and use
each replica to perform a different computation or analysis, and then correlate these streams to derive some
insight. StreamInsight provides a multicastoperator to replicate a stream. We will see examples of the multicast
operation in the next section on Joins.
6.Join
StreamInsight is a temporal query processing engine, and one of the key capabilities of the engine is to correlate
event occurrences with respect to time and their payload values. Many of our everyday queries involve a
temporal correlation: “what was the S&P index when this stock went up by 15%?”, “when did this incident
happen (before/after another event)”?, “what caused this to happen?”, “what should happen when events A, B
and C happen in that order (pattern)”? The JOIN operation enables this correlation between event streams in
StreamInsight.
The Join operation in StreamInsight is temporal, and is characterized by the following behavior. A Join between two events from the
respective streams is valid:
(a) for the duration of overlap of lifetimes of the events,
(b) when the payload values of the events satisfy the join predicate.
Let's consider this simple application.
[Inner Join] “Report the output whenever Toll Booth 2 has processed the same number of vehicles as Toll Booth 1, computed over the
last 1 minute, every time a change occurs in either stream.”
Step 1 Model input and output events from the application’s perspective.
We will use the input stream used for [Partitioned Sliding Window].
Step 2 Understand the desired query semantics by constructing sample input and output event CHTs.
Assume that we are able to separate the streams for Toll Booth 1 and 2. The objective is to correlate the first stream against the second,
and generate an output whenever the count of vehicles between the two streams is the same.
The time plot for this input is shown in Figure 16. The phrase “every time a change occurs in either stream” motivates us to define a
snapshot window on the streams. To get count computation over the last 1 minute we will need to extend event end times by 1 minute
just like in previous examples (and also depicted as dotted-line extensions). And finally we would need to filter the results separately for
toll booth 1 and toll booth 2. The resulting stream for toll booth 1 is depicted in blue in the top track, and for toll booth 2 in red in the
bottom track. Each of these tracks shows the snapshot windows in green which are annotated for convenience with the resulting count
corresponding to the window. The Join operation then compares the two streams. The final result is depicted in the middle and
represents overlapping intervals where the count of vehicles for toll booth 1 matches the count for toll booth 2. Again, participating
events are included in the event annotations for convenience.

In [Partitioned Sliding Window], the ability of the Group and Apply operator to partition events based on a key enabled us to group the
events by TollId, and compute the counts for every snapshot window. In this example, we will replicate that stream, and filter out the
events that correspond to TollId 1 and 2 respectively from each replica of the stream. Then we will correlate these two streams using the
join based on the event count. StreamInsight provides a multicastoperator to replicate streams.
We’ll reuse the logic from [Partitioned Sliding Window] to get a stream of per toll booth aggregations. The two streams stream1 and
stream2 are results of filtering just the events for toll booth 1 and 2 respectively. The logic in each of them is applied over the replicas of
the output stream from partitionedSlidingWindowthat are generated as a result of the Multicast operation. We then correlate these
two streams using a Join operation based on the vehicle count from both streams.

Figure 16. Inner Join between two streams (equal vehicle count is the JOIN predicate)
varstream1 = from e inpartitionedSlidingWindowwhere e.TollId == "1"select e;
varstream2 = from e inpartitionedSlidingWindowwhere e.TollId == "2"select e;
varquery = from e1 instream1
join e2 instream2
on e1.VehicleCount equals e2.VehicleCount
selectnewTollCompare
{
TollId_1 = e1.TollId,
VehicleCount = e1.VehicleCount
};

The Multicast operator is implicitly invoked as a consequence of the LINQ statements that define stream1 and stream2. Note that any
number of such replicas can be generated. The JOIN that you see above is a case of inner join.
A cross join (aka Cartesian join) involves combining payload fields of every event in the first stream with payload fields of every event
from the other, joined stream. It has no join condition.
varcrossJoin = from e1 instream1
from e2 in stream2
{
};
Cartesian joins bring up the specter of an explosion of rows in relational systems. In this case, that risk exists only if the events that are
joined also have exactly overlapping lifetimes on an event-by-event basis.
StreamInsight natively supports Cross Join, Inner Join, and Anti Join. We will take some time to discuss how we can realize other join
types from these three primitives.
A theta joinderives from inner join, where the join condition is something other than equality of values in the specific fields. Consider
this query as an example.
[Theta Join] “Report the output whenever Toll Booth 2 has processed lesser number of vehicles than Toll Booth 1, computed over the
last 1 minute, every time a change occurs in either stream”.
varquery = from e1 instream1
from e2 in stream2
where e1.VehicleCount > e2.VehicleCount
{
};
An equi join is a special case of the theta join – where the join condition is an equality of values in specific fields from the payload of
both streams. This is the example we saw above in [Inner Join].
A self-joinderives from inner-join – it is a join of the stream to itself. Combined with multicast, it can be used to correlate different fields
in the payload (which by itself, one might argue, does not require a join) with different temporal constraints being imposed on each
stream. A commonly recurring pattern in streaming queries is to replicate a stream using a multicast, having one of the replicated
streams compute an aggregate or perform another complex operation, and then use the other replica, typically with temporal
transformations, and to combine the two results to achieve the desired application semantics.
6.1Alerting and Anomalies – Left Anti Join (LAJ)
A common need in streaming applications is the ability to detect non-occurrences of events. Relevant to our toll station scenario, note
that several toll stations offer automated tollbooths. In the US, there is a network of highways that provide this service called EZPass.
Commuters who subscribe to this service can drive through the automated tollbooths with lesser delay. A scanner in the automated
tollbooth reads the tag whenever the vehicle passes through one of the toll booths. Given this background, a major pain point for EZPass

is lost revenue due to toll violations on two accounts: (1) toll evasion – where the vehicle that passes through the toll does not have a tag
(2) speed – where the vehicle passes too fast through the toll. We will look at the toll evasion case here.
In general terms, we detect non-occurrence in the following manner:
1. Define a reference stream of events, each characterizing the occurrence of something of interest.
2. Compare the stream that is being observed against this reference stream on a continuous basis.
3. Only when the observed stream has an absence (i.e., non-occurrence) of the reference event, you report this non-occurrence from
the reference stream.
Figure 17. Left Anti Join in action
Consider the query:
[Left Anti Join] “Report toll violators – owners of vehicles that pass through an automated toll booth without a valid EZ-Pass tag read”.
Let's walk through the query building process.
Step 1 Model input and output events from the applications perspective.

One of these will be the reference stream – it indicates the physical passage of a vehicle through the tollbooth – where an event will be
generated whenever a vehicle crosses the tollbooth. The other will be the observed stream – it represents a valid tag read – where an
event will be generated whenever a valid tag read occurs when the vehicle crosses the tollbooth.
The Start time of the point event in the reference stream will be the point in time when the pressure sensor underneath the asphalt at
the entry to the toll booth senses a vehicle. For the sake of simplicity, let's assume that the hardware in the toll station is fast enough
to read the tag on the windshield at the exact point in time when the vehicle itself was sensed. So the start time of the point event in
the observed stream will be the same, provided it was a valid read – if the tag was missing, expired, or is invalid, there would be no
event generated.
Step 2 Understand the desired query semantics by constructing sample input and output CHTs.
The time plot in Figure 17shows a reference stream with events named eRi and indicates the passage of a vehicle through the toll booth
1. The middle track showsan observed stream with events named eOiand indicates valid tag reads in toll booth 1. The bottom-most track
on the time plot shows the output events, each of these events indicates a tag violation.
• The reference stream is the left branch of any join that we propose to use here, the observed stream is the right stream.
• We need a join type that outputs an event from the reference stream whenever an event in the left (reference) stream of the join
does NOT find a matching event in the right (observed) stream. The match is defined as the points in time:
1. when an event does not exist on the observed stream;
2. when the event’s payload does not satisfy the join predicate.
This example demonstrates case 1 above.The lifetime of this output event will be the duration over which the two events do NOT
overlap – which implies, again, a point event. [Left Anti Join] demonstrates this behavior.
// Simulate the reference stream from inputStream itself - convert it to a point event stream
varreferenceStream = from e ininputStream.ToPointEventStream()select e;
// Simulate the tag violations in the observed stream by filtering out specific
// vehicles. Let us filter out all the events in the dataset with a Tag length of 0.
// In a real scenario, these events will not exist at all – this simulation is only
// because we are reusing the same input stream for this example.
// The events that were filtered out should be the ones that show up in the output of Q7
varobservedStream = from e ininputStream.ToPointEventStream()
where 0 != e.Tag.Length
select e;
// Report tag violations
varquery = referenceStream.LeftAntiJoin(observedStream, (left, right) =>true);
Note that there is no requirement that the reference stream and the observed stream consist of events with the same duration for Left
Anti Join to work. For cases where these events have different lifetimes, the output from Left Anti Join will have a lifetime that
corresponds to the non-overlapping areas of the two lifetimes. It is, however, important that you make sure that the output behavior
correctly reflects your application semantics – here is where the above process will be immensely helpful.
Note that there was no payload considered in the determination of the join – this is another axis on which you could pivot the join
behavior. Also note that we did not specify any window of time over which to make this correlation. You can easily compose another

query based on [Left Anti Join] to “find out the number of violations over the past 1 hour” using the count and windowing constructs
we discussed earlier.
Left Anti Join is the third join type natively supported in StreamInsight.
7.Let’s Play Ball – Heavy Hitters
Recall the query [PartitionedSlidingWindow] from Section 4.1.
[Partitioned Sliding Window] Find the most recent toll generated from vehicles being processed at each station over a one minute
windowreporting the result every time a change occurs in the input.
Now, assume that a toll administrator wants to obtain some aggregate statistics on the output from this query. Assume that she is
interested in knowing the top 2 toll amounts every 3 minutes from the output of [PartitionedSlidingWindow], as described in the
following query:
[TopK] “Report the top 2 toll amounts over a 3 minute tumbling window over the results obtained from [Partitioned Sliding Window].”
We will retain the same event models as [PartitionedSlidingWindow].
Step 2 Understand the desired query semantics by constructing sample input and output event tables.
The input to this example is the output from [PartitionedSlidingWindow], shown in Figure 18.
Figure 18. Top-K query with tumbling windows built using output from [Partitioned Sliding Window]

Event presentation follows established convention – events related to toll booth 1 in blue, toll booth 2 in red, and green line segments
show tumbling windows of 3 minutes. We named events with an "r" prefix as they are the results from the [Partitioned Sliding Window]
query.Provided total tolls are in parenthesis.
The expected output is shown in the bottom track of the time plot. We also used convention to put events ranked 1 in the top row and
those ranked 2 in the second row below.
Step 3 Consider the different elements of query logic required to compute the output above.
We need a mechanism to rank in descending order the TollAmount in each of the buckets defined by the tumbling window. The orderby
clause with the descending qualifier in the StreamInsight query achieves this.
We need to take the top 2 of these rank-ordered entries – the Take() operator in LINQ allows us to do this. LINQ also provides us the
rank in the output.
The resulting query is shown below.
varquery = from window inpartitionedSlidingWindow.TumblingWindow(TimeSpan.FromMinutes(3))
fromtopin
(from ein window
orderby e.TollAmount descending
select e
).Take(2,
e =>newTopEvents
{
TollRank = e.Rank,
TollAmount = e.Payload.TollAmount,
TollId = e.Payload.TollId,
VehicleCount = e.Payload.VehicleCount
})
select top;
8.Filter and Project
We have quietly used filter and project in most of our examples, without specifically highlighting them. Let's take a minute to do so.
Simply put, the Filteroperation is the where clause specification in a StreamInsight LINQ query, along with the select clause that
outputs the filtered events. Consider the query [TopK]. Suppose we wanted to see the output only for TollStation 1. The time plot in
Figure 19 shows the effect of this operation on the output of the [TopK] query.

Figure 19. Filter for toll booth 1 applied to the result of [TopK] query
The query becomes:
varquery = from e intopK
where e.TollId == "1"
select e;
The filter clause here is e.TollId == "1". This can be any logical expression allowed by LINQ. UDFs (User Defined Functions, which we
will see shortly) find their common use here.
A Projectoperation “transforms and projects out” the payload fields for the stream output, packaged into a known type (as we have seen
in examples so far) or into an anonymous type (when used for query composition) as part of the select clause. Projection allows you to
transform the payload – in terms of in-place changes to a field, or in adding or removing fields. Similar to the case for filters, there are no
limitations on the kinds of LINQ and C# expressions that can be supported for projection transforms.

9.Operators as Legos – building up Outer Join
So far, we have followed the top-down, four step query writing methodology to arrive at the LINQ query. If you are able to complement
this top-down approach with a bottom-up introspection of the query in Step 3, this will greatly improve your ability to write efficient
queries. In this section, we will use this bottom-up approach to constructing a query to build up an Outer Join operation. First, let's
understand how a StreamInsight query works.
StreamInsight enriches LINQ with temporal query capability through LINQ extension
methods. On the client, the C# compiler runs HelloToll.cs through a StreamInsight
LINQ Provider, which translates a LINQ query into query template, query binder, and
query definition XML documents. The StreamInsight engine understands only this
XML-based representations of the query– it generates the execution plan, loads and
links the required assemblies that constitute a running query, and executes the
query.
In its compiled form, a StreamInsight query is an event flow graph made up of
operators and streams. Each Operatoris a stateful or stateless machine that performs
a specific operation on each event arriving via one or more input streams, generates
new events as a result of this operation, and outputs the event via one or more of its
output streams. Streamsare efficient in-memory couriers of events fromone operator to another.
For any given query, you can view this query plan using the StreamInsight Debugger. In the simplest use case, start a query trace, run the
query, stop the trace, and then view the trace file in the debugger (we will not digress into Debugger setup here – please consult the
documentation).Let's consider the query [TopK] for this exercise.

Figure 20 Query Graph of [TopK] as shown in the Event Flow Debugger
c:Program FilesMicrosoft StreamInsight 2.1Tools>trace start c:q9-trace.etl
The command completed successfully.
Now execute the query from Visual Studio. Once the query completes…

c:Program FilesMicrosoft StreamInsight 2.1Tools>trace stop
The command completed successfully.
Next, start the debugger from the Start menu, and open the file c:q9-trace.etl. The resulting query graph is shown in Figure 11. The plan
shows events flowing from top to bottom through an operator chain IMPORT -> CLEANSEINPUT -> ALTERLIFETIME -> GROUPING
(which contains an AGGREGATE and PROJECT) ->PROJECT -> TOP-K -> PROJECT -> CLEANSE. We have looked at all of the LINQ level
constructs that correspond to these operators, except for CLEANSEINPUT and CLEANSE. Note that there are other powerful uses for the
debugger – we will not digress into those capabilities now.
Now, imagine generating such a plan manually, using just the higher level constructs we have learned so far, to build either other
operators or complete queries. This is the bottom-up approach to query building that we referred to earlier. The repertoire of primitive
operators in StreamInsight is not very large, and you do not have to know all of them. What is important is the process of trying to
compose a query by working bottom up to construct a query using a flow diagram substituting for the actual query plan.
In the earlier section on Joins, we left out one important join type for later:Outer Join. In this section, we will build this operator bottom
up by combining other primitive operators such as inner join, multicast, and project - we can put them together like so many Lego pieces
to compose the outer join query.
An Outer join derives from inner join, and has the following characteristics.
• The left outer join results in an output event that contains all the fields of the payload from the left stream, with matching values
from the fields in the right stream, and NULL values for the fields where no matches are found.
• The right outer join contains all fields of the payload from the right stream, and matching values from the left stream, and NULL if
no matches are found.
• The full outer join results in events that contain all the fields from payloads of both the streams, with NULL values for non-matching
fields.
Let's consider a left outer join.
We will use interval events for this example.
Step 2 Understand the desired query semantics by constructing sample input and output event tables.
We will first consider the payload values to illustrate the outer join. From a temporal perspective, the semantics are that of an inner join.
Assume we have two streams with payloads shown side by side below.
Start
Time
End
Time
License
Plate
Make Model Start
Time
End
Time
License Plate Toll TagId
12:01 12:03 XYZ 1003 Toyota Camry 12:03 12:08 BAC 1005 5.50 567891234
12:02 12:04 XYZ 1001 Toyota Camry 12:05 12:07 ABC 1004 5.00 456789123
12:03 12:07 ZYX 1002 Honda Accord
12:03 12:08 BAC 1005 Toyota Camry
12:05 12:07 ABC 1004 Ford Taurus
12:07 12:09 NJB 1006 Ford Taurus

The result from a left outer join of the above two streams on equality of License Plates, and based on an intrinsic comparison of the
timestamps done by the inner join semantics, is shown below.
Start
Time
End
Time
License
Plate
Make Model Toll TagId
12:01 12:03 XYZ 1003 Toyota Camry NULL NULL
12:02 12:04 YXZ 1001 Toyota Camry NULL NULL
12:03 12:07 ZYX 1002 Honda Accord NULL NULL
12:03 12:08 BAC 1005 Toyota Camry 5.00 456789123
12:05 12:07 ABC 1004 Ford Taurus 5.50 567891234
12:07 12:09 NJB 1006 Ford Taurus NULL NULL
Given that this operator is not available out of the box, we will emulate left outer join behavior with the primitives at hand.
• An inner join between the two streams will give us the matching rows from the two streams
• A Left Anti Join between the two streams will give all the rows in left stream that do NOT have matching rows in the right stream.
Since the left stream is the reference stream, this will give us the payload fields (License Plate, Make, Model). Now we need to add
the NULL columns to this payload – this is done very easily using a Project operation on this output.
• Combine the above two streams with a Union operator to get the output we want.
This is shown graphically in Figure 20 below.
Figure 20. Outer Join composed from primitives
We will use the same input stream to generate the left and right streams using suitable projections. See how the query matches the
graph provided above.
// Simulate the left stream input from inputStream

varouterJoin_L = from e ininputStream
selectnew
{
LicensePlate = e.LicensePlate,
Make = e.Make,
Model = e.Model,
};
// Simulate the right stream input from inputStream – eliminate all events with Toyota as the vehicle
// These should be the rows in the outer joined result with NULL values for Toll and LicensePlate
varouterJoin_R = from e in inputStream
where e.Make != "Toyota"
selectnew
{
Toll = e.Toll,
TollId = e.TollId,
};
// Inner join the two simulated input streams
varinnerJoin = from left inouterJoin_L
from right inouterJoin_R
where left.LicensePlate == right.LicensePlate
selectnewTollOuterJoin
{
LicensePlate = left.LicensePlate,
Make = left.Make,
Model = left.Model,
Toll = right.Toll,
TollId = right.TollId,
};
// Left antijoin the two input simulated streams, and add the Project
varleftAntiJoin = outerJoin_L
.LeftAntiJoin(outerJoin_R, left => left.LicensePlate, right => right.LicensePlate)
.Select(left =>newTollOuterJoin
{
LicensePlate = left.LicensePlate,
Make = left.Make,
Model = left.Model,
Toll = null,
TollId = null
};
// Union the two streams to complete a Left Outer Join operation
varquery = innerJoin.Union(leftAntiJoin);
Such graphing of event flows, followed by query composition in LINQ, is the bread and butter of StreamInsight query development.

10.Union
The stream outerJoin is the result of a Union operation. This operator joins two streams which have the same
event payload type (aka schema). The Group and Apply operator has an implicit Group Union operator to
group the results from various apply branches. The Union operation is very useful in such scenarios where the
requirement is just to merge the two streams, without performing any correlation checks like Join.
11.Extending the Query with Smarts
Simple aggregates such as Sum, Count, and Avg are built into StreamInsight extensions for LINQ. If you have the need the compute
more custom aggregates and functions, you have four choices:
• User Defined Functions– useful for per-event computations. The function is executed for each incoming event and an expression is
returned. It can be used wherever expressions can be used.
• User Defined Aggregate– useful for computing aggregates over a set of events. It returns a scalar value from a computation based
on a set of events. Typically, the events that fall within a window of time constitute the set of events over which you are
aggregating.
• User Defined Operator– useful for defining complete event transformers. A UDO processes a set of events and returns another set
of events. In most general terms, this feature can be used to build custom operators that perform unique operations not covered by
the existing set of operators such as join, union, group and apply, project, and so on. The more common use case is to build time
sensitive and time insensitive operations against a set of events.
Let's explore these capabilities with toll station based examples.
11.1User Defined Function
There are roughly 19 million EZ-Pass tags in circulation in the US (source). This surprisingly low number indicates that this technology has
not yet captured the user’s imagination as something that could be a more useful commuting tool than a mere payment device.
Depending on how the backend data store is designed, you can obtain different tag information (expired, reported-stolen, other) from
the account database in anywhere from a few seconds to a minute.
The opportunities for immediate action – for customer service (email or mail notification, traffic updates, auto service reminders and
promos “time for oil change, get 10% off at Firestone”), or prevention of misuse or crime (flagging authorities) – are limitless. What takes
several days using traditional batch processing can be shortened down to seconds or minutes after the occurrence of the event of a
vehicle crossing a toll station. Consider another variant of the toll violation query.
[UDF] For each vehicle that is being processed at an EZ-Pass booth, report the toll reading if the tag does not exist, has expired, or is
reported stolen.
The query for the above problem statement will look like this:
varquery = from e ininputStream
where 0 == e.Tag.Length ||TagInfo.IsLostOrStolen(e.Tag) ||TagInfo.IsExpired(e.Tag)
selectnewTollViolation
{
Make = e.Make,

Model = e.Model,
State = e.State,
Tag = e.Tag,
TollId = e.TollId
};
The two user defined functions IsExpired and IsLostOrStolen are defined below.
// Assume this is the reference database; the user defined function will search against this.
publicstaticTagInfo[] tags = newTagInfo[]
{
newTagInfo {TagId="123456789", RenewalDate = ParseDate("2/20/2009"),
IsReportedLostOrStolen=false, AccountId="NJ100001JET1109"},
newTagInfo {TagId="234567891", RenewalDate =ParseDate("12/6/2008"),
IsReportedLostOrStolen=true, AccountId="NY100002GNT0109"},
newTagInfo {TagId="345678912", RenewalDate =ParseDate("9/1/2008"),
IsReportedLostOrStolen=true, AccountId="CT100003YNK0210"}
};
publicstaticbool IsLostOrStolen(stringtagId)
{
returnTags.Any(tag =>tag.TagId == tagId&&tag.IsReportedLostOrStolen);
}
publicstaticbool IsExpired(string TagId)
{
return Tags.Any(tag => tag.TagId == tagId && tag.RenewalDate.AddYears(1) >DateTime.Now);
}
Note that these routines can be of arbitrary complexity. The motivation behind the simple array of TagInfo objects in this example is to
demonstrate something self-contained within this program. In a production system, this will most likely be replaced by LINQ to SQL
query against a database, or against an in-memory cache like Windows AppFabric.
User defined functions are generally used in the filtering clause of the LINQ query, but they can be applied wherever an expression is
returned, so they can be used in Project operations. For example, you can imagine a function that is used to retrieve the VIN (vehicle
identification number) of a vehicle given its license plates –a sample query might be as shown below:
varquery = from e in inputStream
where 0 == e.Tag.Length || TagInfo.IsLostOrStolen(e.Tag) || TagInfo.IsExpired(e.Tag)
selectnewTollViolation
{
Make = e.Make,
Model = e.Model,
State = e.State,
Tag = e.Tag,
TollId = e.TollId
Vin = GetVinFromLicensePlate(e.LicensePlate) // UDF in a Project operation
};
GetVinFromLicensePlate is a UDF that does referential access against some external store. In this place, any other complex operation
that returns an expression will help.
11.2User Defined Aggregates
Out of the box, StreamInsight supports Min, Max, Avg, Sum, and Count. But there are several other custom aggregates that you may be
interested in computing against an event stream. StreamInsight provides the capability to develop such aggregates via the extensibility
mechanism of user defined aggregates.

In the eastern part of the US, Eastern New York, New Jersey, and Connecticut form what is called the Tri-State area, and southern and
central NJ has a similar affinity with the state of PA. Given the population and geographic density, it is common for people to live in one
state and work in another. Assume that for a given toll station, we want to compute the ratio of out-of-state vehicles to in-state vehicles.
Note that this example can be solved using a single query with simple counts of in-state, out-of-state, and total number of vehicles, and
simply dividing them to obtain the result – it does not require the heavy machinery of a UDA. But the goal in this section is to show the
different pieces of crafting a UDA, so we will consider this simple application logic.
[UDA] Over a 3 minute tumbling window, find the ratio of out-of-state vehicles to in-state vehicles being processed at a toll station.
The query for this problem statement will look like this:
var query = from winininputStream.TumblingWindow(TimeSpan.FromMinutes(3))
selectwin.UserDefinedAggregate<TollReading, OutOfStateVehicleRatio, float>(null);
The OutOfStateVehicleRatiois a User Defined Aggregate that is externalized to the LINQ surface through the
UserDefinedAggregate operator extension. The actual class that contains the UDA implementation looks as follows:
publicclassOutOfStateVehicleRatio : CepAggregate<TollReading, float>
{
publicoverride float GenerateOutput(IEnumerable<TollReading> tollReadings)
{
floattempCount = 0;
float totalCount = 0;
foreach (vartollReadingin tollReadings)
{
totalCount++;
if (tollReading.State != "NY")
{
tempCount++;
}
}
return tempCount / totalCount;
}
}
Some notes about the implementation of OutOfStateVehicleRatio class:
• The input into this aggregate is a set of events with payload of type TollReading. The output of this aggregate is a value of type
float (by virtue of this class inheriting from CepAggregate<TollReading, float> and the IEnumerable<TollReading>).
• Alternatively, since we know that the out-of-state is determined based on the payload field State of type String, we can define a
more restrictive class as CepAggregate<string, float> as shown below (changes from the previous class is highlighted in bold)
publicclassOutOfStateVehicleRatio2 : CepAggregate<string, float>
{
publicoverride float GenerateOutput(IEnumerable<string>stateReadings)
{
float tempCount = 0;
float totalCount = 0;
foreach (var statein stateReadings)
{
totalCount++;
if (state != "NY")
{
tempCount++;

}
}
return tempCount / totalCount;
}
}
Another operator extension,UserDefinedAggregateWithMapping,enables an additional formal parameter of an Expression that maps to
the type of a single field of the payload (this mapping could simply be a reference to a specific field in the payload itself). This will enable
us to invoke the UDA with a lambda expression that maps the event payload to a string – in this case, by simply referencing the State
field in the payload. This causes an IEnumerable collection of just the State values to be sent into the UDA instead of the entire event
payload.
var query = from winininputStream.TumblingWindow(TimeSpan.FromMinutes(3))
selectwin.UserDefinedAggregateWithMapping<TollReading, OutOfStateVehicleRatio2, string, float>
(e => e.State, null)
UDAs are not incremental. For every aggregate computation for a given event input, the full set of events that “belong” to the window
over which the aggregate is computed is considered. However, this provides an acceptable performance for most applications.
11.3User Defined Operator
User Defined Operators are useful for processing a set of events and returning another set of events. This can be used to build custom
operators that perform unique operations not covered by the existing set of operators such as join, union, multicast, and so on. UDOs
are also useful when you need to deal with the complete event structure,i.e., inclusive of the Start and End Time fields, rather than
dealing just with payload fields.
Consider a hypothetical example where the toll station also acts as a weigh station for trucks and other commercial vehicles. If the vehicle
is hauling a load that is not in line with its “class” – in terms of number of axles and such parameters - weigh stations charge an extra toll
for the additional tonnage. Assume that for each truck or commercial vehicle that passes through the EZ-Pass toll booth over a particular
time period, the toll station has to compute this score and output an event that has a different structure – with the vehicle’s tonnage,
license plate, weight compliance score, and other information.
[UDO] Over a one hour tumbling window, report all commercial vehicles with tonnage greater than one ton (2K lbs), along with their
arrival times at the toll, and any charges due to weight violation. Overweight charges during the rush hour (7am to 7pm) are double that
of non-rush hours.
We will use the same input events as discussed in previous examples, augmented with weight-related information for this example. The
output event will be structurally different than the input event.
Step 2 Understand the desired query semantics by constructing sample input and output event CHTs.
This is similar to the tumbling window example we noticed earlier. The emphasis in this example is on the output of an event with
different payload type with different start and end timestamps.
While the windowing constructs are simple, the output event structure is dependent on the timestamp of the event. This requires a
construct called time-sensitiveUser Defined Operator.

varquery = from winininputStream.TumblingWindow(TimeSpan.FromHours(1))
from e in win.UserDefinedOperator(() => new VehicleWeights());
select e;
The UDO VehicleWeights() is defined as follows. In this example, we have made the weight charge a constant value – it is possible to pass
this into the UDO implementation as a configuration parameter (in case the weight charge varies every month or every week depending
on some other external factors). This is only to keep the example simpler to understand.
publicclassVehicleWeights : CepTimeSensitiveOperator<TollReading,VehicleWeightInfo>
{
double weightcharge = 0.5;
publicoverrideIEnumerable<IntervalEvent<VehicleWeightInfo>>
GenerateOutput(IEnumerable<IntervalEvent<TollReading>> events,
WindowDescriptor windowDescriptor)
{
List<IntervalEvent<VehicleWeightInfo>> output = newList<IntervalEvent<VehicleWeightInfo>>();
// Identify any commercial vehicles in this window for the given window duration
foreach (var e inevents.Where(e => e.StartTime.Hour >= 0 && e.Payload.VehicleType == 2))
{
// create an output interval event
IntervalEvent<VehicleWeightInfo> vehicleWeightEvent = CreateIntervalEvent();
// populate the output interval event
vehicleWeightEvent.StartTime = e.StartTime;
vehicleWeightEvent.EndTime = e.EndTime;
vehicleWeightEvent.Payload =newVehicleWeightInfo
{
LicensePlate = e.Payload.LicensePlate,
Weight = e.Payload.VehicleWeight,
// here is the interesting part; note how the output is dependent on
// the start and end timestamps of the input event. The weight charge
// is a function of the rush hour definition, the weigh charge factor
// and the vehicle tonnage itself
WeightCharge =
((e.StartTime.Hour >= 7 && e.StartTime.Hour <= 14) ? 2 : 1)
* weightcharge
* e.Payload.VehicleWeight
};
// output the event via the IEnumerable interface
output.Add(vehicleWeightEvent);
}
return output;
}
}
Some important points to consider from both the User Defined Aggregate and UDO discussions above:
• UDAs and UDOs can be defined to be either time sensitive, or time insensitive. The UDA example is time insensitive, and the UDO is
time sensitive.
• The main difference between the two is that time sensitive UDAs or UDOs expose the complete event to the UDO or UDA
implementation – meaning that you can access the StartTime and EndTime of the event that “belongs” or “falls into” a window of

time, and you can define your output event to be dependent on these timestamps. In other words, the event timestamps and
payload fields of the output event can be based on these input event timestamps. In addition, the time sensitive UDA or UDO can
be designed to provide the window’s start and end time values.
In contrast, a time insensitive UDO or UDA exposes only the payload to the UDO or UDA implementation. The output event from
these extensibility mechanisms can be dependent only on the payload field values of the incoming event.
• For a time sensitive UDO or UDA, an important caveat is that ALL values in the implementation of the UDO or UDA MUST
be deterministic
• UDO or UDA is invoked once per window, IF the events arrive into the UDO/UDA in an ordered fashion. We will defer the discussion
on UDO behavior with out of order events until after the next section.
. In other words, you cannot use constructs like Random(), or DateTime.Now() whose values can change from one
invocation of the UDA or UDO to the next.

Part II Understanding Streaming Behavior of a Query
The discussion in Part 1 centered around query development – the process of designing a query, the various operators that constitute a
query, and the composition of the query itself. We deliberately ignored any discussion on stream behavior – i.e., how an event stream
flowed from its source to the sink in a query. We did not delve into how and when query output is generated, and the factors that impact
the correctness and responsiveness of the query.
This is the final step in our query development process:
Step 5 – Consider the timeliness of query output, balanced against correctness of query output
12.Notes on Events and Event Modeling
12.1Event Shapes
A StreamInsight query can support three event models (aka shape) – Point, Interval, and Edge – based on which you can model your
application events. It is worth mentioning that a Projectoperator can change the shape of the payload only, not the event shape itself.
StreamInsight provides you with the flexibility to model the shape of your events:
• In the input adapter: Using the Adapter API in the StreamInsight SDK, you can implement your input application event as a
PointEvent, IntervalEvent, or an EdgeEvent and enqueue it into the query. The input adapter instance is represented in a logical
query graph as an IMPORT operator.
• In the query itself: Inside the query, it is still a good practice to think of all events as interval events, irrespective of its shape. But
from the standpoint of programmability, once an event crosses this threshold into the query, constructs like AlterEventDuration(),
AlterEventLifetime(), and ToPointStream() enable explicit transformation of the event shape, while operations like JOIN can also
implicitly change the shape of the event being output from them. A shape change from a point to an interval event is reversible, but
changes to other event types may be irreversible. For example, if you transform an interval event to a point event, it is impossible
within the context of the query execution to “go back” and discover the validity interval of the event you just transformed. Of
course, you can record the event flow and inspect what happened using the Event Flow Debugger. We have seen examples of these
constructs earlier.
• In the output adapter: Similar to the input adapter, the output adapter can also model the event along one of the three shapes for
output. An output adapter instance is represented in the query graph as an EXPORT operator. Once the event crosses this
threshold, the output adapter instance influences the shape of the output event via the binding of the query with the instance.

Figure 21. Point Event Example
We will briefly look at the various event models in the context of query development, and to set the stage for discussions around stream
behavior in upcoming sections.
PointEventis used to model application events that typically occur at a point in time. A simple example is a web click.
From a query development standpoint:
• Point events used in conjunction with Hopping/Tumbling windows convey the idea of events “falling into” or “belonging to” a
particular, fixed time slice, as shown in the Figure 21.
• Point events have the lifetime of a single tick. When you want to report the impact of a point event over a particular time period (as
in “web clicks over the past one week, one hour” etc.) every time a change occurs in the stream (i.e., in an event-driven manner, using
snapshot windows) the programming technique is to “stretch” or alter the duration of the point events uniformly by the time
period, and then define the snapshots. You can use the same trick when you have to correlate point events from two streams.
The figure below shows an example of correlating point events from two streams in a data-driven manner looking back over a 2
minute window.
• Conversely, you can use point events to vertically slice an interval or an edge event through a JOIN to generate a point-in-time
result.
IntervalEventis used to model application events that span a duration of time. We have built most of our examples in the previous
chapter with interval events. Some key features from a query development standpoint:
• Events modeled as intervals are a quick means to correlate and detect co-occurrence of events across two different streams. The
overlap of lifetimes is an indicator of this co-occurrence.

• Interval events influence CTI events to exhibit specific temporal behavior (discussed next).
Figure 22. Edge Event
EdgeEventis used to model application events for which the end time is unknown at the start time of the event, but is provided at a later
point in time. The input source must have clear markers against each event indicating if a given event that has to be enqueued is a Start
edge or an End edge. The input adapter can then use this marker to set the EdgeType property in the CreateInsertEvent() method.
The two main caveats in enqueueing edge events are:
• The End edge must have the same start time and the payload field values as the Start edge.
• If an End edge event arrives without a corresponding Start edge event it will be ignored.
These are some of the most common use cases and motivations for using Edge Events.
• Edge events are ideal for modeling “long-running” events, such as the uptime of a process or a machine, given the flexibility of
deciding the termination of the event on demand. Against this “anchoring” event, several small events such as CPU cycles, I/O,
memory usage, and such short term metrics can be correlated and summarized.
• A popular application of edge events is in modeling signals – to implement step functions, discretize analog signals, or model event
patterns with arbitrary event lifetimes.
• Following the same theme of “anchoring” event, edge events are used for modeling anomaly detection applications in conjunction
with Left Anti Join. The edge event can be the anchoring event in the left hand side (or reference) stream, while the event stream
with the “anomaly” such as non-occurrence of the event, gaps/mismatch in event payload values can be the right stream. An output
from the left stream is sent out whenever an event in the left (reference) stream of the join does NOT find a matching event in the
right (observed) stream. The match is defined as the points in time (1) when an event does not exist on the observed stream (2)
when the event’s payload does not satisfy the join predicate.
12.2 Event Kinds
You can enqueue just two kinds of events into a StreamInsight query from the input adapter. You do not explicitly tag these events as
such – they are generated by virtue of your using different methods in the Adapter API.
• INSERT event– This is the event kind that gets enqueued when you choose one of the above event shapes, populate the payload
structure, and use the Enqueue() method in your input adapter program.
• CTI event– At the outset, know this much – if CTI events are not enqueued into a query, the query will NOT generate any
output. This is a special punctuation event that gets enqueued when you use the method EnqueueCtiEvent() in your input adapter
program, or define their enqueueing process declaratively. We will understand the significance and purpose of this event in the next
section.

12.3Event “CRUD”
Of the familiar CRUD (Create, Read, Update, Delete) operations, only Create will resonate with users familiar with persisted storage or
messaging systems – the other operations are distinctly different in a streaming system.
Create: You use CreateInsertEvent() to create an INSERT event of a given shape, populate its application timestamp(s) and payload fields,
and enqueue the event.
Read: Recall from our introduction that the streaming model is very different from the query–response model seen in persisted systems.
There is no equivalent of a “row/tuple store” from which you can selectively read an event based on a tuple identifier. The only way you’d
“read” the exact event that you enqueued into a query is at the output of the query – assuming the query was a straight pass through, as
in:
varquery = from e in inputStream select e;
Update: StreamInsight events are immutable. Once an event is enqueued, there is no means to reach into the system and delete or
modify that specific enqueued event in-place in the stream.
The only event model that supports a semblance of an update is the Edge Event. This is demonstrated in Figure 23 Edge Event, and
shown in this table below. You can enqueue a start edge event with EndTime ∞ and then submit an end edge event with a modified end
time. Then you can submit another start edge in a gapless manner with the new payload value. A sequence of such start and end edge
events mimics the behavior of a step function, or an AD convertor.
EventKind Event Edge StartTime EndTime Payload Comment
INSERT e1 Start 12:01 ∞ 1
INSERT e1 End 12:01 12:03 1 Same start time and payload value as e1
INSERT e2 End 12:03 12:06 2 -- ditto for e2 --
INSERT e3 End 12:06 12:08 1
INSERT e4 End 12:08 12:11 3
INSERT e5 End 12:11 12:13 2
… and so on…
Inside the query processing engine, StreamInsight supports RETRACT and EXPAND events. These events are not exposed to the external
user. The only time you will confront these events is while examining a query plan from the Event Flow Debugger. Event retraction and
expansion, and the related operators responsible for cleaning up these events – CLEANSEINPUT and CLEANSE – is outside the scope of
this document.
Delete: Similar to update, there is no concept of deleting an enqueued event. The closest equivalent to a Delete operation is the
ReleaseEvent() method. You will invoke this in the output adapter after you Dequeue an event and copy the results to the event structure
in your local/managed memory.
This short digression on event models and structures is a prelude to the main topic in this chapter – which is to understand stream
behavior and query advancement. This is also the final step to consider in completing the query development process.

13.Step 5 – Time and Commitment
Until now, our discussion of query semantics has been devoid of any notion of timeliness of output delivery. We have proceeded with the
notion that any input event that enters the system is used in the computation and the correct results are streamed out. However, for any
output to be streamed out of a StreamInsight query, the input stream must contain special events called CTI events. This section explains
the need for these special events in your input stream.
StreamInsight is a temporal query processing engine over complex event streams. It is built for applications where time – as in real clock
time (“now” versus “over the past few seconds” versus “past two hours”) – plays a central role in the computation and comprehension of
results. Each event that is processed by a query represents a payload (i.e., data elements) along with start and end timestamps that
demarcate the lifetime of that event.
StreamInsight expects these start and end timestamps to be assigned by the application – in other words, these are application
timestamps. StreamInsight does NOT implicitly assign any values to these timestamps – the program that enqueues events into the
server sets these timestamp from a source. We will see why this is an important distinction shortly.
So far, all of our discussion of query processing has been at the “logical” level. To define query semantics, we used an input event table,
built the time plot to understand how events interact with each other, and derived the output event table. In all of our examples, the
StartTime of the most recent event input into the system was equal to, or at least one tick greater than, the event that arrived in the
system just before it. The End Time of a particular interval event may have overlapped or exceeded that of its preceding event, which is
not a problem. In short, we assumed that all incoming events arriving into a StreamInsight server have their application StartTimes
ordered in relation to the system clock of the host server – i.e., they were Ordered Events.
Ordered Events are achievable when the StreamInsight server is on the same machine as the event source. But in many real-life
deployments, the server and the event sources/sinks are in disparate machines connected by networks for various reasons. Such multiple
systems contribute to latency in the delivery of events from one system to another. This gives rise to the possibility of events arriving into
a StreamInsight server with their timestamps out of order in relation to the system clock of the host server. StreamInsight queries are
designed from ground-up to enable the user to accommodate such Out Of Order Events.
13.1Why does order matter?
Note that a StreamInsight query may have both stateful (Joins, built-in and user-defined Aggregates, user-defined Operators) and
stateless operators (Project, Group and Apply, Union, Filter) in its queries. The ordering of events can directly impact the results of
stateful operators, as shown in the example below of a Count with tumbling windows.

First, let's understand the time plot better. The X axis shows the progression of application time. The event labels, however, are directly
representative of their relative ordering with respect to theclock on the host server. For simplicity, we have considered point in time events.
The values of these local timestamps do not matter – but their relative ordering matters. The following table explicitly calls out what “out
of order” means – in the second table on the right, e4 arrives with an application timestamp that is out of order relative to the
system/wall clock time. Since e3 is the first incoming event, no order is yet established. But once e3 has happened, then the next event
could be in order or out of order. Note that tumbling windows report results for the next 3 minutes by default.
Event Application
System/WallClock
Time
StartTime
Event Application
System/WallClock
Time
StartTime
e3 12:03:00 10:00:00
e4 12:07:00
10:02:00
e3 12:08:00 10:00:00
e4 12:07:00
10:02:00
Given this understanding, the count values for the five tumbling windows from the two time plots are:
Event StartTime Count Event StartTime Count
o1 12:00 2 o1 12:00 2
o2 12:03 1 12:03 --
o3 12:06 1 o2 12:06 2
o4 12:09 2 o3 12:09 1
o5 12:12 1 o4 12:12 1
If you have agreed with these results, it is because of two implicit assumptions in your understanding of the example:

1. That the count is computed for a window defined by specific start and end times, spanning 3 time units – this is correct, and a
constant.
2. That we are asking the query to report the output once every three minutes aligned with the window boundaries (i.e., we are asking
for the output at each horizontal line in the table).
Assume that our application acts on the output of this query with this logic:
if (count < 2)
{
BuyOrder();
}
else
{
SellOrder();
}
Then, with no out of order events (table on the left), o2 would have triggered a BuyOrder by virtue of e3’s arrival. With e3 arriving out of
order (table on the right), the action would be SellOrder – even if at a different time interval (12:06). If this application were a trading
station, we will discover this mistake in a post-audit of the transaction, and undertake some compensatory action.
Now, assume that we know that the network transmission is unreliable, and there is a possibility of events arriving out of order into the
StreamInsight server. Instead of every 3 minutes, assume that we request that the output be reported once every six minutes. Then the
second table becomes:
Event StartTime Count
o1 12:00
12:03
2
<empty>
o2 o3 12:06
12:09
2
3
o4 12:12 1 <tbd>
Again, each horizontal line in the table signifies the reporting of the output. By doing this, we have accommodated for the late arrival of
event e3, and thereby presented a correct (SellOrder) result even in the face of unordered input. But this correctness of output comes at
a price – we have doubled the latency (or put another way, reduced the “liveliness”) of the query output by a factor of two.
The key takeaways from this discussion are:
1. StreamInsight confronts out of order processing of events intrinsically in its query algebra
2. StreamInsight provides a mechanism for the user to make a conscious compromise between liveliness of the query – at the expense
of possible incorrect output, or absolute correctness – at the expense of slowness in reporting of output.
The mechanism for conveying this compromise via the input event stream to the query is a special punctuation event called Current
Time Increment. This is the topic of our next section.

13.2 Current Time Increment (CTI) Events
A CTI event provides a watermarking functionality in the stream to affect two fundamental attributes of the query:
1. The liveliness of the query – the pace at which the query is able to emit output.
2. The statefulness of the query – the pace at which the query is able to release events that contribute towards state (for joins,
aggregates,etc.).
You can explicitly issue the CTI event using the EnqueueCtiEvent() method – with the timestamp as a parameter.
When the streaming application enqueues a CTI into the input stream, you accomplish two things:
1. It provides a guarantee to the query that, from the point of issuance of the CTI, it will not enqueue any event whose Start Time is
less than the timestamp of the CTI Event (i.e., the query will not see an event that is earlier in application time than the CTI event). If
a situation does arise where such an event is enqueued into the query, the query will return a “CTI Violation” exception.
The figure below depicts a scenario where event e5 arrives with a timestamp of 12:07 after a CTI event that has been issued with
timestamp 12:09, which is a cause for a CTI violation.
IMPORTANT– In your program, note that the exception due to CTI violation will happen only in the scenario where you explicitly
enqueue the CTI Events without using AdvanceTimeSettings() to define your CTI progression. This interface is explained below.
2. It informs the query that any stateful operator that has processed events that arrived preceding the CTI can release the events –
under the following specific conditions:
a. an interval event that preceded this CTI event can be released IFF the CTI's timestamp is greater than the End Time of the
interval event.
b. a point event which preceded this CTI event can release the point event IFF the CTI's timestamp is greater than the Start
Time of the point event.
c. an edge event (Start or End Edge does not matter) which preceded this CTI event can release the edge event IFF the CTI's
timestamp is greater than the Start Time of the edge event.
13.2.1Time and Commitment revisited
Given the above background, it is easy to see how CTI events help in achieving the compromise between blocking (i.e., waiting for
unordered events) for correctness of output versus achieving a lively stream (i.e., not waiting for events, but accepting that some of the
output results may be incorrect). In the above figure, point events e3 and e4 are unordered – if the CTI was issued once every event,

then there would have been an error in the output. But if this error was acceptable, then a CTI after every event (as opposed to once
every 3 minutes) would have resulted in a maximally responsive query.
Important: If you know that the events arrive into a query ordered on time, then it is highly recommended that you issue a CTI
coinciding with the Start Time of every event that is being enqueued.
13.3AdvanceTime Settings
A practical problem with CTIs is that most of us cannot exactly predict when we want to issue a CTI explicitly. There are a couple of ways
a developer can implement the streaming application (or to be more precise, the input adapter) to issue CTIs:
1. Populate the input stream itself with CTI events after one or more INSERT events. This is feasible if the event source is persisted, but
even then, it is a cumbersome process or an overhead to seed the input stream with CTI markers – which the input adapter can then
use as an indicator to call EnqueueCtiEvent().
2. Programmatically enqueue CTIs from the input adapter. You can specify a CTI frequency in terms of one CTI every N events through
the adapter configuration, and the adapter can use this to invoke EnqueueCtiEvent()once every N events. This however makes the
adapter implementation complex – if the EnqueueCti fails because of pushback from the engine (see Adapter documentation), then
you have to keep track of enqueueing the CTI event along with INSERT events at the next opportunity provided by the engine.
So the solution is to automate the enqueue of CTIs in a declarative fashion. StreamInsight provides the mechanism to specify this as an
interface implementation in the adapter factory (see Adapter documentation) or as part of the query binding specification itself. See the
product documentation on AdvanceTimeSettings.
14.Conclusion and Future Work
In this paper, we provided a developer’s introduction to the Microsoft StreamInsight Queries. We discussed a systematic 5 step process
by which a developer can think through the formulation of a problem, and working through the steps to arrive at a StreamInsight query.
We introduced the components of a StreamInsight application through a simple tutorial, and used this tutorial to discuss typical needs in
a streaming application, showing how StreamInsight queries addressed these needs. We covered the windowing constructs and other
language elements that enabled filtering, computations, and correlation over events. We showed examples of both top-down and
bottom-up analysis and understanding of composing queries. We then expanded on the various forms of stream behavior and how
StreamInsight enables the user to arrive at a compromise between latency requirements and the need for correctness of results.
While this document provides a good introductory coverage, it is not complete by any means. We have not covered some useful
constructs in Count Window and Dynamic Query Composition capability. The Join construct is immensely powerful for various forms of
event correlation – the examples provided in this document are simplistic. The query behavior section could be expanded to explain the
temporal algebra in greater detail. Particularly, we have not touched upon the Edge event and signal processing using StreamInsight.
15.About the Sample Code
This document ships with sample code packaged in a file called HitchHiker.zip. This file contains a Visual Studio 2010 solution called
HitchHiker.sln, which contains the main project, including a routine that lets you iterate through the various queries discussed in the
document and watch their execution.

Feedback and Further Information
We hope that this paper, along with the sample code, will give you a quick start on programming your event-driven applications using
StreamInsight. We’d greatly appreciate your feedback on this document. If there are errors or bugs in the document or code, please
excuse us, and let us know – we’ll arrange to address the error immediately.
Please send your questions, suggestions, comments and errata on this document to Alex.Raizman@microsoft.com.
We’d appreciate your feedback on the problems that StreamInsight is currently able to address, and more importantly, on any pain
points that it does NOT address. For general product feedback, we’d encourage you to post your questions and comments at the
StreamInsight Forum: http://social.msdn.microsoft.com/Forums/en-US/streaminsight
For the latest news and updates on the product, please subscribe to our blog at: http://blogs.msdn.com/b/streaminsight/
Microsoft StreamInsight ships as part of Microsoft SQL Server 2012, and it is available as a separate download
here: http://go.microsoft.com/fwlink/?LinkID=253700
Acknowledgments
We thank Torsten Grabs, Galex Yen, Georgi Chkodrov, Balan Sethu Raman, and Anton Kirilov for their diligent review of this document
and for their time and guidance.
Copyright © 2012, Microsoft Corporation

Ahg microsoft stream_insight_queries

Recommended

More Related Content

What's hot (16)

Viewers also liked (20)

Similar to Ahg microsoft stream_insight_queries (20)

More from Steve Xu (18)

Recently uploaded (20)

Ahg microsoft stream_insight_queries