Would Microsoft Azure Stream Analytics Be a Suitable Foundation

for an Event Processing Network Model?

Arne Koschel

1 a

, Anna Pakosch

1 b

, Christin Schulze

Irina Astrova

, Christian Gerner

and Matthias Tyca

Hochschule Hannover, DataH, University of Applied Sciences and Arts, Hannover, Germany

Department of Software Science, School of IT, Tallinn University of Technology, Tallinn, Estonia

Keywords:

Event Processing Network (EPN), Event Processing Network Model, Azure Stream Analytics.

Abstract:

This article looks at a proposed list of generalized requirements for a uniﬁed modelling of event processing

networks (EPNs) and its application to Microsoft Azure Stream Analytics. It enhances our previous work in

this area, in which we recently analyzed Apache Storm, Amazon Kinesis Data Analytics and earlier also the

EPiA model, the BEMN model, and the RuleCore model. Our proposed EPN requirements look at both: The

logical model of EPNs and the concrete technical implementation of them. Therefore, our article provides

requirements for EPN models based on attributes derived from event processing in general as well as existing

models. Moreover, as its core contribution our article applies those requirements by an in depth analysis of

Microsoft Azure Stream Analytics as a concrete implementation foundation of an EPN model.

1 INTRODUCTION

Intelligent data management and processing has

changed: Collecting large amounts of data from var-

ious sources happens in every company today, of-

ten called ’Big Data’. It is not longer sufﬁcient to

store data in relational DBs, log ﬁles or events sep-

arately. Information from data combined from dif-

ferent sources is important for the competitiveness of

enterprises.

Batch Processing is an established approach for

processing ’Big Data’. At its core data is collected

and processed ’in batches’ (Shaikh, 2019). Therefore,

data is collected for a certain period of time before be-

ing processed. The drawback is that no real-time pro-

cessing is possible. First, data is collected for some

time, before processing takes place.

Recently, (Event) Stream Processing joined the

ﬁeld (Shaikh, 2019). An approach to process data di-

rectly after generation. Through this near real-time

processing, an action can happen immediately after

processing. Enterprises can react faster to changes.

For the implementation of Stream Processing a

modeling technique called (complex) Event process-

ing network (EPN) found its way into practice. This

https://orcid.org/0000-0001-5695-2893

https://orcid.org/0009-0001-6867-4488

approach gives a guideline, included components and

also requirements, how such a Stream Processing

should occur.

Along with the rise of Stream Processing tools

were developed to model and implement EPNs.

Therefore, we contribute here an evaluation of differ-

ent recent tools, which support EPN realization as au-

tomatically as possible. The evaluation also addresses

the following questions: What are EPns and which

requirements are important to provide them? A de-

tailed look is taken at Apache Storm, Amazon Kinesis

Data Analytics and – in this paper – Microsoft Azure

Stream Analytics.

The following article enhances our work from

(Schulze et al., 2023) and (Koschel et al., 2023) and

provides the following contributions: First, we brieﬂy

look at our – compared to our work from (Koschel

et al., 2017) – more formally structured list of general-

ized EPN model requirements (as shown in (Schulze

et al., 2023) in detail). Second, we provide – as the

key contribution of the present article – an in-depth

evaluation of Microsoft Azure Stream Analytics with

respect to our EPN model requirements. Moreover,

we also brieﬂy compare Microsoft Azure Stream An-

alytics to Apache Storm and Amazon Kinesis Data

(AKD) Analytics.

The remainder of this paper is structured as fol-

lows: After discussing related work in Section 2 we

Koschel, A., Pakosch, A., Schulze, C., Astrova, I., Gerner, C. and Tyca, M.

Would Microsoft Azure Stream Analytics Be a Suitable Foundation for an Event Processing Network Model?.

DOI: 10.5220/0013073000003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 1, pages 15-22

ISBN: 978-989-758-737-5; ISSN: 2184-433X

place a brief introduction to the topic and provide our

EPN model requirements in Section 3. Next, we take

an in-depth look at Microsoft Azure Stream Analyt-

ics in Section 4 followed by a brief comparison to

Apache Storm and Amazon Kinesis Data (AKD) An-

alytics in Section 5. Finally, Section 6 summarizes

our results and leads to a conclusion.

2 RELATED WORK

The basis of our project builds on authors in the

scope of EPN and Complex Event Processing (CEP)

for example, Dunkel and Bruns (Dunkel & Bruns,

2015) and (Bruns & Dunkel, 2010). We also used

foundations from our earlier work on EPNs, namely

(Koschel et al., 2017), (Koschel et al., 2018) and (As-

trova et al., 2019). There we more informaly establish

the requirements for EPNs and apply them to different

EPN modeling approaches and tools (EPiA, BEMN,

RuleCore). With the present paper we extend our

work with slightly reﬁned and more formally struc-

tured requirements as well as a deep look at more

recent tools, here in particular at Microsoft Azure

Stream Analytics.

Compared to our earlier work, here we cast the re-

quirements into a template from (Rupp & Pohl, 2021)

that means we somewhat formalize them. We use a

template to deﬁne the requirements in a standardized

form and to show their importance. To ensure the

quality of the requirements, we validated them against

the quality criteria from (IEE, 1998).

Furthermore, we use the IEEE830-1998 stan-

dard (IEE, 1998) for quality criteria for requirements.

There exists a newer standard IEEE/ISO/IEC29148-

2011, which describes the quality criteria from

IEEE830-1998 in more summarized form. We still

meet all the quality criteria from both versions,

except for the singularity. That one is new in

IEEE/ISO/IEC29148-2011.

For the description and evaluation of the differ-

ent tools we used the documentation of the publishers.

Apache offers large documentation in (Str, 2021a) for

Apache Storm, Amazon Web Services provides in-

formation in (Str, 2021b) for Amazon Kinesis Data

(AWD) Analytics and Microsoft introduces Microsoft

Azure Stream Analytics in (Microsoft, 2021b).

In contrast to other publications, we distinguish

ourselves by standardizing and validating the require-

ments of EPNs and by evaluating various tools with

different open or closed source characteristics, ef-

fort, and costs. With this variety, we aim to give an

overview of different tools and support the decision

for an individually suitable tool.

3 EPNs AND REQUIREMENTS

This section curtly presents the basics of EPNs and

the requirements for this kind of systems.

According to (Luckham, 2002) an event is a sig-

niﬁcant change of state. Event Processing Networks

(EPNs) can be seen as generalized software systems

that allow the processing of events. However, EPN

models lack standardization, which is where our work

aims to contribute.

3.1 Basics of Event Processing

Networks

EPNs are build on the basis of the Event-Driven Ar-

chitectures (EDA) and Complex Event Processing

(CEP). These both approaches

’[...] represent a new style of enterprise ap-

plications that places events at the center of

the software architecture - event orientation as

an architectural style.’ — Dunkel and Bruns

(Bruns & Dunkel, 2010, p. 4)

In this context, EDA is more about the design of

event-driven architectures as a design style. CEP de-

scribes a technology for dynamic processing of large

datasets (Bruns & Dunkel, 2010). Thus, CEP is a

part of an EDA, which can be used for processing

data within it. In detail CEP describes the dynamic

processing of large data streams (also called event

streams) in real-time. An event is any happening in

the system. Here the change of state of a fact or an

object is represented (Bruns & Dunkel, 2010).

The processing of events within a CEP is realized

by using rules. These rules contain knowledge about

handling events or event sequences (Dunkel & Bruns,

2015). For the realization of these rules and the pro-

cessing of the data CEP contains Event Processing

Agents (EPA).

An EPN is a set of EPAs, which are intercon-

nected and exchange information during and about

processing of the data (Dunkel & Bruns, 2015). An

EPN can be interpreted as a graphical tool for mod-

eling the ﬂow of events for event processing systems

(Koschel et al., 2017). Thus, the main components

of EPN are EPAs in order to be able to perform CEP.

EPAs contain various components like Event Model,

(Event) Rules, and Event Procession Engine (Dunkel

& Bruns, 2015).

Other components, for example producers, are

also further elements of EPNs and can be taken from

(Dunkel & Bruns, 2015) and (Koschel et al., 2017).

The next part explains the requirements for EPNs.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

3.2 Requirements

We evaluate the selected tools following an identical

set of requirements, which we have put into a stan-

dardized form as shown next.

3.2.1 Handling the Requirements

The set of requirements origins from our work in

(Koschel et al., 2017). We have standardized the form

of these requirements in (Schulze et al., 2023) by ap-

plying (Rupp & Pohl, 2021) and (IEE, 1998). The

reason was that the original requirements were just

described as bullet points, had no formal structure and

were partially a little ambiguous.

To address these issues, we evaluated various re-

quirement templates how they address issues such as

writeability, readability and learnability and common

use (Robertson & Robertson, 2012). We have chosen

(Rupp & Pohl, 2021) because it provides a straight-

forward structure for requirements.

In addition, we apply the quality criteria of IEEE

830-1998 (IEE, 1998) to achieve high quality require-

ments in structure and content. Speciﬁcation of the

quality criteria according to (IEE, 1998) are require-

ments, that are correct, unambiguous, complete, con-

sistent, veriﬁable, modiﬁable, traceable and ranked

for importance and/or stability.

The requirements are formulated according to a

template and fulﬁll all quality criteria. Those tem-

plates achieve writeability, readability and learnabil-

ity and are therefore efﬁcient. This also satisﬁes the

modiﬁable criteria from IEEE 830-1993.

Correctness, unambiguousness and completeness

are achieved by splitting, expanding and substitut-

ing specialist terms. Requirements are checked to be

consistent, veriﬁable, traceable and they are ranked

by importance (see more details in (Schulze et al.,

2023)).

3.2.2 The Requirements

Our standardized requirements are as follows:

• EPNR1. The tool shall offer the developer to

model events with their inherent attributes as the

central component of the engine.

• EPNR2. The tool shall map real world descrip-

tions to events as scenarios.

• EPNR3. The tool shall offer event structures as

simple, complex or aggregated. Simple events can

be created and used independently. In addition,

complex events have dependencies and references

to other events. Also, aggregated events can be

grouped logically.

• EPNR4. The tool shall offer possibilities to ex-

press the relativity of events and their temporal

and causal relationships, e.g. sequence, precondi-

tions and postconditions.

• EPNR5. The tool shall process and show the ﬂow

of events through the system.

• EPNR6. The tool shall offer the modeling of EPN

by components, their properties and used patterns.

• EPNR7. The tool shall offer the modeling of

components outside the system boundary and the

behavior between inside and outside components.

• EPNR8. The tool should be expressive in usage,

about readability, writability, learnability and efﬁ-

ciency.

• EPNR9. The tool should offer the developer fur-

ther possibilities to create the model, e.g. IDE,

graphical event programming.

In (Schulze et al., 2023) we evaluated Apache

Storm (Str, 2021a) and in (Koschel et al., 2023) we

evaluated Amazon Kinesis Data Analytics. As the

major contribution of the present paper we will eval-

uate Microsofts Azure Stream Analytics against our

requirements.

4 MICROSOFT AZURE STREAM

ANALYTICS

In this section, Microsoft Azure Stream Analytics is

examined and evaluated. It was chosen by the au-

thors for its closed source characteristics and because

of Microsofts signiﬁcant position in the market.

4.1 Overview of Microsoft Azure

Stream Analytics

Azure Stream Analytics is a real-time CEP engine

that is designed to analyze and process high volumes

of fast streaming data. Data can be obtained from

multiple sources simultaneously. It can be catego-

rized as a Platform-as-a-Service (PaaS) and thus it is

fully managed within Azure (Microsoft, 2021b).

Azure Stream Analytics guarantees exactly-once

event processing semantics and at-least-once event

delivery, to avoid losing events. It also provides built-

in checkpoints to maintain the status of your job and

delivers repeatable results (Microsoft, 2021a).

For realizing the stream processing, Azure Stream

Analytics provides the following components (Mi-

crosoft, 2021b):

Would Microsoft Azure Stream Analytics Be a Suitable Foundation for an Event Processing Network Model?

4.1.1 Jobs

The core of Azure Stream Analytics are jobs. Jobs

have to be created by developers themselves and con-

sist of Input, Query Processing Engine and Output

(see Figure 1). A job is a self-contained unit. How-

ever, jobs can be attached to each other so that the

output of a previous job would be the input of the

following job. This allows complex events to be pro-

cessed in multiple stages. Furthermore, a job can have

multiple inputs and outputs.

4.1.2 Input

As Input Azure Stream Analytics provides IoT Hubs,

Event Hubs, Azure Blob Storage or Azure Data Lake

Storage. Hubs and Storages also provide an interface

for external producers to transfer data into a job. Hubs

are classic message queues that collect data, store it

temporarily, and transmit it to the Query Processing

Engine on demand.

4.1.3 Output

The Output can provide Azure resources for storage

and archiving, data warehousing, dynamic visualiza-

tion or sending alerts and executing actions. Azure

resources are possible as output, however these can

transfer the output to external targets.

4.1.4 Query Processing Engine

Azure Stream Analytic and its Query Processing En-

gine is built on Trill – a streaming query processor de-

veloped by Microsoft Research (Chandramouli et al.,

2014). In order to examine the Query Processing En-

gine it is necessary to take a closer look at Trill, since

the engine is subject to the concept.

Trill (Trillion events per day) is an in-memory

streaming analytics engine, which reads input data

exactly once. It is designed to handle a wide range

of data, both real-time and ofﬂine, and is based on a

temporal data and query model. Trill can be used as a

streaming engine, a lightweight relational in-memory

engine, and a progressive query processor.

In addition, Trill was developed to address the

following speciﬁc requirements (Chandramouli et al.,

2014):

• Query Model. A Stream Analytics Engine must

be able to process various data. On one hand, of-

ﬂine data should be processed, such as logs; on

the other hand, real-time data should be processed

in order to immediately publish warnings in case

of possible problems. Real-time data should be

able to be processed with historical data to detect

correlations. Finally, progressive data should be

processed to provide fast and continuous informa-

tion.

• Fabric & Language Integration. A Stream Ana-

lytics Engine must be easily usable by a high-level

language (HLL) in order for the engine to be used

by an application. HLLs such as Java and C# offer

a rich volume of data types, libraries and custom

logic that must be supported by the engine.

• Performance. Processing large ofﬂine data sets

requires high throughput, while real-time moni-

toring requires low latency.

Trill meets all the above requirements with the help of

the hybrid system architecture (Chandramouli et al.,

2014). There are two modes of Trill deﬁned as Li-

brary mode and Multi-core mode. In Library mode,

which is the default, Trill is only executed in one

thread and can thus be integrated into other applica-

tions. Compared to the Library mode, higher perfor-

mance is achieved in Multi-core mode. Here sev-

eral threads are executed in parallel. It supports a

new two-level streaming temporal MapReduce oper-

ation, executed using a lightweight optional sched-

uler (Chandramouli et al., 2014).

Trill offers a query language called Trill-LINQ.

This language enables CEP and also forms the ba-

sis of Query Processing Engine. The following core

concepts are provided by Trill-LINQ (Chandramouli

et al., 2015):

• Filtering and Projection. Similar to SQL, the pro-

jection of an event can be complete or transform

e.g. selected attributes. Furthermore, events can

be ﬁltered based on conditions.

• Windowing. Time intervals within the processing

are deﬁned so that each event is assigned to a time

window and processed inside of it. Events always

contain a timestamp since they are processed by

Trill strictly non-decreasing. Based on the length

of the time intervals the developer controls the

size of the batches and thus also the performance

with latency and throughput. There are several

windows, which are described in the context of

Azure Stream Analytics.

• Aggregation. Close to SQL aggregation can be

used to combine events and transform them into

new values, e.g., number of events.

• Grouped Computation. Similar to SQL a group-

ing is performed, where the result of a sub-

operation is provided to a speciﬁc key.

• Correlation and Set Difference. Trill allows to

perform temporary join operations on different

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

streams or datasets and merge them. Any combi-

nation of real-time data, historical data or ofﬂine

data is possible.

• Data-Dependent Windowing. The windows are

selected based on the data. The window can be a

time interval, as in classic windowing, or an event

that can occur and thus shorten the time interval.

Trill has only been described rudimentarily, tai-

lored to the article to explain Azure Stream Analytics.

A detailed description would exceed the scope of this

paper, but can be found in (Chandramouli et al., 2014)

and (Chandramouli et al., 2015).

By incorporating Trill as a library, Azure Stream

Analytics was up and running within 10 months.

Queries are created using Transact SQL and then

translated into Trill expressions by a SQL compiler

(Terwilliger, 2018).

Since Trill is designed for high compatibility with

high-level languages and their numerous data types,

the Query Processing Engine can process data in

JSON, AVRO or csv format. Due to the possible

nesting of JSON and AVRO, events of arbitrary struc-

ture and complexity can be processed. This ﬂexibil-

ity means, that Transact SQL can be extended with

JavaScript and C# user-deﬁned functions (Microsoft,

2021c). As known from SQL, this language also has

operations to easily ﬁlter, sort, aggregate, and join

streaming data.

As in Trill time is important since it is essential

in CEP. Thus, it must be ensured that it is easy to

work with this time component. Windowing was in-

troduced for this case. In Azure Stream Analytics

events are also processed based on their timestamp.

In total there are ﬁve different time windows in Azure

Stream Analytics (Microsoft, 2022):

• Tumbling Window. Divide into disjoint time in-

tervals, events, whose time stamps are greater than

the interval start and less than or equal to the in-

terval end, are assigned to this interval. E.g. 12:00

PM – 12:05 PM window will include events that

happened exactly at 12:05 PM, but will not in-

clude events that happened at 12:00 PM.

• Hopping Window. Intervals of ﬁxed length over-

lapping and offset by one HOP parameter. E.g.,

interval length is 10 minutes and HOP parameter

is 5 minutes, which means the ﬁrst interval is from

12:00 PM - 12:10 PM and the second interval is

from 12:05 PM - 12:15 PM. Events whose times-

tamp is 12:06 PM - 12:10 PM are in both intervals.

• Sliding Window. A ﬁxed length interval is shifted

using the time axis. In order not to generate an

inﬁnite number of possible shifts, a new window

is only created when events have left or joined this

window.

• Session Window. The intervals depend on the

arrival of the events and the maximum interval

length. It is deﬁned by a timeout and the inter-

val length. If no further event occurs within the

timeout after the arrival of an event, the interval is

closed. As soon as a new event arrives, a new in-

terval is opened. However, even with permanently

arriving events, the interval length is never longer

than the maximum length.

• Snapshot Window. All events with the same

timestamp are assigned to one snapshot.

In addition to windowing, queries can be nested. So,

it is possible to deﬁne sub-queries and process them

within the query of a job. Within the sub-queries, dif-

ferent sources can be used and correlated.

The Query Processing Engine offers multiple

functions through Trill to process streams. Beside

the windowing and the accompanying division of the

streams, events can be processed with a SQL based

query language. Queries can be nested to structure

the queries as needed. Functions can be created to

make the processing more precise. Finally, jobs can

be attached to each other. In addition to the stream,

other data such as historical, reference, or scoring can

be used to implement processing or correlate relation-

ships (see Figure 1).

This part introduced the functionality of Azure

Stream Analytics and Trill. By using a SQL-based

language, streams can be examined and analyzed.

The next part examines the requirements.

4.2 Evaluation of Microsoft Azure

Stream Analytics for Modeling EPN

This part will argue, which requirements are fulﬁlled

or not fulﬁlled by Microsoft Azure Stream Analytics.

• EPNR1 - fulﬁlled:

Events can be processed in JSON, AVRO or csv

format. Both, JSON and AVRO data, can be struc-

tured and contain some complex types such as

nested objects (records) and arrays.

• EPNR2 - fulﬁlled:

Due to the freely selectable structure of JSON and

AVRO data, things of the real world can be de-

scribed and mapped in scenarios.

• EPNR3 - fulﬁlled:

Events can be grouped and aggregated by queries.

In addition, real-time data can be correlated with

historical data to identify patterns. Finally, there

is the windowing concept, in which groupings can

be created based on time intervals.

Would Microsoft Azure Stream Analytics Be a Suitable Foundation for an Event Processing Network Model?

Figure 1: Stream processing concept in Azure (based on

(Microsoft, 2021b)).

• EPNR4 - fulﬁlled:

Events can be processed causally on multiple lev-

els. Several jobs can be executed in sequence,

furthermore within one job there can be several

queries and processing steps.

• EPNR5 - fulﬁlled:

Events are processed in a granular manner using

at-least-once delivery and exactly-once process-

ing. Job diagrams within the Azure interface vi-

sualize the ﬂow of events from exactly one job.

• EPNR6 - fulﬁlled:

The given components in Azure can be used to

model an EPN. As described, there are Input,

Query Stream Processing and Output. In essence,

the Query Stream Processing offers the possibility

of processing using the Stream Analytics Query

Language. Extensions are supported by user-

deﬁned functions.

• EPNR7 - fulﬁlled:

Event producers do not need to be part of the sys-

tem. Any kind of data can be imported from out-

side, e.g. from IoT devices, weather data, proto-

cols.

• EPNR8 - fulﬁlled:

Various tutorials, from step-by-step beginner to

advanced concepts and examples for end to end

scenarios.

• EPNR9 - fulﬁlled:

Implementation is possible in Azure Portal or Vi-

sual Studio (Code), Azure CLI, Terminals.

Azure Stream Analytics also meets all the require-

ments for modeling an EPN. Due to the PaaS charac-

ter, developers contain a platform that enables them

to speciﬁcally implement the use cases of stream pro-

cessing. However, the given functions and features of

Azure Stream Analytics, even though it is SQL-based,

are very complex and can be a barrier to entry.

5 COMPARISON

In this article, we analyzed Microsoft Azure Stream

Analytics in Section 4 in-depth with respect to our

standardized set of requirements from Section 3.2.2.

As we stated there, Microsoft Azure Stream Analyt-

ics nicely fullﬁls our requirements and thus it is well

suited for the realization of EPNs.

Mainly the collected information for the presenta-

tion of Microsoft Azure Stream Analytics (as well as

other tools) was taken from the documentation of the

publishers or developers of the tools. For this reason,

some information may be presented subjectively, as

companies would like to widely distribute their tool

in any case. Moreover, – in contrast to Apache Storm

– Amazon and Microsoft Azure are costly tools, so

an advertising factor within the documentation can-

not be ruled out. Furthermore, information may be

incomplete, because companies want to keep their im-

plementations private.

To provide some more distinctive criteria to other

tools we took in particular a more developer-oriented

perspective. The result is summarized in table 1,

which brieﬂy compares all three tools under evalu-

ation, namely Apache Storm, Amazon Kinesis Data

(AKD) Analytics and Microsoft Azure Stream Ana-

lytics. Developers may use this table to identify the

most important criteria that argue for or against a tool.

In particular open source nature, price, convenience,

and potential vendor lock in some distinctive factors.

While Apache Storm is open source and free of

charge, Amazon Kinesis Data (AKD) Analytics and

Microsoft Azure Stream Analytics are not. However,

while Apache Storm has to be maintained by one

selves, Amazon Kinesis Data (AKD) Analytics and

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

Table 1: Tools Criteria and Characteristics.

Apache Storm

Amazon Kinesis

Data Analytics

Microsoft Azure

Stream Analytics

Basis MapReduce MapReduce Trill

Support no support

Plattform-as

-a-Service (PaaS)

Plattform-as

-a-Service (PaaS)

Costs no costs costs based on usage costs based on usage

Effort high low average

Eventformat Tuple Tuple JSON, AVRO, csv

Environment JVM AWS Azure

Language

topology in Java,

others arbitrary

Java, Scala,

Python, SQL

SQL based,

JavaScript or C#

user-deﬁned functions

Distribution Thread-based

Thread-based in

Apache Flink

no details

Maximum Processing Rate

over one million tuples

per second and per node

real-time real-time

Reliability guaranteed by spouts guaranteed by AWS 99.9% promised

Data protection own realization

Server location can

be set

Server location can

be set

Security own realization provided by AWS provided by Azure

Microsoft Azure Stream Analytics are nicely hosted

and maintained by Amazon and Microsoft and possi-

bly easier to be used.

Thus, there is no clear winner between the three

tools, but more a question of individual developer

taste and skills as well as company preferences. For

example, if a company prefers open source tools and

has good development and hosting skills in house,

then Apache Storm seems to be more suitable. If a

company is an AWS shop anyway, wants likely less

maintenance effort and is able to pay the costs for Ki-

nesis, then Kinesis could be more favorable. Like-

wise Microsoft Azure Stream Analytics could be the

favourite option, if a company is a strong Microsoft

Azure user anyway.

6 CONCLUSION

In this article we had a deep look at Event Process-

ing Network Models as a foundation of Event Stream

Processing tools (cf. (Schulze et al., 2023)) and pre-

sented our enhanced (compared to our earlier work)

standardized set of requirements for EPN models in

Section 3.2.2.

As the key contribution of this article we apply

those requirements, in order to have an in-depth look

at Microsoft Azure Stream Analytics in Section 4. It

turns out that Microsoft Azure Stream Analytics is a

well suited tool for modeling and implementation of

EPNs. Additionally we summarized our comparison

of Microsoft Azure Stream Analytics, Amazon Kine-

sis Data Analytics (Koschel et al., 2023), and Apache

Storm (Schulze et al., 2023) in Section 5.

Since all those tools mostly fulﬁlled our require-

ments comparisons may need other criteria as well.

The suitability of a tool depends on more individual

circumstances, such as which kind of ’shop’ you are

– for example, Amazon vs. Microsoft –, but also how

high the own development and administration effort

should be.

Therefore, the decision for a suitable tool is based

on the effort, the control and the costs involved. For

these reasons, no absolute recommendation can be

made. Rather the authors recommend examining each

individual use case or at least a set of typical ones, in

order to select the ideal tool. Since the ﬁeld of event

processing is more important than ever, we will con-

tinue to evaluate upcoming tools in this space within

future work of ours.

REFERENCES

(1998). IEEE Recommended Practice for Software Re-

quirements Speciﬁcations. IEEE Std 830-1998, (pp.

1–40).

(2021a). Apache Storm. Apache Software Foundation. On-

line: https://storm.apache.org/ [retrieved: 04, 2022].

(2021b). Streaming Data Solutions on AWS. Amazon Web

Services Inc. Online: https://docs.aws.amazon.com/

whitepapers/latest/streaming-data-solutions-amazon-

kinesis/welcome.html [retrieved: 04, 2022].

Would Microsoft Azure Stream Analytics Be a Suitable Foundation for an Event Processing Network Model?

Astrova, I., Koschel, A., Kobert, S., Naumann, J., Ruhe, T.,

& Starodubtsev, O. (2019). Evaluating RuleCore as

Event Processing Network Model. In Proc. 15th In-

ternational Conference on Web Information Systems

and Technologies (WEBIST 2019) (pp. 297–300). Vi-

enna, Austria: SCITEPRESS -– Science and Technol-

ogy Publications.

Bruns, R. & Dunkel, J. (2010). Event-Driven Archi-

tecture - Softwarearchitektur f

ur ereignisgesteuerte

Gesch

aftsprozesse (Software architecture for event-

driven business processes). Springer.

Chandramouli, B., Goldstein, J., Barnett, M., DeLine, R.,

Fisher, D., Platt, J. C., Terwilliger, J. F., & Werns-

ing, J. (2014). Trill: A high-performance incremental

query processor for diverse analytics. Proceedings of

the VLDB Endowment, 8(4), 401–412.

Chandramouli, B., Goldstein, J., Barnett, M., & Terwilliger,

J. F. (2015). Trill: Engineering a library for diverse

analytics. IEEE Data Eng. Bull., 38(4), 51–60.

Dunkel, J. & Bruns, R. (2015). Complex Event Processing

- Komplexe Analyse von massiven Datenstr

omen mit

CEP (Complex analysis of massive data streams with

CEP). Springer Vieweg.

Koschel, A., Astrova, I., Kobert, S., Naumann, J., Ruhe, T.,

& Starodubtsev, O. (2017). Towards Requirements for

Event Processing Network Models. In Proc. 8th In-

ternational Conference on Information, Intelligence,

Systems, Applications (IISA 2017) (pp. 27–30). Lar-

naca, Cyprus: IEEE.

Koschel, A., Astrova, I., Kobert, S., Naumann, J., Ruhe,

T., & Starodubtsev, O. (2018). On Requirements

for Event Processing Network Models Using Business

Event Modeling Notation. In Proc. 2018 Conf. Intelli-

gent Computing. Advances in Intelligent Systems and

Computing (SAI 2018) (pp. 756–762). London, UK:

Springer.

Koschel, A., Astrova, I., Pakosch, A., Gerner, C., Schulze,

C., & Tyca, M. (2023). Is Amazon Kinesis Data Ana-

lytics Suitable as Core for an Event Processing Net-

work Model? In Proc. 16th International Confer-

ence on Agents and Artiﬁcial Intelligence (ICAART

2024) (pp. 1036–1043). Rome, Italy: INSTICC,

SCITEPRESS.

Luckham, D. (2002). The Power of Events. Addison Wes-

ley, USA.

Microsoft (2021a). Event Delivery Guarantees (Azure

Stream Analytics). Microsoft Documentation.

Online: https://docs.microsoft.com/en-us/stream-

analytics-query/event-delivery-guarantees-azure-

stream-analytics [retrieved: 09, 2024].

Microsoft (2021b). Introduction to Azure Stream

Analytics. Microsoft Documentation. Online:

https://docs.microsoft.com/en-us/azure/stream-

analytics/stream-analytics-introduction [retrieved:

04, 2022].

Microsoft (2021c). Parse JSON and Avro data in Azure

Stream Analytics. Microsoft Documentation. On-

line: https://docs.microsoft.com/en-us/azure/stream-

analytics/stream-analytics-parsing-json [retrieved:

04, 2022].

Microsoft (2022). Windowfunctions (Azure Stream

Analytics). Microsoft Documentation. Online:

https://docs.microsoft.com/de-de/stream-analytics-

query/windowing-azure-stream-analytics [retrieved:

09, 2024].

Robertson, S. & Robertson, J. (2012). Mastering the

Requirements Process: Getting Requirements Right.

Addison-Wesley Professional.

Rupp, C. & Pohl, R. (2021). Basiswissen Requirements En-

gineering (Basic knowledge Requirements Engineer-

ing). dpunkt.verlag.

Schulze, C., Gerner, C., Tyca, M., Koschel, A., Pakosch,

A., & Astrova, I. (2023). Analyzing Apache Storm

as Core for an Event Processing Network Model.

In Proc. International Conference Intelligent Systems

(IntelliSys 2023), LNNS 824 (pp. 397–410). Amster-

dam, Netherlands: Springer, Cham.

Shaikh, T. (2019). Batch Processing — Hadoop Ecosys-

tem. K2 Data Science and Engineering. Online:

https://blog.k2datascience.com/batch-processing-

hadoop-ecosystem-f6da88f11cae [retrieved: 04,

2022].

Terwilliger, J. (2018). Microsoft open sources

Trill to deliver insights on a trillion events

a day. Micro-soft Blog Developer. Online:

https://azure.microsoft.com/de-de/blog/microsoft-

open-sources-trill-to-deliver-insights-on-a-trillion-

events-a-day/ [retrieved: 09, 2024].

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence