AI and Software Architecture - Two Sides of the same Coin

Introduction

The whole media worldwide is currently jumping on the AI bandwagon. In particular, Large Language Models (LLM) such as ChatGPT sound appealing and intimidating at the same time. When we dive deeper into the technology behind AI, it doesn‘t feel that strange at all. In contrast to some assumptions of the yellow press, we are far away from a strong AI that resembles human intelligence. This means, blockbusters such as Terminator or Bladerunner are not becoming true in the near future. 

Current AI applications, while very impressive, represent instantiations of weak AI.  Take object detection as an example, where a neural network learns to figure out what is depicted on an image. Is it a cat, a dog, a rabbit, a human or something different? Eventually, neural networks process training data to compute and learn a nonlinear mathematical function that works incredibly well for making good guesses (aka hypotheses) with high precision about new data. 

On the other side, this capability proves to be very handy when dealing with big or unstructured data such as images, videos, audio streams, time series data, or Kafka streams. For example, autonomous driving systems strongly depend on such kind of functionality, because they continuously need to analyze, understand and handle highly dynamic traffic contexts, e.g., potential obstacles.

In this article, I am not going to explain the different kinds of AI algorithms such as types of artificial neural networks and ML (Machine Learning) which may be part of a subsequent article. My goal is to draw the landcape of AI with respect to software architecture & design.

There are obviously two ways of applying AI technologies to software architecture:

• One way is to let AI algorithms support software architects and designers in their tasks such as requirements engineering, architecture design, implementation or testing - which I’ll call the AI solution domain perspective.
• The other way is the use of AI to solve specific problems in the problem domain, why I’ll name it the AI application domain perspective.

AI for the Solution Domain

LLMs are probably the most promising approach when we consider the solution domain. Tools such as GitHub Copilot, Meta Llama 2 and Amazon CodeWhisperer help developers generate functionality in their preferred programming language. It seems like magic but comes with a few downsides. For example, you never can be sure whether an LLM learned its code suggestions from copyrighted sources. Nor do you have any guarantee that the code does the right thing in the right way. Any software engineer who leverages an application like Copilot needs to look over the generated code again and again to ensure the code is exactly what she or he expects. It requires software engineering experts to continuously analyze and check LLM answers. At least currently, it appears rather unlikely that laymen may take over the jobs of professional engineers with the help of LLMs. 

Companies already have began to create their own LLMs to cover problem domains such as industrial automation. Imagine, you need to develop programs for a PLC (Programmable Logic Control). In such environments, the main languages are not C++, Python or Java. Instead you’ll have to deal with domain-specific languages such as ST (Structured Text = Siemens SCL) or LD (Ladder diagram). Since there is much less source code freely available for PLCs, feeding an LLM with appropriate code examples turns out to be challenging. Nonetheless, it is a feasible objective. 

AI for the Application Domain

In many cases Artificial Neural Networks (ANNs) are the basic ingredient for solving problem domain challenges. Take logistics as an example where cameras and ANNs help identity which product is in front of a camera. Other AI algorithms such as SVNs (Support Vector Machines) enable testing equipment to figure out whether a turbine is behaving according to its specification or not, which is commonly coined Anomaly Detection. At Siemens we have used Bayes-Trees to forecast the possible outcome of system testing. Reinforcement Learning happens to be useful for successfully moving and acting in an environment, for example robots learning how to  complete a task successfully. Another approach is unsupervised learning such as k-Means Clustering which classifies objects and maps them to different categories. 

Even more examples exist:

Think about security measures in a system that comprise keyword and face recognition. Autonomous driving uses object detection and segmentation in addition to other means. Smart sensors include ANNs for smell and gas detection. AI for preventive maintenance helps analyzing whether a machine might fail in the near future based on historical data. With the help of recommender systems online shops can provide recommendations to customers based on their order history and product catalog searches. As always, this is only the tip of the iceberg.

Software Architecture and AI

An important topic seldomly addressed in AI literature is how to integrate AI in a software-intensive system. 

MLOps tools support different roles like developers, architects, operators and  data analysts. Data analysts start with a data collection activity. They may augment the data, apply feature extraction as well as regularization and normalization measures, and select the right AI model which is supposed to  learn how to achieve a specific goal using the data collection. In the subsequent step they test the AI/ML-model with sufficient test data, i.e. data the model has not seen before. Eventually, they version the model & data and generate an implementation. Needless to say that data analysts typically iterate through these steps several times. When MLOps tools such as Edge Impulse follow a No/Low-Code approach, separation of concerns between different roles can be easily achieved. While data analysts are responsible for the design of the AI model, software engineers can focus on the integration of the AI model in the system design process, as the MLOps envoronment generates implementation of the model.

Software engineers take the implementation and integrate it into the surrounding application context. For example, the model must be fed with new data by the application which reads and processes the results once inference is completed. For this purpose, an event-driven design often turns out to be appropriate, especially when the inference runs on a remote embedded system. If the inference results are critical, resilience might be increased by replicating the same inference engine multiple times in the system. Docker containers and Kubernetes are perfect solutions, in particular when customers desire a scalable and platform-independent architecture with high separation of concerns like in a Microservice architecture. Security measures support privacy, confidentiality, and integrity of input data, inference results and the model itself. In most cases, inference can be treated from a software engineering viewpoint mostly as a black box that expects some input and produces some output. 

When dealing with distributed systems or IoT systems, it may be beneficial to execute inference close to the sources of input data, thus eliminating the need to send around big chunks of data, e.g., sensor data. Even embedded systems like edge or IoT nodes are capable of running inference engines efficiently. In this context, only inference results are often sent to backend servers.

Operators finally deploy the application components onto the physical hardware. Note: a DevOps culture turns out to be even more valuable in an AI context, because more roles are involved.

Input sources may be distributed across the network, but may also comprise local sensor data of an embedded system. In the former case, either Kafka streams or MQTT messages can be appropriate choices to handle the aggregation of necessary input data on behalf of an inference engine. Take processing of weather data as an example where a central system collects data from various weather stations to forecast the weather in a whole region. In this context we might encounter pipelines of AI inference engines, where the results of different inference engines are fed to a central inference engine. Hence, such scenarios comprise hierarchies of possibly distributed inference engines.

Architecting AI models

Neural networks or other types of AI algorithms expose an architecture themselves, be it a MobileNet model leveraged for transfer learning, a SVN (Support Vector Machines) with a Gaussian kernel, or a Bayes decision tree. The choice of an adequate model has significant impact on the results of AI processing. It requires the selection of an appropriate model and hyperparameters such as learning rate or configuration of layers in an ANN (Artificial Neural Network). For data analysts or those software engineers who wear a data analytics hat a mere black box view is not sufficient. Instead they need a white box view to design respectively configure the appropriate AI model. This task depends on the experience of data analysts, but may also imply a trial-and-error approach for configuring and fine tuning the model. The whole design process for AI models closely resembles software architecture design. It consists of engineering the requirements (goals) of the AI constituents, selecting the right model and training data, testing the implemented AI algorithm, and deploying it. Consequently, we may consider these tasks as the design of a software subsystem or component. If an aforementioned MLOps tool is available und used, it significally can boost design efficiency.

Conclusions

While the math behind AI models may appear challenging, the concepts and usage are pretty straightforward. Their design and configuration is an important responsibility that experts in Data Analytics and AI should take care of. MLOps helps separate different roles and responsibilities which is why I consider its use as an important development efficiency booster. 

Architecting an appropriate model is far from being simple, but resembles the process of software design. Training an AI model for ML (Machine Learning) may take weeks or months. As it is a time consuming process, the availability of performant servers is inevitable. Specialized hardware such as Nvidia GPUs or other dedicated NPUs/TPUs helps reduce the training time significantly. In contrast to the amount of required training efforts, optimised inference engines (-> Tensorflow Lite or Lite Micro) often run well and efficient on resource constrained embedded systems which is the concept behind AIoT (AI plus IoT).

 Systematic Re-use

Re-use is based upon one of the fundamental principles not only for lazy software engineers: DRY (Don‘t Repeat Yourself). Instead of reinventing the wheel developers and architects may re-use existing artifacts instead of reinventing the wheel again and again. 

Re-usable assets come in different flavors:

• Code snippets are small building units developers may integrate in their code base. 
• Patterns are smart and proven design blueprints that solve recurring problems in specific contexts.
• Libraries comprise encapsulated functionality developers may bind to their own functionality.
• Frameworks also comprise encapsulated functionality. In contrast to libraries developrs integrate their own code into the framework according to the Hollywood principle (don‘t call us, we‘ll call you).
• Components/Services include binary functionality (i.e., they are executables) that developrs may call from their own application.
• Containers represent runtime environments that provide functionality and environments to applications in an isolated way.

Apparently, these are different levels of re-usable assets with varying granularities, complexities, and prerequisites.

Software engineers may not only use re-usable software assets, but other types as well. For instance:

• Tests, Test units, Test plans
• Documents
• Production plans
• Configurations
• Business plans
• Software architectures
• Tools

While some assets such as code snippets may be used daily in the code-base, patterns or software architecture templates need to be instantiated in an easy way. 

The more impact re-usable assets have on applications and the more abstract they are, the more systematic the re-use approach must be. The most challenging projects are product lines and ecosystems that require different assets at different re-use levels. For example, they introduce the need for a configurable core asset base that is re-usable across different applications. Furthermore, they support a whole class of applications that share the same architecture framework and other assets. A core asset in a product line or ecosystem affects not one application but a whole system family.  Thus, its business impact is very high. 

In such scenarios, core assets often are inter-dependent and must be configured for the specific application under development.  As a prerequisite for the development of a core asset base, a Commonality/Variability analysis is necessary that determines what applications sharing the same core assets have in common and how they differ. A core asset needs a common base relevant for all applications that use it as well as configurable variation points to adapt it to the needs of an application. 

A bad or insufficient  Commonality/Variability analysis incurs higher costs a may even lead to project failure. 

Core asset development and application development might happen separately by different teams or  by the same teams. Each approach has its benefits and liabilities.

Due to the high business and technical risks of these advanced approaches, all stakeholders need to be involved in the whole development process. Building a product line or ecosystem without management is not feasible. Managers need to re-organize their organisation, spend budget for core asset development and evolution. 

Most product lines and ecosystems fail, because:

• lack of management support,
• insufficient consideration of customer needs,
• inappropriate organisation,
• inadequate Commonality/Variability analysis,
• insuffient or low-quality core assets,
• underestimation of testing or inadequate quality assurance,
• bad software architecture,
• neglectence of competence ramp-up activities,
• no re-use incentives,
• missing acceptance by stakeholders.

Consequently, product lines and ecosystems need a systematic approach for re-use  and must involve  different types of stakeholders. They need a manager  who is able to guide the approach and has the capability to decide, for example, on budget, organisation restructuring, competence ramp-up activities, or business strategy. 

Re-use comes in different flavors and the higher its impact the more systematic the re-use process needs to be

[to be continued]

 

Models and Modelling - A Philosophical Deep Dive

Motivation

Not only in software architecture we use models for designing and documenting systems. Models are also indispensible in other engineering disciplines and in natural sciences. We all experienced good and bad models in our daily lifes. What is a model really about? And how does a good model look like? Let us enter a (philosophical) discussion about this topic.

What is a model?

A model captures the essence of a domain. It focuses on the core entities and the relationships within a domain from a specific viewpoint, i.e., serving a specific purpose. A model contains rules that must hold for its constituents. Models are used by humans or machines to communicate about the respective domain for a particular purpose.

Examples of models include:

a UML diagram

a street map

a floor plan

an electronic circuit diagram

a problem domain model (DDD)

quantum theory

mathematical formulas

Consequences: 

(i) The same domain can be represented using different models, each capturing another viewpoint of that domain. This viewpoints are often briefly called views.

(ii) Models can be informal or formal depending on their usage as a means for communication. Thus, they must be easily understandable and comprehensible by stakeholders.

(iii) Models introduce abstraction layers by using generalization and specialization leaving out „unnecessary“ respectively irrelevant details. 

(iv)  A model does not describe reality but a subset of reality viewed from a specific angle.

(v) Languages are based upon models. A model can be viewed as a language, and vice versa. 

(vi) A model may support a graphical presentation or a textual presentation, it even may include both.

The complexity of a model is directly proportional 

• to the number and types of its entities and their relationships,
• to the kinds and numbers of abstractions being used,
• to the complexity of its underlying rules.

A good model:

• provides a proper separation of concerns (SoC)
• consequently applies principles such as the single responsibility principle (SRP), Don’t-Repeat Yourself (DRY), KiSS, or the Liskov Substitution Principle (LSP) in order to gain the highest understandability and comprehensiveness
• uses expressive names for all its abstractions, entities, dependencies
• provides an effective and efficient means of communicating among stakeholders
• focuses on essence and leaves out everything that does not serve the required purpose of the addressed viewpoint
• avoids accidental complexity strictly and consequently 
• allows to model simple things in a simple way, while being capable of expressing complex things in a doable way

Stakeholders

The creation of a model should be guided by its (types of) stakeholders, in particular by the way they intend to use the model. In this context a meta model helps define how the set of creatable models should look like. Thus, meta models constitute modeling languages. They help create different models or views.

To define an adequate model that serves an intended purpose all (human) stakeholders should be involved. UML is an example of a modelling language that serves the needs of software engineers but (often) not those of many domain experts. In fact, domain experts might have their own models readily available. While a model might be perfect for machine-machine communication, it aint’t necessarily adequate whenever humans are involved. The more formal a model is, the easier it can be processed by computers. Humans often need more informal and expressive models instead. If both kinds of stakeholders are involved, we need to balance between formal and informal approaches. 

Emojis are an example of an informal model. They can be immediately understood by a human, but may be more difficult to process by a machine.

Artificial Neural Networks albeit “simple” can be processed by machines very well, but are hard to be understood by a human - i.e., with respect to what they actually do and how they work. 

UML is somewhere in the middle of these extremes. 

Fortunately in many mature domains, models already exist. An electrical circuit defines a proven concept of a model. Mathematics is often considered a uniquitous language with predefined notations. In the context of software engineering, domain models are often implicitly defined and have been established as common sense in an organization. If software engineers with no or little domain expertise start to develop software applications for the respective domain, they need to make the implicit model explicit. Otherwise they cannot design a software architecture that meets the customer requirements. This is what DDD (Domain-Driven Design) is all about. It tries to come up with a domain-specific model using generic building blocks such as DDD patterns and techniques.

The representation of a model should fit the needs of its stakeholders. For humans graphical notations often work very well, because they explicily reveal their structure in an easy manner and are good to grasp and to handle. Due  to productivity reasons, textual models may be more beneficial and flexible in some cases. As an example consider software code. For a beginner graphical code blocks might work very well, while advanced programmers prefer coding textually, because they can mentally map seamlessly between the “graphical” design and the textual code representation. Handling code graphically might just reduce their productivity, effectiveness and flexibility due to all clutter and constraints.

Model Transformations

To keep many stakeholders satisfied a possible approach might be to introduce different models for different types of stakeholders and also create mappings between these models, for example an easy to understand UML model that is transformed into a machine readable XML schema. 

Actually, software engineers are used to handle different models that are mapped onto each other. In software engineering a compiler represents a model transformation from a high level language to a system language or interpreter. A UML diagram might be transformed into high level language code. A low-code/no-code environment creates domain-specific applications from high level user specifications. However, model-to-model transformations can be quite complex, in particular when the gap between models is very large and if no common-off-the-shelve solutions for the transformations are available. Moreover, the more models the more transformations are necessary. Note: a model transformation might also be done manually if the model is not too complex and mapping rules are pretty straightforward.

Model sets

In domains such as building contruction or software engineering multiple views are necessary to represent information from different angles. Take design view, deployment view or runtime view as examples in the software engineering domain. In addition, their might be different model abstraction layers, for example, an in-depth design view versus a high level software architecture view.  In other words, to solve a task we need a model set instead of a single model that captures every detail from every perspective.

No matter how the views differ from each other, there needs to be meta information to tie the different views together. Prominent examples are the mapping from a view to code, and the implicit or explicit relation of views with each other. Note: there might be different solutions respectively model kits for the same problem context, e.g., RUP’s 4+1 view in contrast to TOGAF might not be the (only) solution of choice for designing an enterpise system. 

No matter what model set you choose, make sure that it is used consistently. In most cases tool support is strongly recommendable. Models can become very complex. Therefore you need a tool to draw, check and communicate the concrete models. This is the main reason why most software engineering activities rely on some sort of UML environment such as Enterprise Architect or MagicDraw. 

A ground plan is different from an electricity plan. All models together are necessary for building construction.  In this example, there might  also be rules and constraints across all models respectively views. For example, an electrical cable should have a minimum distance to a water pipe. Consequently, we need some kind of verification algorithm to check whether rules/constraints are violated. 

Model Creation

Models shall never be created in a big bang approach. They are living entities that change over time the more experience you obtain. They may start very simple but become more complex over time. Whenever they are overengineered, they need to be simplified/refactored again. Model creaters need to ensure that models can be facilitated and handled by stakeholders easily. If stakeholders have different viewpoints at the same problem, create a model set where each model view serves a particular set of stakeholders.

To start creating a model for a domain context, we should figure out whether such models already exist, and if this is the case, whether these models can serve the desired purpose(s). It is always beneficial to use existing models, in particular due to the experience and knowledge they carry. So, don‘t reinvent the wheel if not absolutely necessary, especially if you are no expert in the domain.

If no model exists, stakeholders should jointly create a model (set). It is helpful if at least one of the stakeholders is experienced in creating models while at least some other person is a domain expert.

If models exist that do not serve the intended purpose, we might change and adapt these models to fit our needs.

Note: a common mistake is to first focus on the syntax of a model. Instead, initially think about its semantics and find a good syntactical representation afterwards.  

No matter how a new model is created,  learning its representation should happen in a quick and straightforward process, even for unexperienced stakeholders.

Interestingly, most graphical models consist of rectangular or other symmetrical shapes, arrows, lines and textboxes, while textual models often use regular or context-free grammars. The reason for this observation is that this way the models are comprehensible and their handling is easy. It should also be possible to draw a model manually in order to discuss it with other stakeholders before documenting it.  Sitting around a modelling tool significantly decreases productivity, at least in my experience. A whiteboard or a flipboard is by far the best tool for modelling. This can be complimented by an AI software that recognizes manually drawn models and transforms them to clean and processable data representations. 

Summary

In this blog posting I did not reveal any new or innovative stuff you didn‘t already know. Neither was it my intent to provide anything revolutionary. It is just a summary of modelling and how to approach it. And if you started thinking about modelling from this more philosophical view, I‘d be happy. 

 

 Cohesion and Coupling

Sometimes what appears to be very simple and clear, is not that clear at all. A prominent example is the difference between cohesion and coupling.

• Cohesion indicates the relationships between components in a module. These relationships are not just syntactic sugar but have a semantic implication.  For example, if a module provides different separate functionalities it will reveal low cohesion. In constrast, modules with high cohesion implement functionalities that are tightly interweaved with each other.  In other words, modules with a low cohesion provide different things that are not related with each other. Thus, they violate the SRP principle (SRP = Single Responsibility Principle). The best refactoring in such cases is to split up the module into different modules that obey the SRP.
• Coupling defines the number of relationships between different modules. A module A that introduces a lot of relations to module B, strongly depends on module B, and vice versa. Strong respectively tight coupling of modules indicates that functionalities in B could be integrated in A. The obvious refactoring would integrate all functionalities of B into A, thus decreasing coupling of the system. Of course, merging of modules might reduce cohesion in the resulting module A. But what if another module A* also depends heavily on B. Should we then merge A and B as well as A* and B? Or should we merge A and B and reroute (A*, B) relations to A, thus possibly increasing coupling between A and A*? The best option is to strive for high cohesion and low coupling throughout our system and its constituting modules. Merging B into A and A* at the same time might violate the DRY principle (Don‘t repeat Yourself). But as we know, the DRY principle is not always appropriate. In Microservices architectures the diffenent microservices should hide their implementation details and do not share the same information but be independent of each other. Nonetheless, the principles of high cohesion and low coupling still apply.

Fortunately, architecture introspection tools can visualize cohesion and coupling and even provide hints and mechanisms how to restructure a given system. 

IoTDeepDive

IoTDeepDive - Infos zum OOP 2019 Ganztagstutorium Fr4

Diese Webseite ist auch erreichbar über folgende URL: https://tinyurl.com/oop2019-iotdeepdive

Hardwareanforderungen

1. Windows, Mac, Linux
2. Der Rechner benötigt WLAN-Fähigkeit über WPA/WPA2. WLAN-Verbindung stellt der Veranstalter bereit.
3. Ein externer USB-Port sollte vorhanden sein.
4. Zum Zugriff auf den seriellen Port mit dem auf dem Board verwendeten Silicon Labs USB-zu-UART, ist ein Treiber nötig. Bitte Version für eigenes Betriebssystem herunterladen und installieren!

Softwarevoraussetzungen

1. Im Tutorium kommt die Arduino IDE zum Einsatz. Download über https://www.arduino.cc/en/Main/Software und Installation gemäß der dortigen Beschreibung (abhängig vom Betriebssystem).
2. In der Arduino IDE muss ein zusätzlicher Boardsmanager für ESP32-Boards installiert werden. Für eine Beschreibung des Vorgehens siehe: http://esp32-server.de/
3. Adafruit-Bibliotheken in der Arduino IDE installieren, falls nicht über Library Manager verfügbar: 
4. Allgemeine Sensorbibliothek hier. Als Voraussetzung für die nachfolgenden Bibliotheken wichtig. Jede Bibliothek  per .ZIP Download holen (GitHub Clone or Download) und dann mittels Sketch>Include Library>Add.ZIP Library der IDE hinzufügen.
5. DHT 11: hier. 
6. BME680: Die Bibliotheken von Adafruit funktionieren auch für das Watterott-Breakout.  Zugehörige BME680-Bibliothek hier herunterladen. 

Source Code für Beispiele

Die Quelldateien für die Übungen finden Sie hier

Handouts

Bitte hier Handouts als PDF herunterladen.
Handouts in Farbe.

IoT Hardware Bezugsquellen

• ESP32 Entwicklungsboard im Makershop.
• BME680 Sensor (Wetter/Umwelt) Breakout-Board bei Watterott.
• DHT11 (Temperatur/Feuchtigkeit) bei exp-tech.
• DHT22 (genauerer Bruder des DHT11) bei exp-tech.
• Breadboard bei exp-tech. Besser, aber auch teurer, sind freilich Labor-Breadboards wie das von pollin.
• Jumper (Dupontkabel) zum Beispiel auf amazon.
• LEDs zum Beispiel als Sortiment auf amazon.
• Widerstandssortiment ebenfalls über amazon.
• Anschlusskabel Micro USB auch auf amazon.

Wer sparen will, versucht in China (ebay, Banggood, Alibaba) entsprechende Komponenten zu beziehen. Dauert of lange und ist eventuell mit einem Gang zum Zoll verbunden.

What the hell is Software Architecture

Since the Nineties many definitions of software architecture have been proposed, most of them being very vague. Eventually, the common denominator of these definitions suggests that software architecture denotes a set of cooperating components. In my opinion, this view is rather simplistic. Most entities in the universe follow the same definition which makes the definition useless from an engineering perspective.

Instead of providing yet another definition of software architecture we should think about its properties.

(i) Software Architecture is both a process and a thing: the process of architecting comprises a sequence of strategic decisions, while the thing is the result of this process. The goal of software architecture is to create the backbone for implementing the specification. Note, that we do not assume a specific software development paradigm here such as Lean or Agile Development.

Remark 1: Strategic decisions refer to requirements and additional forces that affect the whole architecture and result in design artifacts that are tightly coupled with the rest of the architecture which makes them difficult and expensive to change. Examples include mandatory functional properties of the system, operational qualities such as performance or security, infrastructures for modifiability such as a Plug-in Architecture, constraints caused by the system context such as hardware prerequisites.

Remark 2: All architecture design decisions must be driven by the specification and the business goals. Put briefly: No decision without a (good) reason.

(ii) Architecture design spans the whole lifecycle of a system not just its creation. It starts with initial planning and requirements engineering and ends when the system(s) built upon this architecture reach their end. Between these various points in time it covers creation, maintenance and evolution.

(iii) In a naive sense every software-intensive system reveals a software architecture, even if it has been created in a complete unsystematic or unintentional way, i.e., using ad-hoc decisions. What we need is systematic architecture design driven by well defined and prioritized architecturally-relevant requirements as well as risks. Sometimes, some or even all parts of an existing system with an unknown or partially known software architecture need to be (re-)used. In this case software archeology methods are required to help extract the hidden software architecture and make it explicit.

(iv) There are two kinds of architecture quality, external quality and internal quality. While the former one defines the externally visible behavior as demanded by quality attributes, the latter defines the habitability of architectural artifacts (such as simplicity or expressiveness) by developers, testers, etc.

Remark: a consequence of habitability is the limitation of software architecture design to a small number of hierarchical entities such as system, subsystem, components. This entities should follow the Single-Responsibility Principle. Accordingly, the responsibility of fine design and implementation is to refine and extend these abstractions to provide executable artifacts.

(v) For architecture as a process a consistent set of guidelines and tools shall guide the architecture design in order to ensure high quality and business alignment. Without such guidelines the architecture will be overloaded with multiple styles, idioms, patterns, concepts, technologies, paradigms, conventions, all of which reduces internal quality.

Remark: One major challenge is taming inherent complexity while avoiding accidental complexity. The latter one can be caused by using the wrong solutions, applying the right solutions incorrectly.

(vi) Software architecture as a thing is a means for communicating design decisions to other stakeholders. Thus, all decisions must be made explicit in a comprehensive way. The various constituents of the architecture shall be provided to readers in adequate ways depending on their roles and goals and their responsibilities and expectations. For this purpose, software architecture documentation must offer a consistent and complete set of architectural views.

Remark 1: Since software architecture is a thing and a process, its documentation includes the sequence of design decisions and their rationale which relates architecture views and process and which enables requirements and decision traceability.

Remark 2: Software architecture creates a base for design & implementation. It defines an initial (walking) skeleton for deriving the code base. This is why habitability is of foremost importance.

(vii) A software architecture is not an island but embedded into a context. Thus, it is important to strictly separate the architecture from its environment, while at the same time considering and defining the interfaces and interactions between both. Otherwise, it won't be possible to come up with an appropriate software architecture. A context view and use case views are examples that help address these aspects.

(viii) Software architecture must cover both problem domain and solution domain. For this reason, a Multi-Tier design is not an architecture. To avoid monolithic designs both domains should be hierarchically structured into subdomains and their relationsships. The organization of software architecture activities shall be driven by the aforementioned subdomains, not by the line organization. With other words, mind Conways law!

(ix) Software architecture is not necessarily constrained to a single system. It might also define the base for a set of systems within a specific problem domain. In such reuse contexts a Commonality/Variability analysis is required to define a common base architecture which will be modified for a particular implementation context before fine design and implementation start.
Examples: product lines, ecosystems, platforms/infrastructures, libraries.

Remark: This is what reference architecture and product-line architectures are about.

(x) Architecture design must follow a test-driven approach and communicated in such a way that the architecture can be tested. Testing provides relevant information to the architects such as revealing quality issues or design flaws. It includes quantitative and qualitative architecture reviews.

I can't resist. Let me conclude this posting with yet another definition of software architecture:

Software Architecture
is a process, i.e., a sequence of intentional strategic design decisions which map specification and business goals to architecture design.
is a thing, i.e., a set of views that address different stakeholders and are the results of the architecture process.

The main challenge for software architects is: there are different ways to design a good software architecture, but there are infinite ways to create a bad one.

Addendum July, 17th

Some may wonder why software architecture should be considered a process as well.

It is insufficient to obtain some architecture views, i.e. the What. In addition we might need information how the architecture has been created and why it is as it is, i.e. the How and the Why. This is not essential for some stakeholders such as users, but it provides relevant information to developers, customer service, and testers.

One notable example is the detection of design flaws. When we encounter a flaw in our system, we'd like to know when the design flaw has entered the "crime scene" and what other decisions depend on this flawed decision. This gives us traceability and the possibility to rollback in a systematic way. Likewise, process knowledge is important for architecture reviews.

A Matter of Waste

If a queue has a capacity of 100 units, we may enqueue 10 entities with a volume of 10 units per entity. 10 times 10 equals 100, right?

If we find a parking space with exactly the same length as our car, our search will come to an end - assuming the parking car must be parked parallel to the sidewalk. Right?

Hmmmh, the last one does not feel right. But what is the problem? It is that we need some additional space to maneuver. This extra space could be considered as waste, but it is in fact a precondition for parking operations. We may call this quality attribute "parkability".

A similar problem is the issue of a rectangular container with a volume of 100 ought to be filled with 100 spheres of volume 1 each. Obviously, we can't expect 100 spheres to fit into a container with the size of all spheres summed up. Again, the challenge is caused by inherent extra space needed by spheres.

Do we encounter such challenges in software design as well? Of course!
To be precise, we experience "waste" in two different areas: in our own activities as engineers and in our systems.

Let me start with the latter one that is primarily caused by technology and design constraints as well as by inherent properties of the problem domain. If you got this cool new 64-core machine with Terabytes of RAM, you may want to increase the efficiency of your apps by 64 times. Unfortunately, waste comes into our way again. Competition for resources and inherent dependencies in the problem context will cause the performance increase to be significantly lower than the desired factor of 64. And we did not even consider accidental complexity in this scenario.

Interestingly, it is such dynamic overhead that makes it difficult to deal with operational qualities or to build realistic simulators.

But how does this waste issue relate to software engineering activities? Basically, it is the same calculus again.

If we are committed to two different activities with overlapping time spans that need our attention, there is a chance of conflicts. Two ressources are competing for your time. As you cannot focus on two activities at the same time, you'll have to focus on one, which leads to a kind of time debt in the other one.

Another approach is switching back and forth between two activities. However, this requires the engineer to stay synchronized with the current state of each activity. Experts in multithreading call this "context switch". Time needed for such housekeeping duties inevitably causes waste.

Another typical root of waste is when a consultant or engineer of a company is supporting a custumer, because it is necessary to deal with two organizations such as organizational overhead for vacation, payments, ordering items, time accounting, meetings.

Often engineers do not consider such overhead, because they underestimate the problem instead of making waste explicit and finding ways to reduce it. But keep in mind: It is impossible to remove all waste, in particular inherent one.

Experience tells us that at least (!) one fifth of time allocated to an activity will be consumed for "non productive" issues. Thus, we always need to explicitly add a waste factor representing idle time.

Another reason for waste is communication. Think about reading mail, joining meetings, handling phone calls, chating with colleagues, drinking coffee. In a magazine I recently read that such interrupted work is a major issue in companies causing incredible amounts of overhead. This is due to the fact that when humans are experiencing a non maskable interrupt they'll need to reboot their brain before refocusing and continuing with the interrupted work.

So, if you need to reduce productivity of colleagues significantly play the interrupt game. Tom DeMarco once told me about two printer companies in the US where he could relate meeting time and productivity. The more (mandatory) meetings the less productive a company will be. Maybe, this pattern could be called "death by meetings". If you appreciate further patterns, start reading Dilbert by Scott Adams.



- Posted using BlogPress from my iPad

Technical Debt - The Downside of Metaphors

The term "debt" is a metaphor from economy. Its use for software development seems to be very reasonable. Whenever developers fail to address quality issues in their software, they will have to pay this debt back. That is quite simple, isn't it?

However, we can easily find weaknesses of the term "technical debt":

When a system reaches its end of lifecycle, all debt will be gone. Try this for financial debt.

Financial debt is a separate entity, i.e. two debts do not intertwine, while software debt sometimes can't be easily located, isolated or separated, because it is woven into the system.

Technical debt may stay unpaid, if software engineering comes to the conclusion that the cost for refactoring would be higher than keeping the debt untouched.

Financial debt is created intentionally. This is certainly true for some technical debt issues as well such as temporary workarounds. However, many kinds of technical debt are accidental without architects even recognizing it.

Technical debt is destructive. If possible, you would like all of it to be eliminated. In contrast, financial debt often is more constructive - such as increased cash flow.

Financial debt is independent of other system parts, while technical dept in one place can be affected by other parts, and vice versa. An example would be modification of the system that makes the code parts containing the technical debt obsolete.

Interest rates are mostly determined by banks and usually stay fixed over the specified time. In the context of design debt the interest rates increase the longer the debt is not treated. The pay back time is usually not predetermined. And the same kind of debt may be more expensive to pay back in one system part than in other parts.

While almost all forms of financial debt are paid back continuously until the debt plus the interest rates are fully covered, each single technical debt in most cases must be paid back at once.

There is one kind of debt in the financial domain, while technical debt has many different shapes.

If you think about it, you'll find other metaphors that are more convincing. For example, a tumor or infection has many similarities with what we call technical debt: It can occur accidentally, or intentionally if you don't take care of your health. It must be paid back at once - you are either ill or not, which isn't quite true in practice, but very close. The longer it is not treated, the worse the health condition will become - think of viruses that spread and damage other organs. Infections and technical debt are mostly destructive. Both can be monitored in imaging devices or labor diagnostics. There are tumors that are difficult to get rid of and that cause severe damage, while others are harmless if treated early. The worst ones can even become lethal. So my first guess would be to call it software infection instead of software debt.

Another possibility is to use environmental pollution as a metaphor. Like technical debt it may occur incidentally or (most of the time) accidentally. You must pay it back, but you can do this part by part. If left untreated, the situation worsens. It is purely destructive. Often, it can't be easily separated from the environment. Thus, another possibility would be to call it software pollution. Software Architecture Sustainability is a metaphor that is related to environmental issues. And so is the term "Software Ecosystem".

Other metaphors could be considered as well such as terms for similar issues in hardware or building architecture (consider design erosion).

My personal favourite is Software Infection (respectively technical infection, code infection, design infection). Of course, this metaphor has its own liabilities, but it is much closer to its cousin in software development than the debt metaphor is.

And it sounds appropriate to speak about a sick component or architecture. Or to visualize an infection using a modality like in medicine.

I know that it is too late to get rid of the expression "technical debt", but at least we should handle it with care. Some metaphors that sound good in the beginning may turn out to be not well aligned with what we like to visualize.

Location:Auenstraße,Munich,Germany

Are Patterns like Mummies?

In 1994, the Gang of Four, Erich Gamma, Richard Helm, Ralph Johnson, John Vlissides published their groundbreaking book on design patterns. At almost the same time we were creating the first POSA book (Pattern-Oriented Software Architecture) which was eventually published in 1996. The software engineering community got flooded by the Pattern Wave. And soon there arrived an inflation of pattern books, some of them excellent, but many of them of mediocre quality.

I remember, that almost every software development magazine addressed patterns regularly and enthusiastically in the Nineties. The GoF became the Beatles of architects and developers. And most experts anticipated a myriad of new patterns rising at the horizon.

Actually, several pattern books have been published until now, but without the impact the GoF had. If you ask software engineers about the patterns they know, almost all will mention GoF patterns, many may illustrate some POSA patterns as well, and only a few will come up with other patterns such as those in Martin Fowlers book on Enterprise patterns.

This could suggest, that no other important patterns can be found in software habitats anymore, because GoF already got them all. But even if this were the case for general purpose design patterns, shouldn't there be some excellent patterns lurking in more specific domains? Pattern experts have tried to come up with complete pattern languages for their domains. In theory, such pattern languages are beneficial. In practice, it is impossible to cover a medium to large-sized domain with a pattern language because of the inherent complexity involved. Thus, it is not surprising that existing languages cover only tiny domains or fail to completely cover larger domains.

Does the whole universe only know the GoF patterns and that's it? In this case the seminal GoF book would be the holy grail of software development. If not, where do all the unknown patterns hide?

Let us assume we are asked to improve the GoF book, what exactly would we change. Patterns like the Null Object Pattern or the Extension Interface pattern (see POSA Vol. 2) could be added, the Singleton pattern removed and other patterns such as Abstract Factory improved. Patterns are not carved in stone but subject to future changes, albeit not with a high evolution speed.

If you look at patterns from 30000 feet, you'll recognize that there are some benefits of patterns not related to coding.

Pattern forms are a good mental tool to document design. For example, they comprise a name, a context, a problem with forces, and a solution. This helps document all kinds of architecture decisions in a structured way

Patterns are also applicable to document best practices for transformations, data representations, and many other topics. For example, each refactoring can be considered a transformation pattern. Best Practices are ideal candidates for patterns

Patterns introduce an idiomatic viewpoint, as they define a language. Effective usage of software platforms in terms of best practices for frameworks, libraries, APIs and protocols are idiomatic as well. Experience shows that idiomatic approaches help better understand structures and concepts. In software engineering all structures are idiomatic. Languages change and so do Patterns.

The value of patterns is not their content, but also their usage as good mental tools, with the capability of addressing all activities. Patterns help share best practices with others. They unfold a language of idioms that provide understandability, maintainability.

Patterns are dead. Long live patterns.




- Posted using BlogPress from my iPad

Location:Auenstraße,Munich,Germany

Architecture-as-a-Service

Recently, I discussed architecture design with some colleagues who are involved in Product Line Engineering projects. When the term "Reference Architecture" came up, it became obvious after a while that they used the term specifically for Product Line Architecture, while I had Reference Architectures in mind, which are typically created to steer and guide standardization. They might also present architecture styles that are commonly accepted within a domain. Think of compilers which are in most cases structured as a Pipes & Filters architecture which is composed of a lexer, a parser, a semantic evaluator, an optimizer, and a code generator.

A Reference Architecture (RA), on the other hand, is a coarse grained architecture template different organizations use for guiding their design activities. You may remember the OSI 7 Layers model or the OMG Reference Architecture for CORBA, both invented in the Stone Age of software engineering. A RA is a blueprint which doesn't include any further assets but documents. In fact, it is not very specific but rather abstract.

A Product Line Architecture (PLA) is an architecture for the product line of an organization. It captures the commonalities of a set of similar products and defines variation points using feature models or meta models. It contains different core assets, some of which are ready for use. A PLA is much more specific than a RA and defines a framework with (partially) implemented artifacts and variation mechanisms.

An RA can be used as the core of a PLA. For this purpose, the RA is concretized in Domain Engineering by addressing requirements and constraints of an organization. In the process, engineers may provide additional variation points or bind existing variation points. In the latter case the PLA transforms a variation point to a commonality. If many organizations do this for the same variation point in similar ways, a new extended RA is born.

A PLA can become the base of a RA if it is subject to abstraction and considered as an established practice in the domain, i.e. if most PLAs will end up in the same RA.


- Posted using BlogPress from my iPad

Location:Wotanstraße,Munich,Germany

I am back

After a long time absence due to health issues I am finally back. I will add new content in the following months with the publication frequency increasing over time.

At the OOP 2015 conference which ended last friday I was in charge of the architecture track. In my own talks I covered internal & external quality as welll as ways to ruin a project by (wrong) architecting. The latter tutorial introduces failure patterns for harming software development from the perspective of architects. Unfortunately, there are uncountable ways for failure, but only few for success.

A software system may be considered as a living organism. If the organism is not robust enough, even small illnesses may cause large damage. Architecture is basically the infrastructure necessary for metabolism, neural transmission or structural integrity (such as the bones). If parts are damaged, e.g., organs, we may either put in new or artificial organs, or maybe only repair a small part.
The software organism is created in an evolutionary process with biological patterns and is subject to design erosion during its whole lifecycle. Obviously, it is important to check the organism regularly to identify potential health problems as soon as possible. With Architecture Tomographs these checks can be fast and intense. In some cases design erosion or cancer dissemination is high enough to make the system impossible to repair. In these situations we may better create a new organism. However, for less harmful kinds of design erosion we may refactor using patterns. In this context, external quality is every property we can observe from outside such as speed, reliability, ... Internal qualities comprise heart rate, blood pressure, insuline level, bone strength, ....

Of course, the model has its limitations, but for me it turned out to be really useful.


- Posted using BlogPress from my iPad

Micro Management for Micro Brains - why Micro Services suck

It sounds like a great idea. Instead of building a monolithic enterprise application, just split functionality into small components and run them independently in their own processes. These micro services need to cooperate with each other to provide more advanced functionality. For that purpose, they are going to communicate with other micro services.



Such an approach promises better flexibility in changing and deploying systems. If only a small part of the functionality is changed, this will affect only a small number of micro services. Or, as we like to say, "small is beautiful".



What a wonderful new architecture style. Eventually, we found the silver bullet. Agreed, micro services may not be a silver bullet, but, at least, they are silver shotgun shells.



This idea is not new, though. For example,  "actors" are providing the same kind of solution. They encapsulate fine grained services behind message-based interfaces that enforce argument and result types to be value types, so that no side effects may occur. In order to achieve a common goal, actors interact with each other.



"Micro services" contains the term "services" for a reason. Micro services basically introduce a stripped down resurrection of SOA.



However,  the delicious looking and tempting apple may be poisoned. There are several "challenges" when using this architecture style:

• Complexity:  If we are building a non-trivial application, it will reveal inherent complexity.  For the moment, let us assume, there is no accidental complexity involved. When this application is based on micro services, where does the complexity go? Unfortunately, it does not go away, but manifests itself  in complex connection and cooperation patterns between tiny services. 


• Modifiability: If our enterprise application is partitioned into small micro services, we only need to touch and redeploy small services for changing the system, instead of facing the mess of application monoliths. In theory, this is a perfect solution. In practice, it is not, because for non-trivial changes of micro services, we may need to rewrite  a lot of other micro services that are connected with the modified micro services. Remember, the complexity does not disappear but shows up in the topology of the micro services network. This does not only affect design but also evolution, assessment and refactoring activities.


• Internal  Architecture Quality: Architecture is about strategic design. Its main purpose is to map the problem domain to the solution domain. If micro services are used as the primary concept for structuring functionality, the problem domain will be mixed up with the solution domain, especially, if the usage of micro services is not transparent. Thus, the architecture will very likely be technology-based instead of problem-based. Note, that this is another variant of Conway's Law: "show me your architecture and I will know the technologies it depends on".


• External Architecture Quality: As we all have learned the painful way, main reason for architecture failure typically are not functionality aspects.  Main reason are quality attributes like performance, extensibility, fault tolerance, and so forth. For each of these quality attributes, well known design strategies and design tactics are available that present alternative solutions how to introduce the respective quality into the architecture. Architecture design uses utility trees and scenario diagrams to specify external quality requirements. These quality attributes are often crosscutting, which is why complex networks of services make it hard and sometimes even impossible to design and implement  quality attributes.


• Infrastructure: We may use a technology stack for leveraging the micro services architecture pattern or we may build it ourselves. The latter option is a no go, because we are not in the middleware business, at least most of us aren't.