A first example of rules - introducing Archichect

It's high time for a few examples. I'll draw them from two programs:

• First, the medium-sized "Archichect" tool that I wrote for checking and exploration purposes (with substantial input some years ago from Thomas Freudenberg; and requests and ideas by colleagues at Pharmatechnik in Germany, my employer). Archichect is right now a proof-of-concept software, which means, among other things, that I change anything quite freely on a daily basis (if you nevertheless want to take a look, you can search it on GitHub and look into it).
• Second, the large flagship IXOS product of our company, which I will use later to demonstrate large-scale architectural aspects.

Let me start with a small example from the prescriptive architecture of Archichect. Here is a rough sketch of some part of its intended (therefore, prescriptive) architecture:

Obviously, this is a "type 1 sketch": It is a mixture of definitive information ("there are three subpackages called Reading, Transforming, and Rendering") and illustrative, but incomplete information ("There is a class called DependencyChecker implementing interface ITransformer, but probably there are more implementors of this interface"). There is nothing at all wrong with such diagrams—except that one must be extremely careful to draw conclusions from them: The "existential assumption" is usually ok ("All the things in the diagram will be there in reality"), but even this is sometimes risky ("oh, I just meant that as an example").

So let us write some rules for this part of Archichect that lift the information from the diagram to a "type 3 declaration". Here they are, in Archichect syntax:

$ DOTNETITEM ---> DOTNETITEM

(**)                    ---> ::\1
Archichect.Reading      ---> ::Archichect
Archichect.Transforming ---> ::Archichect
Archichect.Rendering    ---> ::Archichect

(You are not happy with the notation?—you would like to express this in code? Well, as I said, I am not too stubborn about notation, so I will show how to write the same rules with code in one of the next postings).

One can actually run these rules over Archichect itself and get a nice and very long result telling us that Archichect (in its current version) has 35505 dependencies, 26690 of which violate the rules above: "Sad", as some well-known guy would tweet. Obviously, transferring an architectural diagram to strict rules requires a little more work than just more or less faithfully copy it to text—but it's only a little more, I promise (and will show you).

Before that, however, let me explain the rules above, and the assumptions behind them, a little more.

First of all, the diagram did not spell out how the concept of a UML package is mapped to the language. Most modern languages have at least one concept of nestable groups for naming things; for example, Java has packages, and the .Net languages have namespaces. In addition to these logical nestable constructs, the physical units of runtime environments, e.g. JAR files or .NET assemblies, are typically named with filenames, which can again use a hierarchical naming concept. For example, in .Net, there are assemblies named System.dll, System.Threading.dll, System.Threading.Tasks.dll, System.Threading.Tasks.Dataflow and System.Threading.Tasks.Parallel.dll etc.

In Archichect, I chose the standard approach of mapping packages to the naming concept of the implementation language, i.e., to .Net's (and C#'s) namespaces. Thus, I would put

• Program, Item, Dependency as well as the three base interfaces into namespace Archichect;
• DotNetAssemblyReader in namespace Archichect.Reading,
• DependencyChecker in namespace Archichect.Transforming, and
• ViolationsWriter in namespace Archichect.Rendering.

The type DOTNETITEM defined by Archichect's DotNetAssemblyReader nicely defines the fields Namespace, Class, Assembly.Name, Assembly.Version, Assembly.Culture, and Member.Name, and therefore the rules above refer to the Namespace field without any further syntactical ado.

Still, an alternative architecture decision could be to distribute the classes and interfaces into multiple assemblies, and then, the rules would have to be written differently, e.g. (there are more possibilities) as follows:

$ DOTNETITEM ---> DOTNETITEM

::(**)                    ---> ::\1
::Archichect.Reading      ---> ::Archichect
::Archichect.Transforming ---> ::Archichect
::Archichect.Rendering    ---> ::Archichect

The two colons indicate that the strings are assumed to refer to the third field of DOTNETITEM, i.e., the Assembly.Name field. Assembly rules collect, at least in .Net, also important architectural information and should therefore be part of the "rule set"; but I will ignore them for the moment and continue with namespace-based rules.

Secondly, regarding the diagram, it does not say anything about any dependency restrictions inside each package. A typical implicit assumption is that, on this level of granularity, each item inside a package may use any other item in the same package. To allow such dependencies, the rule

(**) ---> \1

is added. The \1 notation here is borrowed from regular expressions (and actually, internally, the rule checking is mostly done by creating regexes from the rules and then matching the read-in dependencies against them).
This Archichect rule says that items from some namespace can use items from exactly the same namespace, but not from a child or parent namespace. We will see in a later posting that this would, in some cases, prevent the useful organization of namespaces, and therefore the actual architecture rules of Archichect are somewhat different. For the moment, we leave it at that.

Still, we should understand why our apparently so useful rules gave us so many invalid dependencies, shouldn't we? That's some stuff for another posting.

Prescriptive, descriptive and experimenal architectures - what you want, what you have, what you like

In the previous posting, I described a frugal model for architectural descriptions; and ended with the question for which purposes it can be used. This should have been followed by examples—but I decided that before that, I want to explain a very important high-level view concerning architectural descriptions.

The software we have at a concrete point of time during its evolution is often not the software we want. That is true for external attributes, i.e. features and non-functional qualities, but it is also true for the many architectural aspects of a software or system. From this alone it follows that we have to deal with two different sets of architectural descriptions:

• The prescriptive architecture is the set of rules that we want some software to follow. Typical prescriptive constraints are "the GUI models must not access the database directly", "all event handlers must be aynchronous", "there must be no cycles in the dependencies of modules of type X" and the like.
• The descriptive architecture is the set of constraints that is actually adhered to in the software. These constraints are typically much more muddy than the prescriptive ones—"the generated GUI models do not access that database, but in the startup module, a model directly reads configuration from the database, and for some plugin modules, we actually do not know whether they access the database or not" might be an honest description of some set of dependencies.

In a well-functioning architectural process,

• the relevant aspects of the prescriptive architecture are known and documented unambiguously;
• also the corresponding aspects of the descriptive architecture are routinely extracted from the software and documented;
• and both are compared to detect when the latter diverges from the former at some critical point.

Maybe it should be noted that there are two categories of reasons why the prescriptive and the descriptive architecture might differ: On the one hand, the prescriptive architecture might be stable, but for whatever reasons actual development does not follow it. On the other hand, the prescriptive architecture might change, because architecture-driving requirements change. This should only highlight that it is not necessarily someone's "fault" if the two do not match.

Are we done with the architectural process?

No, we are not: When a critical difference emerges, something has to happen to align them. This might be a change to the prescriptive architecture, or a change to the actual software (which changes the descriptive architecture), or changes to both of them. In many cases, this alignment will be painful. After all, both the reasons why the prescriptive architecture is as it is, and the reasons why the actual software is as it is are profoundly embedded in the requirements and processes and people building the software. In almost all cases I have seen or taken part, planning for the reconciliation of "what we want" and "what we have" was hard and frustrating. Typically, the consequences of such a "re-architecting effort" were, and are, not at all clear for two very important factors:

• How much would the modification cost? The potentially recursive ripple effects that one change creates could lead to a nightmare of subsequent changes, and that prospect alone often considerably reduces the possibility to get the "funding" for such a change.
• How much benefit would the modification yield? Aligning the software with some prescriptive architecture may sound great, but there are typically good (but maybe not well-understood) reasons why the software is as it is; and so "following the rules" might actually make the software worse. The same is also true in the opposite direction: Just changing the prescriptive architecture to "what we have" may result in a set of "rules" that is so large and chaotic that following them is practically impossible.

One important reason for these uncertainties is that we mostly approach such alignment tasks with only two tools:

• "Dive in": This works by direct modification of the software (where the goal is to keep the prescriptive architecture) or the prescriptive documents (when the software, or some aspects of it, should be kept, but the prescriptive architecture should change). For changes in the software, this is typically (and hopefully) done in a feature branch to shield the productive software from modifications whose adverse and potentially fatal consequences are seen only later in the modification enterprise.
  For changes to the prescriptive architecture, the same should be true—which requires versioned handling of architectural documents, including "branching" and "merging". As I understand it, current tools and notations are not well-prepared for this—I would shudder to find out what an automatic merge of two UML diagrams might produce. But maybe I am too feeble here.
  In sum, irrelevant of which side has to change, "dive in" is an expensive undertaking.
• "Panorama": This approach works using informal knowledge and notations that try to capture only the essential aspects and consequences of modification variants. From these, often shaky grounds, decisions on how to proceed are derived. Typically, these are very conservative, and often limit themselves to "pilot projects" or some "drill-downs" which are supposed to be fed into another loop of the modification process.
  In sum, "panorama" often requires many iterations to get a useful result, and is therefore also arbitrarily expensive.

Both the "dive in" and the "panorama" approaches are valuable tools. However, they seem to work only with changes of a limited size. For larger systems and changes, their "sort-of-quadratic effort" (practically try out a subset of all interaction pairs between any two components) limits their usefulness and possibilities.

Thus, there should be a third possibility, namely to simulate changes on an abstract representation of the software. I think we would like to do something like the following:

If we move all the controllers into a new package, then we could separately unit-test them. Let's do it ...
... Ah, but now we see that some controllers have dependency loops with their models, and others dont. We do not want a rule 'allow loops between models and controllers'; but cleaning up all these loops right now is no option.
... But wait, it seems that the loops are mainly on trivial controllers that do not have service dependencies; whereas the controllers on top of services are typically cycle-clean ... so let's make two groups, the 'simple controllers' and the 'service controllers'. Ok, we do it ...
... and now there are only two service controllers with loops. But could we carve out their 'looping' code into a simple controller—let's try it ...
... Ok, so we can agree on new rules for controllers: 'Simple controllers must not access services, but can have looping dependencies with their models' and 'Service controllers may access services, but must not have loops with their models'.
etc.etc.

Such an exploration into the possible changes and their consequences obviously needs to be done on a model of the architecture—a model that can be trusted, i.e., mimics the actual software; is simple, i.e. can be intuitively understood by at least the architects; can be efficiently handled by tools; and what not else (I'll have to come back to these properties later, won't I?).
To distinguish this model from the two introduced at the very beginning, I call it the (or rather, an) experimental architecture.

So, in a nice(r?) world, we end up with three architectural models of a software:

• The prescriptive architecture—what we want (or believe that we want).
• The descriptive architecture—what we have (for some interesting abstractions).
• Experimental architectures—where we try what we might want and have.

But don't we have all these right now, you might ask? After all, the first is in some documents and the minds of all architects and developers, the second is "in the software", the third is on flipcharts during discussions about architecture. Of course, that's true: But I would argue that we need to be able to move information—and this will be loads of information; remember the "telephone directory property"!—between these models, and move it reliably and quickly. That's why a common notation and tooling for all three would appear worthwhile, at least to me.

To coin another term, I will call this the "escalator property" of languages for architectural descriptions: Namely that they can be "escalated" from the descriptive (what we have) to the prescriptive (what we should have) to the experimental (what we might have). And just as an escalator can also go downwards, it should be possible to map the results of an architectural experiment easily to a prescriptive architecture, i.e., to an enhanced set of rules for the system under consideration.

Examples are now really what is needed!

Models, notations, and languages

In previous postings, I used the terms "notation" and "language" too sloppily. Here is a short explanation how I intend to use them in subsequent postings—I hope that this is in line with common usage:

• Whenever we want to work with some real world things, we need a model of it. The model is a more or less rigid (mathematical) abstraction of the object(s) under consideration (which are called the "universe"). For example, a real CPU might be modelled via an abstract processor, which only considers its assembly-level commands, but not e.g. its heat emission. A model of an SQL database might only consider tables, columns and views, but "abstract away" triggers, stored procedures and everything else the vendor might have added as a feature.
• A notation is some sort of symbols that adhere to some syntax. Many notations are linear text notations (all programming languages I know of), but there are graphical notations like UML's notation.
• A language is a combination of a model and a notation, where the notation is mapped to the model or modifications of it.

The last definition implies that one can have many languages for the same model. Here is a simple example of this: Let our model (and also universe) be expressions of integral numbers, with e.g. subtraction, multiplication, and evaluation. Three possible notations are

• parenthesized infix expressions, e.g. "(5 – 3) * (5 – 2)"
• postfix expressions, e.g. "5 3 – 5 2 – *"
• and a tree notation that shows the expression tree.

All three (and many other) notations can be mapped to the model in a way that they "compute the result" of an expression correctly.

Having a good notation is important, but getting notation right is better done by a series of experiments with real people than a conceptual process. I will, therefore, not be very stubborn about notations for architectural problems. On the other hand, I will try to find a single small representation and manipulation model for architectural problems; and then try to argue that the chosen model is sufficient and practical and, well, good.

In spite of my laissez-faire approach to notation, I do hold a few beliefs about notation that I will try to argue more or less emphatically.

The most important is that any notation must scale to large descriptions. Thus, it must be possible to describe, in a manageable and legible way, a system that consists of a "flat 1000 different parts". By "flat 1000 different parts", I mean that the notation must not force the writer and the reader to introduce any sort of abstractions solely because the notation becomes unwieldy. I call this the "telephone directory property": A useful notation must be capable of practically notating a large, boring list of slightly different things "just so".

As a special case, I will not consider any diagrammatic notations for the moment (later, I'll come back to diagrams). For almost the complete history of software engineering, people—intelligent people—have tried to come up with a graphical replacement for formal textual languages like programming languages. There is a complete theory and much practical experience with two-dimensional diagram languages—but on the whole, they have never replaced textual languages in anything but small, and often not-too critical, software systems. The reason is exactly that diagram notations do not have the "telephone directory property"—diagrams describing a 1000-part system are, for all purposes, unusable: They cannot be viewed easily (especially if they contain longer, winding line paths), cannot be printed easily, and they cannot be manipulated easily. The morale: Designing and maintaining diagrams that are not useless from the outset is very hard.

(If you think "UML", and especially "UML according to all the software architecture textbooks out there", I remind you that my focus is not the use of diagrams for informal or "semi-formal"—whatever that means—purposes. For this, many diagram notations are perfectly fine. But I consider only what I called "use case no.3", i.e., languages for describing and maintaining architectures that have a strict semantics that can be used to prove or maintain something interesting in a software system).

There are a few more aspects—important aspects— that will influence all the many parts I want to assemble for useful "rule-based architecturing", but in order to keep the suspense low, I will now immediately give away what my proposed model is: (Finite) directed graphs with labelled edges and nodes. I will not restrict this quite general model much more, except that I have to define the allowed labels. They are:

• A node is identified by a label that is a tuple of strings. In addition, a node can have informational tags, each of which consist of a name and a real number.
• An edge is only identified by the nodes at its ends. It has three counts that are called the overall count, the questionable count, and the bad count. In addition, it can also have name+number tags like nodes.
• Both nodes and edges can have an arbitrary source information that is intended to find the object from which the node or edge were derived at some time.

For "historical reasons" (we invented the basics of this model some 10 years ago), I use the following terms:

• Nodes are called items.
• Edges are called dependencies, and the two items at the end of a dependency are called the using item and the used item.
• The identifying strings of items are called values, and the non-identifying tags of items and dependencies are called markers.

I hope that these terms do confer a rough meaning of the purposes for which they are used.

But—for which purposes are they used?

What made xunit testing successful?

The xunit revolution introduced

• a very simple notation (actually, two notations);
• a reasonable benefit for every developer;
• and, later, a culture that extended "mere xunit testing" to various "development philosophies" like TDD, TDD with baby steps, or BDD.

The notations have a set of important properties:

1. They define a small language of a few important concepts:
   • At the core, only testcases that run in a predefined test harness framework; and—almost unrelated to that framework—assertions;
   • for scalability, testfixtures and setup and teardown of test cases and fixtures.
2. The building blocks are very small: A single assertion is atomic; a single testcase can also be made atomic (i.e. just test a very tiny segment of the intended behavior).
3. There is a simple tool that efficiently does the mundane job of collecting and executing all notated items (test fixtures and test cases).
4. The tool can be easily run by any developer at any time.
5. The tool can also be easily integrated into existing automated build processes.
6. And, finally, the automatic execution can have a drastic feedback on the processes: Tests that do not pass halt the delivery process (by resulting in a "red" build).

The direct benefit for the developer is not that more quality assurance can be done during code development—even though later "xunit philosophies" are, one could argue, roughly founded on this belief (and delivered arguably better processes for direct support of development). On the contrary, more quality assurance (in the sense of "trying to find destructive input to check a program against the limits of a specification") during development would actually be an annoyance, because it disrupts the developer's constructive thought processes necessary for constructing code.

Rather, xunit testing helps to solve the problem of "later regression checks" occurring after code changes, when it is necessary to remember and run the simple as well as the tricky test cases that actually allow a developer (or a team) to hold the belief that the modified piece of code still behaves sanely.

The important experience is that that "later" is not only "much later", when a feature upgrade or bug fix requires changing the code, but that it can be right after the next (well or not so well thought out) modification during the initial development of some piece of code. That really helps developers.

Finally, xunit testing is open in multiple ways—how many tests one writes, how much behavior each one ascertains, when they are run in the development cycle, and when in the build cycle, and, last but not least, how writing and executing of xunit tests feeds back into design and code development. Because all this is not enforced by the tooling in any way, a host of "philosophies" could emerge on top of xunit testing, leading to a lively and sometimes heated debate with a huge effect on wide understanding and on "marketing" of xunit testing.

Great.

Could the same be accomplished for some parts of "architecting"?

We should try, at least, shouldn't we?

So, you and I and everyone should start to invent notations and tools for "architecting" along the lines of what made unit testing successful. I'll leave your ideas to you; in the next posting, I'll start to present mine.

A role model: The (x)unit testing revolution

In order to overcome waterfall and moralistic approaches to software architecture documentation and process, one should take a look at other areas that succeeded in establishing new ways of doing things. One paradigmatic example is xunit testing.

Martin Fowler argues that what I describe here should not be called "unit testing", but "xunit testing". I follow his advice.

Two decades ago, testing had some problems that were, on a high level, similar to those of software architecture processes and documentation:

1. Scope—what do you test, where do you stop?
2. Documentation—how to make test cases permanently available, e.g. for regression testing?
3. Planning—when and how long do you test?
4. Responsibility—who does the testing?
5. Maintainence—how to upgrade existing artefacts, i.e. test cases with their input and output? 
6. Automation of testing—how can the repetetive parts of testing be done by software?
7. And finally, the conceptual question—how much of quality assurance should a testing philosophy encompass?

Before the "xunit testing watershed", the answers were:

1. The boundaries of what is tested, and where you stop, are ill-defined and arbitrary.
2. Documentation is done by natural language or slightly formalized natural language. Formal approaches, using their specific "testing languages", are research topics and typically way overboard for almost all projects.
3. Planning is mostly done "at the end", i.e. based on a waterfall view of software development (which is wrong), or rather, "between devlopment and delivery". This carries an extreme risk that testing is squashed between development overruns and promised delivery dates.
4. The reponsibility for testing lies with "others"—not developers, but "testers". Therefore, there is no continuous process, with conseqeuent small granularity, between development and testing, but a "break" with corresponding hurdles in communications and planning.
5. Maintenance is a heroic or bureaucratic effort that either fails or is expensive to keep up.
6. Automation is done by external "test harnesses", for example "automated GUI testing tools". This is quite expensive and produces brittle test code.
7. The testing philosophy—which has to be implemented by a testing process—should encompass as many quality assurance aspects as possible (consider the alternative:

   Test lead: "Our testing process only deals with functional expectations, but not non-functional ones."
   Manager: "Do you mean we need yet another organization to do the non-functional testing? Do you know how much your non-productive department already costs us?!?")

The net effect was that testing was essentially either a "bureaucratic" or a "moral enterprise": Either one had set up a separate organization to do the testing; or it was left to the morale of single people who—depending on which book they had read; or which catastrophe had occurred the day before at a customer—would either find testing the most important and underrated activity in the world; or an obstacle to development and shipping at that last possible minute.

The xunit testing movement (which was initiated by Kent Beck with his SUnit tool for Smalltalk) gives, fundamentally, answers that are exactly the opposite of the above:

1. What you test are small units that are easy to handle.
2. The test cases are code.
3. and 4. Tests are run during development by the developers, with an arbitrarily small granularity.
5. Maintenance is just like code maintenance.
6. Automation is inherent.
7. The goal is to cover only some software quality features, mostly correctness aspects. Other test topics like usability or acceptance tests are not targeted by the method.

Because of the last item, xunit testing cannot replace manual and other testing techniques—which was, of course, in the "old philosophy", an argument dealt against it. But over time, this argument eventually vanished.

In sum: Xunit testing introduced a striking new alternative on a subset of the testing problem.

Software architecture in real projects is, it seems to me, in the same situation as testing was twenty years ago:

1. The boundaries of what is to be described by a software architecture are ill-defined and arbitrary.
2. The description is, by the rule, done by informal text and "diagrams"—which are interpreted informally even if they use e.g. UML notation. There are formal approaches, but they are way overboard for almost all projects.
3. The documentation is done "at the beginning", "before design sets in", i.e. with the implicit assumption of a waterfall process.
4. The responsibility lies with "others"—not developers, but "architects"; with a resulting break in the process.
5. Maintenance of documentation and architectural rules is a heroic or bureaucratic effort that either fails or is expensive to keep up.
6. Automation?—there is no automation for architecture.
7. Architecture, by definition, covers everything—sometimes limited to "everything that is important", but that does not really exclude anything.

It is obvious, from that list and from the xunit testing experience, that something can be done.

Two common, and defective, approaches in software architecture

Let me deviate—or actually, approach my target from a different angle—for two more postings before presenting one such "mundane notation" for software architecture documentation (which I have promised in my last posting).

What are the main problems with current (explicit) approaches to software architecture? Very briefly, they might be dubbed

• the "waterfall approach"; and
• the "moralistic approach".

The first one, "waterfall thinking", is the old idea that one "starts" with deciding on basic and important architectural aspects, and "then" goes on to design and write software accordingly. Some parts of software engineering might follow this pattern, but there are at least two major scenarios—or maybe forces—that lead to a different process:

• One is the fact that in almost all cases, a huge software is already in place; and the architectural problem is to modify this software "from inside out". This can and is often be done by small exploratory "experiments" in the software that prove or disprove whether some concept might be worthwhile. And in many cases, this is done implicitly and "under the hood", when some developer starts, on his or her own initiative, to introduce the first RESTful service, a "small NoSQL database on the side", or reuses some executable for production purposes that originally started out as a tool for developers only.
• The second scenario is brought on us by the typically vast capabilities of commercial and open-source frameworks or tools. When you buy SQL server instead of using Postgres (maybe for external reasons, like having a partner status with Microsoft); or when you take Angular instead of some lesser-known JS framework because some graphics library ties in better with it, you also "buy into" a huge feature set that comes with that tool. Your architectural possibilities are suddenly, and at the same time, extended by the tool's many for-free features, and also limited by the grand architectural and technological lines of it. And like your software, such a toolset is often "just there", without any possibility or even wish to ponder any underlying architectural requirements and decisions.

It is by no means clear that bottom-up approaches, as done by the hypothetical developer above, aren't on par or even better than processes that proceed from "grand architectural analyses" "down" to design and implementation. And, in real life, such bottom-up situations are unavoidable anyway. Thus, "waterfall thinking", while certainly an option, should not be the only and preconceived approach to architectural decisions.

The second problem is the "moralistic approach" to architecture (and design). Architecture and design decisions produce rules: "In our system, code on the GUI layer must access the database via an intermediate DAO layer"—or the other way round; "plugin registration happens explicitly by adding an entry to the configuration, and not implicitly by merely placing the plugin at some location"—or not. And somehow, such rules must be enforced. Most of the time, there are only two enforcement regimes in place:

• One is the "build-and-install-regime": The build and, later, installation processes of a software require that certain rules are followed. These rules are often implicit, but at least it is hard to violate them. It is also often very hard to change them.
• The other regime is the "moralistic one"—"you should", or "you must": Without support from tools, it is assumed that developers have the capabilities to follow the current rules. When, later, some disaster happens, one can more or less easily find a person that is the "culprit": "You shouldn't have added that trigger that implicitly calls itself and then fills up the audit table!", "You should not have hardcoded that connection string, but taken it from that (faraway) configuration file to keep database accesses consistent!" But of course, people will only adhere to some sub-100% percentage of rules—and this assumes that the rules are explicitly documented and consistent to begin with. And also of course, culprit-finding does not solve problems well (it might, in some cases, prevent others from violating the same rules in the near future). And finally, we are all versed in putting the fault on the shoulders of the ultimate culprit: "This has grown historically."

Both the "waterfall approach" and the "moralistic approach" are wrong in their fundamentals. But just by saying so, there is no positive alternative in place that replaces them. And, to tread somewhat more carefully, one should certainly not throw out top-down approaches (of which "waterfall" is a special case) and rules-of-thumb (an essentially "human-compatible" method for solving problems, just like "moral") from the portfolio of process building blocks for "doing software architecture": These are worthwhile at the, well, right places.

But some alternative view on "doing it" should be possible.

Purposes of architectural documentation disentangled

I have been a little unfair in my last posting: The eight pages on UML 2.0 in Gorton's "Essential Software Architecture" are more than a mere advertisement for that (then) new UML version 2.0—they do actually contain some core advice about how to document architectural aspects of a program. I'll try to extract a compact view of what architecture documentation is, in Gorton's and, I think, the mainstream architecture's textbooks' view, from these pages and the case study in chapter 7.

First of all, architecture documentation is a collection of artifacts for human beings only. This is in contrast to code, which is targeted both at the "machine" and at human readers. In the background, there looms the idea of model-driven architecture, where an architecture model is used to create code—essentially, a compiler for a new language on some "higher" level than standard programming languages. However, like the book, I will disregard this aspect right now and return to it somewhat later.

The clear target of providing information to humans has lead most of us to the use of informal diagrams and standard prose to describe the architectural aspects of a software—"simple box-and-arrow diagrams", as Gorton calls them. He claims that there is "an appropriate diagram key to give a clear meaning to the notation used" in his examples, but most diagrams in his chapters 1 to 5 don't have such a key, and in any case, most people drawing such diagrams don't include one. The problem with this is that any plan to derive hard facts from such diagrams is then doomed.

Now, one purpose of architecture documentation is to give someone a "feeling of the interplay of things", and for this purpose, informal diagrams with textual or oral explanations are perfectly fine and, I am quite sure, even preferable: They appeal to our intuitive approach to most problems, which includes working with somewhat unclear terms and their relations in order to limit thinking about tricky consequences, so that our mind is free to "suck in the universe" of the problem area at hand.

Maybe it should be noted that formal clarity, precise meaning and even "simple" (mathematical) consistency entail, in almost all cases, "hard thought work", as the history of mathematics has shown:

• Geometry in the plane seems like an easy subject, until you start trying to understand its base and algorithms from Euclid's axioms and definitions, well over 2300 years old: There is nothing easy with concepts like parallels or ratios of line segment lengths! And later formalizations, mainly from about the 1800s onwards, are even more intricate.
• The other, apparently so "simple" basis of mathematics, namely the natural numbers, lost its simplicity also in ancient times with some prime number theory by the Greeks. It was and is by no means obvious what can emerge from simple addition and multiplication, let alone from the algebraic structures and formalizations extracted in the 19th century, leading to Gödel's mind-bending encodings and Turing's work.

Let me state this in my "Axiom 1": Mathematics, by and large, is not what we want in software documentation (and that from me, who majored in theoretical computer science ...).

Still, it seems we all want something more than the informal box-and-arrow-diagrams.

Gorton, like many others, proposes the use of UML. I cannot help the feeling that he is not really happy about it. The summary of chapter 6 has the following two sentences:

I’m a bit of a supporter of using UML-based notations and tools for producing architecture documentation. The UML, especially with version 2.0, makes it pretty straightforward to document various structural and behavioral views of a design.

"A bit of a supporter", "pretty straightforward": This does not really sound like wholehearted endorsement.

So, what is the problem?

The problem is, in my humble opinion, that there is no clear picture of what a notation for architectural documentation should do. The described use-cases typically oscillate between a "better notation" for those informal, easily comprehensible overviews over some aspects of a software system, and a more formal notation that can help derive hard knowledge about a system, with that implied goal of "generating code" in model-driven approaches.

I am, after many years in the field, now certain that we have to structure the use cases for architectural documentation in a threefold classification, with different notations for each area:

1. Informal documentation, from which humans can learn easily and intuitively gather a common understanding and a useful overview about some aspects of the system. In the best case, such a documentation is part of a common culture about "how we name and see things." However, this documentation is not intended to derive any hard facts: Everything shown can be disputed and discussed and viewed differently, and the notation can be extended at will if it helps with that intuitive understanding. All must agree that formal arguments based on such documentation are futile and hence must be avoided.
2. Formally sound and precise documentation that can be used to derive invariants and definitive properties of the documented system. If such documentation is used as the basis for a tool-supported model-driven approach, then there is no difference between a descriptive and a prescriptive architectural documentation for the aspects covered by the process. However, such an approach is very expensive in more than one respect:
   • First, especially without full tool support, keeping such a documentation in line with the system is much work, as even tiny changes on one or both sides require precise updates.
   • Second, as software can exhibit very complex behavior, the notation must be capable of describing many and, usually, deep concepts, which makes it hard and "mathematical" to understand and even harder to write. Such documentation therefore blatantly contradicts "Axiom 1".
   • Last, on a conceptual level, it is not really clear that such a documentation is actually "documentation" in the sense of "humanly accessible information relevant for many decisions in the software life-cycle". Rather, it might be more of a formal specification or even—when used in a model-driven process with code generation—part of the implementation, albeit (maybe) on some higher or "more compact" level than standard programming languages.

Thus, rich informal and deep formal notations are not sufficient for documenting and arguing about architectural aspects of a software.

3. Therefore, we need notations that are somewhere in-between: Not informal, so that they can be used to derive and ensure hard facts. But equally, they must be easily usable so that they can be read and written by the average software engineer under average project circumstances. It should be obvious that this type of notation cannot be very rich and also not very abstract. Only then, it can on the one hand avoid requiring an extensive semantics for formal derivations, and on the other hand being too esoteric to be used for understandable documents. In other words, it must be a quite mundane notation. I'll show my preferred notation for this, and its uses, in later postings—just in case you think that this looks a little like the search for the holy grail.

UML, incidentally and unfortunately, does not work really well for any of these purposes if its complex semantics is taken seriously:

1. For an informal notation, it carries a too heavy backpack of that formal semantics which no-one wants to remember when drawing informative diagrams in a running text (as, e.g., in the case study in Gorton's book).
2. For a formal notation, it is too indirect: One needs to map UML propositions back to the underlying semantic model (like Petri nets or state machines), and only then one can formally draw conclusions; as far as I can oversee it, the number of publications that use UML as a formal base has declined quite a bit over the last years.
3. Finally, as a simple but yet strict notation, UML is much too baroque, because it was lobbied to include every useful diagram and icon. This large notational size would recommend it for many different informal diagrams—if it weren't for that formal semantics ballast ...

But even if  you think that UML does work well (or well enough) for one area, there is the danger of misinterpreting UML diagrams: Is a diagram which your team uses as a basis for a decision a "type 1." diagram?—then it conveys informal concepts, but does not limit the decision strictly or formally. A "type 2." or "type 3." diagram, on the other hand, would narrowly limit some choices you can make—and definitely require a formally (for "type 2.") or at least collectively (for "type 3.") approved update of the diagram for any change in the software or the architecture. But most diagrams do not spell out explicitly their "conformance level".

Nonetheless, our analysts and some of our developers and architects (including me) are happy enough to use UML as a pool of symbols for sketching explanatory diagrams that help us to keep our complex machinery at least somewhat documented. So yes, I am, and we are also "a bit of a supporter of using UML-based notations and tools", as Ian Gorton puts it.

But now, I feel, I am starting to owe you an explanation how to do architectural documentation better. The next posting ... well, after I wrote it, it turned out to still consider some general observations about software architecture and how we deal with it.