layout: post title: Executable Architecture Documentation description: “Executable Architecture Documentation” author: jens_nerche tags: [MPS, Java, Language Workbench, arc42, Asciidoc, PlantUML, Continuous Integration, jQAssistant, Neo4j, Cypher, Architecture, Test]
In a former blog post How to keep architecture and it’s documentation in sync I described how you could ‘execute’ an architecture documentation as a test.
But the diagrams are just pictures (or text to be precise), not models. You cannot:
- navigate from one package in diagram A to the same package in diagram B just like in an IDE
- perform model validations
- define constraints
- execute refactorings (e.g. rename) like in an IDE, you have to do Find&Replace instead
- reference a class define in a class diagram in a sequence diagram
- reference artifacts of a diagram in the text as a real link between them
- find usages of diagram artifacts
- do lots of other interesting stuff
Make it better
Make a model of Package Diagrams
Not so surprising my first goal was to make a model out of the plain text. Being an MPS guy, naturally my first choice was to create and use a Domain Specific Language and create textual models with it. I re-used the PlantUML concrete syntax and published a first prototype of the MPS language for PlantUML to GitHub.
But an architecture documentation does not only consist of diagrams. Equally important is a description of the building blocks, discussions on design decisions and so on.
Bring the documentation to the same platform
The models and the documentation can only be combined if they are on the same platform. So I also had to bring the documentation into the MPS PlantUML model. Therefore another DSL is needed in which the documentation can be written. There exist lots of markup languages which provide excellent rendering results, e.g. the most famous (La)TeX. But for my experiments I chose the new shooting star Asciidoc for several reasons which become clear down below.
For this I created the MPS language for Asciidoc. Models of this language were compiled to Asciidoc files which can be directly used by Asciidoc text processors like Asciidoctor. A nice thing about Asciidoctor is that it can also handle embedded diagrams using Asciidoctor Diagram. Also PlantUML is supported.
Template for Architecture Documentation: arc42
There is a good template for documenting a software architecture: arc42. In the download side there are several versions. But the most interesting resides on GitHub: Asciidoc version of arc42 template This one comes with a Gradle build configuration which makes other formats like HTML, Markdown, and docx. Rumors say that also LaTeX may be supported in the future.
I imported the arc42 Asciidoc files into MPS in an Asciidoc model. Chapter 5 “Building Block View” is the one where the package diagrams fit in. But there is also Chapter 9 “Design Decisions”. Some Design Decisions have consequences in the code that could be monitored easily. Let’s take the example Logging Framework. After discussing the alternatives and making a decision for e.g. log4j, a new design constraint has to be put in place “do not use any class of java.util.logging”.
jQAssistant, Neo4j and Cypher
Until now there are a lot of more or less unrelated building bricks. To join them all, some mortar is needed. The mortar is the MPS project for Executable Architecture Documentation. It combines the MPS languages for Asciidoc, PlantUML, and Cypher. Now it is possible to integrate UML diagram models and Cypher query models in Asciidoc documentation models. To my knowledge, there is no other tool which combines good text processing, UML modeling, a software structure database, and a database client. But that’s exactly the tool set needed by a software architect to define an architecture via models and description and check whether that architecture is reflected in the code. Leave one out and it doesn’t really work. Models without describing text are just nice pictures. An architecture definition without automated checks on every build will sooner or later diverge from the code.
Speaking about the automated checks: how did I achieve that?
Integrate into Gradle build
As described above, the Asciidoc documents serve two purposes:
- inform developers
- contain rules for the code
So the integration into the Gradle build has also to do two things:
- create some user friendly format like HTML or PDF
- check the rules against the code
For the format conversion there is the Installing and Using the Asciidoctor Gradle Plugin guide and also a Gist with example of integrating Asciidoctor in Gradle. The tricky part is to find most versions that work together.
Note if you are still forced to use JDK7: you may need to increase their available “PermGen” space for more complex projects via the gradle.properties setting of: org.gradle.jvmargs=”-XX:MaxPermSize=512m”.
For the rule checking part there is jQAssistant with Kontext E plug-ins. jQAssistant can handle Cypher rules embedded in adoc files out of the box, so the design rules of chapter 9 were supported directly. And for the chapter 5 part of the building blocks, I described in How to keep architecture and it’s documentation in sync how our PlantUML plug-in can be used.
Ok, now there are lots of different building brings, loosely coupled by some mortar. But what gives us that? Is there a downloadable product? No. Is this the one way to make all of us happy? No. Are there some success stories? No. Is there at least something usable in real life projects? Maybe - didn’t try that.
But what is it? It’s the idea. The idea to end the diverging parallelism of a documented and a really existing software architecture. The idea to have architectural rules defined in the architecture documentation and checked by the Continuous Integration build, giving not only the violations but also the section in the documentation containing the details. And all without repeats and with hard, navigable links between all parts. Now the architecture documentation has tightly integrated UML models in the same documents - or did you ever try to modify a UML model embedded in an Office word processor? And the documentation is executable in the sense that (currently some important) rules can be checked in the build process directly using the architecture documentation files.
I did some prototyping to show that idea. To show that it is possible, for having a starting point. What I showed above can be used in different configurations, e.g. without MPS as text only. Or you find a way to embed XMI into Asciidoc to replace PlantUML by your UML tool exports. Or you combine that with a Java project implemented in MPS or a C project implemented with mbeddr. All is open source or creative commons - you can play with it, send pull requests, report issues, create own forks, you name it.
There is still lots of work to do. All the MPS languages are prototypes to explore the idea. They have to be developed to a usable state. Also a downloadable all-in-one release is essential for someone who wants to try out the stuff. And of course gathering hands-on experience in real life projects to get feedback for the direction of the development is really needed. Then there are more chapters in the acr42 template, and I’m sure that some of them can also have testable rules. Last but not least it would be nice to execute the architecture tests from within MPS just like unit tests were executed within an IDE. This already works in the Executable Specification project (in German) and could be ported.
And I’m pretty sure some of you readers come up with lots of additional ideas. So let me know - in GitHub or via mail to jens dot nerche at kontext dash e dot de.